Glean:GO 2026 Keynote  🔗 **Source:** [Glean:GO 2026 Keynote](https://www.youtube.com/watch?v=vGmmt7p2Hyk) > [!note]+ Document Notes > tags: [enterprise-ai, ai-agents, model-routing, token-economics, ai-governance, enterprise-search] > author: Glean > source: Glean (YouTube) > reference: https://www.youtube.com/watch?v=vGmmt7p2Hyk > date created: 2026-09-21 09:00 > > > [!info] Generated analysis — not the author's words. Sections below are machine-written > > from the captured source. Only text inside `> [!quote]` blocks is verbatim from the author. > > Do not quote or attribute any other line to the named author. > > # 🏷️ Title: Glean:GO 2026 Keynote > <!-- deep-read:sections start --> ## 🧭 Table of contents - [📝 Summary](#-summary) - [📖 Main content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning styles assessment](#-learning-styles-assessment) - [📊 Accessibility assessment](#-accessibility-assessment) - [🌍 DEIB assessment](#-deib-assessment) - [✅ Logical validity](#-logical-validity) - [🛡️ DARVO analysis](#️-darvo-analysis) - [🎯 Conclusion](#-conclusion) - [🔄 Refactor appendix](#-refactor-appendix--gleango-2026-keynote) ### 🔗 Source links - [Glean:GO 2026 Keynote (YouTube, 1:48:58)](https://www.youtube.com/watch?v=vGmmt7p2Hyk) - Transcript captured locally: `/Users/courtneyrobertson/.youtube_analyses/transcripts/vGmmt7p2Hyk.txt` (16,470 words) > [!warning] Transcription reliability > This is an auto-generated transcript with heavy speech-recognition noise. "Glean" appears as > "Gleam," "Gleen," "Clean," and "Le." Product and person names are garbled: Glean Tau appears as > "Clean Tow," "Glean Tao," "Gleantow," and "Glean Towels"; Nemotron as "Neotron"/"Neatron"; Kimi > as "Kimmy"; DaVita as "Dvita"/"Dvidita." Dollar figures lose their "$0." prefix. Every number and > name below is recorded as spoken; where the transcript contradicts itself, both readings are shown. ## 📝 Summary - **What happened.** Glean ran a 109-minute opening keynote at Glean:GO 2026 in San Francisco (Fort Mason) on 2026-09-21, streamed to what the host put at 5,000+ attendees. The company announced a proactive-AI product line (interactive dashboards, independent agents, proactive task management, team chat), Glean Transform, auto routing under a new "Glean Intelligence" umbrella, agent governance policies, context-aware AI threat detection, the Glean AI gateway, and Glean Tau, a desktop app for code execution and co-working. - **Why it matters if you use Glean at work.** Three of these change day-to-day operating decisions rather than positioning: auto routing plus AI usage controls change what a Glean seat costs, AI gateway changes whether your memories and skills stay portable across Cursor, Copilot, and ChatGPT, and agent governance policies plus threat detection change who signs off before an agent gets write access. - **Key stakeholders.** Glean: Arvind Jain (founder and CEO, transcribed "Arvin Jane"), Amar Malitira (COO), head of product (name garbled — introduced as "Amrajan," self-introduced as "John"), Tony (token yield), Eddie (evals), Michael (finance), Aaron Grove (VP solutions engineering), Judith Bramwell (EMEA lead, addressed as "Juliet" at the close), plus demo presenters Selene and Blair. External: Phil Lutka (executive director, core IT, General Motors), Kate Rooney (CNBC), Alexander (OpenAI, enterprise products, formerly leading Codex agent work), Carrie (NVIDIA), Evan Spiegel (Snap, recorded), "VC" (SVP engineering, Snowflake), Madu Nasharim (CIO, DaVita). - **Risks and tradeoffs.** The headline 81% token-cost saving and 78% preference rate come from Glean's own benchmark against one named competitor, Claude Co-work, with the query set, harness, and evaluation environment disclosed only in a single sentence and not published on stage. Adoption numbers cited by customers are impressive and internally inconsistent in at least two places (DaVita's 17,000 vs 70,000 users; Snowflake's "16x swing" on a 15/60-to-60/15 NPS shift). Nothing announced carried an explicit GA, beta, or preview label. - **Bottom line.** Treat the product list as a real roadmap signal and the numbers as vendor-supplied until you rerun them on your own corpus. The two items most worth a procurement conversation this quarter are auto routing with department-level spend profiles and AI gateway memory portability, because both have contract and lock-in consequences beyond the feature itself. ## 📖 Main content ### The argument the keynote makes The keynote runs a three-problem structure that Arvind Jain sets up at roughly 13 minutes and every subsequent segment answers in turn. The three problems, in his framing: enterprises can't get the right context to their agents, so quality is poor and business value doesn't land; AI costs more than most enterprises can afford; and AI has created new problems of its own, specifically a proliferation of AI-written documents, skills, and agents that are hard to manage and secure. Each block of the keynote maps to one problem. Context is answered by the context engine (retrieval plus enterprise graph) and by proactive AI. Cost is answered by model choice, auto routing, and AI usage controls, bundled as Glean Intelligence. Sprawl and security are answered by Glean Protect additions, agent governance policies, context-aware AI threat detection, and the AI gateway. ### Context: retrieval plus an enterprise graph, positioned against MCP Glean's definition of context is the collective human intelligence and experience of an enterprise, packed so agents can operate at the level of human experts. The stated data problem is that structured data and data lakes capture only 5 to 10% of enterprise knowledge and that playbooks and SOPs cover only a slice more; the rest lives in people's heads and surfaces in emails, messages, meeting notes, and exception approvals. The mechanism: retrieval surfaces current high-quality information across systems; the enterprise graph maps people, customers, teams, projects, and their relationships. Glean says it builds these ontologies largely automatically and exposes them for you to visualize, work with, and change. The competitive framing here is the most load-bearing and least supported claim in the keynote. Jain positions the graph against "the industry's typical approach of delivering shallow contexts via MCP" and then states Glean is the only context and intelligence company that truly understands your enterprise. No comparison, test, or definition of "shallow" accompanies that claim. Note the tension with a later segment: the AI gateway is itself described as a control plane for LLMs *and MCPs*, and Glean reports 78% of its customers reaching Glean through APIs and MCP. ### Proactive AI: five things, one bet The head of product frames human attention, not AI capability, as the bottleneck, then announces proactive AI in five parts: 1. **Proactive task management** — surfaces what needs attention, prioritizes it, and prepares the next step. Demoed as Glean having already drafted a customer slide deck, carrying the engineering root cause into it. 2. **Interactive dashboards** — called "our newest artifact." A dynamic, shareable view combining CRM structured data, warehouse usage telemetry, and unstructured customer data. Demoed on a renewal-health dashboard where a fictional account, Frasier Automotive, shows red because an API failure hit during performance reviews. 3. **Independent agents** — long-running agents with their own distinct identity, memory, and context, able to act proactively and coordinate across teams, with governance so they "don't go off the rails." Demoed as an on-call agent in Slack identifying an escalation, analyzing the issue, reading code, opening a pull request, and tagging engineers. 4. **Memory and skills** that learn and self-heal continuously. 5. **Team chat** — a shared space where multiple teammates and Glean work in one thread. This is the segment's actual thesis: AI moving from single-player to multiplayer. ### Glean Transform Announced as a living representation of how work gets done in a company, built on the enterprise graph, sold on the premise that transformation starts with visibility into what needs transforming. Demoed on Glean's own instance: 78 workflows transformed, 272 opportunities across all departments, 38 opportunities inside the sales team alone. The workflow drill-down uses a discovery-call and use-case-mapping process, identifying that sales development reps context-switch between Gmail and Google Drive during account research. Transform then builds a blueprint for a call research agent, deployable in one click with instructions, tools, knowledge, and config already populated. On the measurement side, Transform reports that Glean's own customer escalation workflow became 46% faster after deploying a diagnostic agent, saving 40 minutes per ticket through faster resolution and more accurate ticket assignment. ### Token yield: model choice, auto routing, usage controls Tony opens by naming the business-model alignment: Glean's model improves when your tokens do more work, not when you burn more. Three levers follow. **Model choice.** Over 40 open and frontier models, with day-zero support as models drop, described as landing roughly every other week. Eddie frames 2026 as the year open models caught up, on the argument that the industry stopped measuring maximum intelligence and started measuring cost-performance. **The Pareto frontier chart.** Built on enterprise data across departmental task variety. Named positions: GPT 5.6 Luna dominating a large part of the frontier on bang for buck; Kimi K3 (transcribed "Kimmy K3") and GLM 5.2 (transcribed "GLM52") in the middle; Opus 5 at the top right for strongest intelligence. Eddie also notes the weekend's "ox alpha" model is known to be GLM 5.3 Flash and expects it at the frontier, without having re-plotted. He states plainly that the chart would look different a month earlier, because OpenAI's price cut hadn't happened and Kimi K3 didn't exist yet — which is also the argument against single-provider lock-in. Task-level breakdown: information-seeking tasks have a flat cost-performance slope and get served by inexpensive models; coding has a steeper slope with a performance jump at frontier open models and Opus 5. **Auto routing.** Announced during this segment. Routes across the 40+ models and also selects reasoning level — the example given is GPT 5.6 Luna on low reasoning rather than high for simpler tasks, and Opus 5 for visual tasks and visual artifact creation. Runtime routing is performed by **Glean Waldo**, a specialist model Glean post-trained on strong open-source base models, named as the Nemotron family and later specified as Nemotron 3 Nano. **The benchmark.** Glean says it saves 81% on token costs versus Claude Co-work, with Glean's response preferred 78% of the time, and that the preference held across every individual department. Per-task cost is given as $0.58 for Glean against $2.98 for Claude Co-work (the transcript renders the first figure as "58 per task"). Glean attributes the saving to two compounding factors: cheaper model selection, and 70% fewer tokens consumed. The token reduction is credited to a precomputed central index and enterprise graph, versus harnesses relying on federated MCPs that search live across sources with unreconciled search quality, then dedupe, filter, and join while tokens burn. Eddie's own caveat is the honest part of this segment and is worth keeping attached to the number whenever it's quoted internally: these are Glean's results on Glean's benchmark, involving Glean's query-set composition, conditioned on the choice of harness and on the evals in the environment. **AI usage controls.** Caps are described as a backstop, not a strategy — the framing is an engineer hitting a limit on the 20th of the month. What ships alongside: auto routing that is spend-aware and downshifts to cost-efficient models as budget depletes (compared to a phone's low-power mode); admin-set profiles per department across cost-efficient, balanced, and frontier-level intelligence; self-serve analytics over AI usage data answering questions like who needs a nudge off Opus 5 and whose agents are still running unattended; and spend forecasting. Model choice, auto routing, and usage controls are bundled as **Glean Intelligence**. ### Security: from guarding data to watching patterns The security segment's premise is that a year ago AI wasn't accessing your data or writing back to enterprise systems, and now a single user can spin up thousands of agents. At that scale, the speaker argues, guarding data and actions isn't enough — you have to watch for patterns, specifically an agent with valid permissions behaving in ways that put the organization at risk. The keynote's phrase for this is "AI as an insider threat." Shipped since last year's Glean Protect announcement: models detecting sensitive content at 95% accuracy; restricted topic policies; alignment models that verify agent actions against the user's intent at runtime. Announced here: **agent governance policies** (controlling what agents can read, restricting high-risk writes, access controls at the individual agent level) and **context-aware AI threat detection**, which spots high-risk agentic behavior across sessions and uses enterprise context to separate legitimate activity from a potential attack. The demo is the clearest technical illustration in the keynote. A dashboard shows four issues detected two days prior. One is a sensitive-data exfiltration. A support engineer, Alex, asks Glean Assistant to reconstruct a Zendesk issue for customer Atlas Labs, find similar cases, and draft an RCA. Assistant pulls two similar internal RCAs from other customers. Alex then has Assistant reply on the Zendesk ticket with the generated RCA — which contains detailed information from Northstar Health and Pioneer Bank. Every individual step had valid permissions and no malicious intent. The enterprise graph catches it by understanding both customer boundaries and Alex's actual book of work. ### AI gateway and Glean Tau The **AI gateway** is announced as a unified control plane for LLMs and MCPs, securing and governing models, context, skills, and memories across every front door, with apps inheriting the auto-routing efficiencies and cost controls. The second announcement inside it is that memories and skills become portable across apps: Glean aggregates memory from chats, agents, and other AI apps and makes it shareable. The stated grievance is that memory currently locks you into a single app and becomes another data silo. The named front doors Glean says it makes better: Cursor, Copilot, ChatGPT. The stated principle is AI sovereignty — that enterprises should own and govern models, context, skills, and memories, because front doors come and go. **Glean Tau** (transcribed variously as Clean Tow, Glean Tao, Gleantow) is a desktop app for code execution and co-working, built by combining an open-source coding harness with Glean's context platform and AI gateway, named after one of its creators. Jain says the majority of Glean's company already uses it internally as their primary codegen and co-working product. The pitch is that it democratizes coding access for knowledge workers who delegate complex long-running tasks across authorized tools, and that it arrives preconfigured with governed enterprise context, models, MCP servers, skills, and memories in one click. The Tau demo is a sales-prospecting workflow: after an executive event at Wayfair Tavern with 34 attendees, Tau pulls the attendee list and a local voice memo from the presenter's computer, enriches both with enterprise context (is this an existing customer, what's the person's role, what's top of mind), produces a weighted lead score and ranked list, drafts and sends multiple outreach emails, then sets up an automation to monitor buyer intent signals across contacts met in the last six months and ping the seller in Slack when signals change. ### Contested and uncertain claims - **"The only context and intelligence company that truly understands your enterprise."** No defined test, no named alternative measured, no baseline. - **MCP as shallow context.** Asserted as an industry pattern without measurement, and complicated by Glean's own MCP surface area and the 78% API/MCP access figure. - **81% / 78% / 70%.** Single-competitor comparison on a self-built benchmark; methodology described in one sentence and not shown. - **"2026 is the year of open source in AI"** and "open-source harnesses often beat closed ones in benchmarks." Neither benchmark nor harness is named. - **Availability.** Not one announcement in the keynote carried GA, beta, preview, or a date. Treat everything as unstated. ## 🗒️ Notes #### ⏱️ [00:06:28](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=388) Judith Bramwell opens — EMEA lead - Frames the year: frontier and open models now coexist; the price-performance curve is resetting as the cost of powerful AI falls. - Four context scenarios given as the value story: seller walking into a renewal knowing promises made and open issues; developer opening a PR seeing affected systems, a similar outage from last year, and required reviewers; support seeing account history, product issue, fixes tried, and the engineer most likely to solve it; new hires connected to people, tools, practices, and policies on day one. - Customers named for the two days: General Motors, DaVita, Ericsson, eBay. Partners named: OpenAI, NVIDIA, Snowflake, Cisco, Deloitte (transcribed "Deote"), Databricks. - Sets the runtime expectation at "the next 90 minutes." #### ⏱️ [00:11:12](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=672) Arvind Jain — three enterprise problems - Attendance given as 5,000+ in the room and online. - Customer examples: GM engineers designing better cars faster; McCarthy transforming construction management so crews arrive with plans and safety procedures in hand; DaVita helping healthcare employees surface trusted information. - The three problems: context quality blocking business value, AI costs beyond what most enterprises can afford, and proliferation of AI-written documents ("slop"), skills, and agents that are hard to manage and secure. > [!quote] [00:13:53] > This is bots setitting. We are spending too much time working for AI as opposed to AI working for us. #### ⏱️ [00:15:37](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=937) The context engine - Glean founded 2019, positioned as having built the context foundation feeding enterprise AI across millions of users. - The data-coverage claim sets up the enterprise-graph argument. > [!quote] [00:15:37] > Structured data and data links capture only 5 to 10% of all enterprise knowledge > [!quote] [00:17:18] > Glean is the only context and intelligence company that truly understands your enterprise. #### ⏱️ [00:18:52](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=1132) AI sprawl, gateway, memory platform - Diagnosis: multiple AI front doors and agent platforms, each partially and inconsistently connected to company data. - Glean identifies highest-quality content, skills, and memories based on human and successful-agent usage — positioned as search expertise applied to agent inputs. - Memory platform learns employee working preferences from activity across enterprise systems, and learns how agents compound intelligence with human feedback. Framed as AI sovereignty. - The gateway brings closed and open-source models so customers can meet data-retention and sovereignty requirements. > [!quote] [00:19:42] > In fact, 78% of our customers now access glean through APIs and MCP. #### ⏱️ [00:21:27](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=1287) Token costs and the routing argument - Core claim: pointing the most expensive models at the simplest tasks is an architecture choice, not a limitation of AI. - Glean says it built an evaluation and training framework on real-world enterprise data, which produced its model-routing technology. > [!quote] [00:22:17] > We released a benchmark today showing that Glean saves 81% on token costs #### ⏱️ [00:23:08](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=1388) Open source and Glean Tau - Claim: 2026 is the year of open source in AI, covering both models and harnesses; open-source harnesses "often beat closed ones in benchmarks." Neither benchmark nor harness named. > [!quote] [00:24:07] > And one of these products is Clean Tow, a new desktop app that does everything from codegen to general co-working. #### ⏱️ [00:24:57](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=1497) Head of product — proactive AI announced - Attention, not capability, named as the bottleneck: users don't know which feature solves which job at the moment of need. - The five-part announcement: task management, email triage, and meeting coach; interactive dashboards; independent agents; memory and skills that self-heal; single-player to multiplayer. - Email triage and meeting coach are named once here and never demoed or detailed again. #### ⏱️ [00:27:57](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=1677) Interactive dashboards and team chat demo - Fictional HR SaaS scenario. Renewal-health dashboard combines CRM structured data, warehouse usage telemetry, and unstructured customer data. - Frasier Automotive shows red; Glean attributes the risk to an API failure hitting during performance reviews. - Team chat: AE and CSM added to a shared thread, AE asks Glean to draft a communication plan, CSM has Glean update next steps in the dashboard. #### ⏱️ [00:30:31](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=1831) Independent on-call agent in Slack - The agent identifies the escalation, analyzes it, reads code, opens a pull request, and tags the right engineers. - Memory is named as the mechanism that makes future resolutions faster. #### ⏱️ [00:31:21](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=1881) Proactive task management demo - Glean has pre-built a first-draft slide deck for an upcoming customer meeting, carrying the engineering root cause and resolution into it. #### ⏱️ [00:33:55](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=2035) Glean Transform announced > [!quote] [00:33:55] > Today we are announcing glean transform. - Glean's own instance: 78 workflows transformed, 272 opportunities across all departments, 38 within sales. - Discovery-call workflow drill-down finds SDRs context-switching across Gmail and Google Drive during account research; Transform generates a call research agent blueprint deployable in one click. > [!quote] [00:36:32] > Our customer escalation workflow has become 46% faster. - Attributed to a deployed diagnostic agent, with 40 minutes saved per ticket and more accurate ticket assignment. #### ⏱️ [00:37:22](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=2242) General Motors — Amar Malitira with Phil Lutka - Problem framed as access: a 100-year-old company, tens of thousands of employees, many business units, hard to find relevant current information. - Eight months into the partnership at the time of recording. - Adoption method: started from end-user friction, met two populations where they were (thousands who had never used AI, and a large group ready to build agents), invested heavily in enablement training and joining all-hands and staff meetings. > [!quote] [00:39:32] > So they went from zero to 60,000 users. > [!quote] [00:41:27] > We we sort of look at the typical metrics, but 5D30 is where we anchor. - 5D30 (five active days in thirty) still above 80% organization-wide as of that morning. - Named use case: the DR agent — a design release engineer agent built with Glean's forward deployed engineering team, serving roughly 6,000 design release engineers, consolidating issue, warranty, telemetry, and engineering data from disparate systems. A dedicated session was scheduled in room 11 at 11:00. - Forward strategy: don't bolt AI onto processes that shouldn't exist; assume the process can be completely different. #### ⏱️ [00:45:08](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=2708) CNBC panel — Kate Rooney, Arvind Jain, Alexander (OpenAI) - Alexander: led OpenAI's agent work with Codex for about a year, now leads enterprise products. - Jain on why AI isn't proactive yet: even at the most advanced technology companies people barely use AI to its potential because they don't know what it can do; proactivity requires knowing a person's role, OKRs for the next three months, weekly tasks, and how they write and read. - Alexander on the enterprise ask: reduce change management — leaders don't want to train every employee to prompt correctly or manually inject context. Form factor matters; a coding agent, an automation, and a teammate are used differently. - Jain on agent security, three rules: an agent acting on your behalf must not access data you can't see, must not make changes you couldn't make, and should be further restricted to the scope of tasks it was built for. Independent agents with their own identity need their own permissions and guardrails — his example is a 10-person support team gaining an 11th member that is an agent. - Alexander's layered approach: start agents connected to nothing and add context incrementally (slower than the CEO wants, faster overall because security stays comfortable); add observability so admins can review agent actions and tighten them; then shift from agents impersonating people toward agents with role-specific permission sets. - Alexander also predicts agent execution shifting to the cloud and off devices, which he argues is more secure. - Jain's economic caution is the sharpest line in the panel: AI is often becoming more expensive than the business value it delivers, and companies need to not run out of money before that flips. - Alexander's closing advice: hire high-agency people and structure teams to scale their patterns. #### ⏱️ [00:59:01](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=3541) Tony — token yield framing > [!quote] [00:59:01] > Our business model gets better when your tokens do more work, not when you burn more of them. - Three levers named up front: model choice, auto routing, AI usage controls. #### ⏱️ [01:00:45](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=3645) Eddie — model choice and the Pareto frontier > [!quote] [01:00:45] > That's why Glean gives you over 40 open and Frontier models - Day-zero support as models drop, described as roughly every other week. - Pareto definition given on stage: a model is Pareto optimal if nothing beats it on performance at the same cost or matches its performance for less. - Frontier positions named: GPT 5.6 Luna dominating much of the curve; Kimi K3 and GLM 5.2 mid-curve; Opus 5 top right. "Ox alpha" from the weekend identified as GLM 5.3 Flash, not yet plotted. - Obsolescence acknowledged: a month earlier the chart differed because of OpenAI's price cut and because Kimi K3 didn't exist. Used as the argument against single-provider bets. - Task slope: information-seeking flat (cheap models suffice), coding steep (frontier open models and Opus 5 jump). > [!quote] [01:02:01] > So it's important to note you know these are our results on our benchmark. #### ⏱️ [01:04:27](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=3867) Auto routing announced > [!quote] [01:04:27] > So today we're announcing auto routing in Glean. > [!quote] [01:04:27] > So that means we're talking about 58 per task with Glean versus $2.98 with Cloud Co-Work on a similar task. - The "58" is $0.58; the transcript drops the "$0." prefix. - Routing picks both model and reasoning level. Opus 5 for visual tasks and visual artifact creation; GPT 5.6 Luna on low reasoning for simpler tasks. - Glean Waldo: Glean's own post-trained specialist router, built on strong open-source base models from the Nemotron family, routing to the right reasoning level on demand. > [!quote] [01:05:58] > We actually use 70% fewer tokens. - Mechanism: precomputed central index plus enterprise graph versus federated MCP setups that search live across sources of differing quality, then dedupe, filter, and join while burning tokens. > [!quote] [01:06:34] > So, on quality, Glean was actually preferred 78% of the time - Preference reported as holding across every individual department. #### ⏱️ [01:07:55](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=4075) AI usage controls - Caps positioned as a backstop against a surprise bill, not a strategy; the scenario is your best engineer hitting a limit on the 20th. - Auto routing is spend-aware and downshifts to cost-efficient models as budget runs low — analogized to phone low-power mode. - Admin profiles set per department: cost efficient, balanced, or frontier-level intelligence. - Self-serve analytics over AI usage data: who needs a nudge off Opus 5, whose agents are still running in the background unattended. - Spend forecasting added so budgets can be set against trend. - Model choice, auto routing, and usage controls collectively named **Glean Intelligence**. #### ⏱️ [01:10:53](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=4253) Michael (finance) — the token whiplash video - Narrative arc: three years of telling everyone to token-max, it worked, costs rose, the CFO noticed. - Diagnosis: too many employees using frontier models for non-frontier work. Analogy used on stage: asking your CTO to reset a password. - Position taken: CFOs aren't the villain, and this isn't less AI use, it's more responsible use. > [!quote] [01:12:07] > Tokens are just dollars in another currency. #### ⏱️ [01:12:48](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=4368) NVIDIA — Carrie on open models - Case for open models: researchers, startups, and enterprises can evaluate, understand what's inside, customize, guard, and inspect. Enterprises want to own their AI rather than rent it. - Claim: NVIDIA is one of the largest, if not the largest, contributor to open-source AI — releasing open weights, the training datasets, and the exact frameworks and recipes used to train the models. - Sovereign AI builders named: Naver in Korea (transcribed "neighbor"), Sarvam in India (transcribed "Sarbomb"), and a German effort (transcribed "Sufi"). Companies named: CodeRabbit for code reviews, Perplexity for AI search, and Glean Waldo. - Glean Waldo confirmed as trained on Nemotron 3 Nano; Glean also offers Nemotron Ultra in the product. > [!quote] [01:16:08] > like 90% of Nvidians use Glean every single day for this IT search. - NVIDIA's deployment: Glean sits in the back end of IT systems, connected to 20+ enterprise tools, across roughly 42,000 employees, ranked among the highest in an internal IT survey. Over 400 employee-built agents run in Slack grounded in Glean. - Carrie's build argument: RAG is only one part; agents are models, harnesses, and grounded information, plus connectors, permissions, governance, security, and compliance. NVIDIA wants its engineers on core competency and buys trusted back-end tooling. - Favorite internal agents named: NV Info and NV Bugs. #### ⏱️ [01:17:52](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=4672) Snap — Evan Spiegel (recorded) - Internal program where people across Snap share new Glean agents every month so patterns spread. - Agents named: a quarterly playback agent that assembles context on someone's work for performance reviews; a tech lounge agent handling common IT questions and self-serve troubleshooting and routing complex or sensitive cases to a person; an R&D tax credit agent that finds and organizes material so engineers stop reconstructing projects from memory, returning hundreds of hours to engineering. > [!quote] [01:20:04] > our AI powered support agent answers roughly 3.9 million questions from Snapchatters each month and has reduced support ticket volume by about 62%. - Advertising: first-pass image review automation rose from 40% in Q2 last year to nearly 90%. - Engineering: an internal AI code reviewer now reviews 90% of pull requests across Snap, with an estimated 30,000 hours of code review time saved. - Snap is building Specs, described as a new kind of computer for supervising intelligent systems rather than operating computers. #### ⏱️ [01:20:43](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=4843) Security — Glean Protect one year on - Framing: a year ago AI wasn't accessing your data or writing back to enterprise systems; now a user can spin up thousands of agents at a time, so pattern-watching replaces per-action guarding. - Threat model stated: an agent with valid permissions behaving in ways that put the organization at risk. AI itself named as the largest emerging insider risk. > [!quote] [01:22:24] > We've added models that detect sensitive content at 95% accuracy. - Also shipped: restricted topic policies; alignment models verifying agent actions against user intent at runtime. - Announced: agent governance policies — controlling what agents can read, restricting high-risk writes, and setting access controls at the individual agent level. > [!quote] [01:23:18] > That's why today I am super excited to announce contextaware AI threat detection. #### ⏱️ [01:24:14](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=5054) Blair — exfiltration demo - Dashboard shows four issues detected about two days prior; one is a sensitive-data exfiltration request. - Chain: support engineer Alex works a Zendesk ticket for Atlas Labs, asks Assistant to reconstruct the issue, find similar cases, and draft an RCA; Assistant pulls two similar internal RCAs; Alex has Assistant reply on the ticket with the RCA. - The RCA contained detailed information from Northstar Health and Pioneer Bank — two unrelated customers. - Detection mechanism: the enterprise graph understands customer boundaries and Alex's book of work, so it flags one customer's data crossing into another's boundary. Every step had valid permissions and no malicious intent. #### ⏱️ [01:26:22](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=5182) AI gateway announced > [!quote] [01:27:11] > That's why we are announcing Glean's AI gateway, a unified control plane for LLMs and MCPs - Governs models, context, skills, and memories across every front door; apps inherit the auto-routing efficiencies and cost controls. - Second announcement: memories and skills become portable across apps. Argument: memory currently locks you into a single app and becomes a data silo holding your entire history of working with AI. - Named front doors Glean says it improves: Cursor, Copilot, ChatGPT. - Gateway video restates the principle: choose which LLMs run (frontier or open), set budgets by user and by group, keep data, memories, and skills interoperable so the same question gets the same answer at any front door. #### ⏱️ [01:28:48](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=5328) Glean Tau demo - Positioning: democratizes coding access for all knowledge workers so they can delegate complex long-running tasks across authorized tools; ships configured with governed enterprise context, all models, MCP servers, skills, and memories in one click. - Demo: executive event at Wayfair Tavern with 34 attendees. Tau pulls the attendee list and a local voice memo, enriches both with enterprise context (existing customer, role, what's top of mind), produces a weighted lead score and ranked list, drafts and sends outreach emails, then sets an automation to monitor buyer intent signals across contacts met in the last six months and message the seller in Slack when signals change. - The harness is referred to once as "Cody Harness" — likely a transcription artifact of the harness name behind Tau. #### ⏱️ [01:33:40](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=5620) Snowflake — SVP engineering ("VC") - Opening distinction: using AI to save 20 minutes versus using it to accomplish 80% of your work. - Method: treat internal engineering teams like customers — survey them, hypothesize, run experiments. > [!quote] [01:35:15] > So 16x swing in 18 months. - The underlying figures: internal engineer NPS moved from 15 satisfied / 60 dissatisfied to 60 satisfied / 15 dissatisfied. The "16x" label does not follow from those numbers (see Logical validity). - Organizational read: working with coding agents is as much a social and organizational problem as a technical one, because people who spent 20 years mastering a craft now watch an agent do parts of it. - Two-phase playbook, quoting Andy Grove — let chaos reign, then rein in the chaos. Phase one: observe the self-selected "AI superheroes" discovering patterns. Phase two: raise the floor by teaching those patterns to engineers who don't want to spend their day learning new ones, and raise the ceiling by giving the superheroes more time to find new patterns. - The four-question enterprise model: what happened (analytics), what's happening now (Slack, Jira, Teams, Gmail), what will happen (predictive systems), and systems of record (SAP, Workday, Salesforce, Jira). The opportunity: an "enterprise brain" or agentic control plane tying all four together, subscribing continuously to human and agent activity, operating at every memory horizon and every level of the org hierarchy, governed with ACLs and role-based controls. - Team size given as 3,500 people. - Three ambitious problems he'd hand to AI: verifiable work (Snowflake is rewriting its core execution engine and compiler because ground truth exists and an agentic factory can iterate), complex reasoning across many systems, and migrations, which combine both. #### ⏱️ [01:41:25](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=6085) DaVita — Aaron Grove with Madu Nasharim (CIO) - DaVita is a kidney care company; patients frequently have comorbidities such as heart disease or hypertension, so they engage heavily with the health system. - Conversations began about a year prior; deployed across the organization in April 2026. - Population figures are internally inconsistent in the transcript: "all 17,000 people from the front lines in the clinics all the way to the back offices," then "of the 70,000 uh users almost 40,000 engage." Aaron repeats "70,000." - Usage: about 5.4 million queries since launch. > [!quote] [01:42:23] > on average people saved about 19 to 20 minutes - Use cases: synthesizing patient information across paper, faxes, PDFs, and APIs into a summary; proactive outreach to dialysis patients showing signs of falling off a treatment plan that requires three to four hours, three times a week. - Three adoption conditions named for a highly discerning clinical audience: it has to be in the workflow, it has to be reliable every time, and it has to add value — with the observation that healthcare has no return button once care is given. - DaVita runs an innovation marketplace where teammates share agents they've built. > [!quote] [01:47:05] > That's 108 million minutes or 75,000 days. - Immediately reframed: the real ROI is five extra minutes of caregiver time with a patient. #### ⏱️ [01:48:02](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=6482) Close - Attendees directed to breakouts and back at 3:45 for a closing session. The closing host is addressed as "Juliet" though the opening host introduced herself as Judith Bramwell — an ASR inconsistency, not two confirmed people. ## 📚 References - [Glean:GO 2026 Keynote (source video)](https://www.youtube.com/watch?v=vGmmt7p2Hyk) - [Glean](https://www.glean.com/) — vendor; the token-cost benchmark referenced on stage was said to be released the morning of 2026-09-21 and is not linked in the transcript - [OpenAI Codex](https://openai.com/codex/) — Alexander's prior remit before leading enterprise products - [NVIDIA Nemotron open models](https://developer.nvidia.com/nemotron) — Glean Waldo's stated base-model family (Nemotron 3 Nano); Nemotron Ultra offered in-product - [Model Context Protocol](https://modelcontextprotocol.io/) — the "federated MCP" pattern Glean contrasts itself against, and the protocol its own gateway governs - [Anthropic Claude](https://www.anthropic.com/claude) — Claude Co-work is the named benchmark comparison; Opus 5 is named as the top-right frontier point - [Z.ai GLM](https://z.ai/) — GLM 5.2 and GLM 5.3 Flash named on the frontier chart - [Moonshot AI Kimi](https://www.moonshot.ai/) — Kimi K3, named as mid-curve and as not existing a month prior - [Snap Specs](https://www.spectacles.com/) — Evan Spiegel's "new kind of computer" - Not linked because the transcript gives no URL: Glean Transform, Glean Tau, Glean Waldo, Glean Intelligence, Glean AI gateway, context-aware AI threat detection, agent governance policies. ## 🗣️ Paraphrase - Glean's opening argument is that enterprises are stuck coaching AI instead of being helped by it, and that the root cause is shallow context. The company's answer is a retrieval index plus an enterprise graph that maps people, customers, teams, projects, and their relationships, built largely automatically and editable by the customer. - The product answer to context is a set of features that act before you ask: dashboards that fuse CRM, warehouse telemetry, and unstructured data; long-running agents with their own identity and memory; task management that drafts the next artifact; and a shared thread where several people and Glean work together instead of one person chatting alone. - Glean Transform turns the enterprise graph into a map of how work actually happens, ranks workflows by transformation opportunity, generates preconfigured agents from those blueprints, and then reports the before-and-after so the impact claim has a number attached. - The cost argument is that pointing expensive models at simple tasks is an architecture choice. Glean's fix is over 40 models, a post-trained router called Waldo that picks both model and reasoning level, spend-aware downshifting, department-level intelligence profiles, usage analytics, and forecasting. - The benchmark claim is 81% lower token cost and 78% preference against Claude Co-work, with 70% fewer tokens consumed, credited to a precomputed index rather than live federated MCP searching. The presenter states on stage that these are Glean's results on Glean's benchmark, conditioned on harness and evaluation environment. - The security argument is that valid permissions are no longer sufficient, because an agent can chain individually legitimate steps into an exfiltration. Agent governance policies restrict reads, writes, and per-agent access; context-aware threat detection watches behavior across sessions and uses the graph to tell legitimate work from an attack. - The gateway argument is sovereignty: memory and skills currently lock you into whichever app you built them in, and Glean proposes to aggregate and make them portable across front doors, including competitors' front doors. - Glean Tau is the company's own demonstration of the thesis — an open-source coding harness wrapped in Glean context and the gateway, already the internal default, pitched as a way for non-engineers to delegate long-running work. - Customer segments supply the proof: GM's 60,000 users in 90 days with 80%+ five-day-in-thirty retention, NVIDIA's 90% daily usage across roughly 42,000 employees and 400+ Slack agents, Snap's 62% support ticket reduction and 90% PR review coverage, Snowflake's internal NPS reversal, DaVita's 5.4 million queries at 19–20 minutes saved. - The most quotable caution comes from Glean's own CEO rather than a critic: AI is often costing more than the business value it delivers, and the job is making sure companies don't run out of money before that inverts. ## 📰 Newsletter Glean spent 109 minutes on stage arguing that the expensive part of enterprise AI isn't the model, it's the context you make the model rebuild every single time — and then put a price tag on it: $0.58 per task versus $2.98, on their own benchmark, against Claude Co-work, with 70% fewer tokens burned. Their engineer said the quiet part into a live mic: these are our results on our benchmark, conditioned on our harness and our evals. Keep those two sentences stapled together when the number reaches your budget meeting. For anyone running an open-source community, the interesting announcement isn't the routing — it's the memory. Glean's AI gateway claims to make memories and skills portable across Cursor, Copilot, and ChatGPT, on the argument that memory has quietly become the newest lock-in surface. That's a governance question dressed as a feature, and it's the same question WordPress contributors have been asking about plugin data and hosting defaults for fifteen years: who owns the thing you built inside someone else's product? Three items worth borrowing regardless of your stack. NVIDIA's Carrie put the honest version of open-model value on record: evaluate, understand what's inside, customize, guard, inspect — a better contributor pitch than any license explainer. Snowflake's engineering SVP gave the cleanest adoption playbook of the day, cribbed from Andy Grove: let chaos reign, watch who invents the patterns, then raise the floor for everyone who doesn't want to spend their week inventing patterns. And DaVita's CIO named the three conditions for adoption in a discerning audience — in the workflow, reliable every time, adds value — which is a better docs checklist than most docs checklists. Learning resources and links: the [full keynote](https://www.youtube.com/watch?v=vGmmt7p2Hyk) runs 1:48:58 with the pre-roll ending around 6:00. The token-yield segment starts at [59:01](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=3541) and is the densest 12 minutes if you only have time for one. The security exfiltration demo at [1:24:14](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=5054) is the best teaching artifact in the whole keynote — a chain of individually authorized actions producing an unauthorized outcome. Steal it for your next security training, whatever you build on. ## ✅ Action items #### DevRel - Developer advocacy: build a 20-minute internal talk from the [59:01](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=3541) token-yield segment that reproduces the Pareto framing on your own workloads. Measurable outcome: a published price-performance chart for your five highest-volume task types. - Education: run an office-hours session on reasoning-level selection, not just model selection. The keynote's most transferable technical point is that GPT 5.6 Luna on low reasoning beats the same model on high reasoning for simple tasks at the same quality. - Developer experience: write a one-page quickstart for Glean's API/MCP path, given that Glean reports 78% of customers now reach it that way. Include an auth example and a permissions-inheritance note. - Developer marketing: attribute any internal Glean adoption push to the 5D30 metric Phil Lutka named, not to seat count. Seats are a procurement number; five-active-days-in-thirty is a behavior number. - Community management: start a monthly agent-sharing ritual modeled on Snap's program and DaVita's innovation marketplace. Both named it as the mechanism that spread patterns fastest; require each shared agent to declare its tool scope. - Internal/external/cross-project: open a feedback loop with your Glean account team on auto-routing profile defaults before they're set org-wide. The department-level cost/balanced/frontier profile is a policy decision being presented as a setting. #### OSPO - Policy and governance: write the decision rights for who can create an independent agent with its own identity. The keynote treats this as a product capability; it's a headcount-adjacent governance question in practice. - License compliance: Glean Waldo is post-trained on Nemotron 3 Nano and Nemotron Ultra is offered in-product. Confirm the license terms of every open base model in your routing pool, including redistribution and output terms, before enabling all 40+. - Security management: request the methodology behind the "95% accuracy" sensitive-content detection claim — what corpus, what precision-recall split, what false-negative rate — before treating it as a control in a risk register. - Software inventory: add agents to your SBOM thinking. NVIDIA reported 400+ employee-built agents in Slack; nobody on stage described how those are inventoried, owned, or expired. - Developer support: publish guardrails for the Glean Tau pattern specifically. A desktop app that reads local files (the demo pulled a voice memo off the presenter's computer) and sends outbound email is a materially different risk class than a search box. - Community engagement: NVIDIA's claim to be the largest open-source AI contributor, publishing weights, datasets, and training recipes, is worth verifying independently before citing it in your own upstream-funding case. - Strategic alignment: define the KPI you'd use to falsify a "46% faster workflow" claim internally. Glean Transform reports its own impact; pick an independent measure before deploying it. #### Community - Maintainers: run a governance clinic on the exfiltration pattern from [1:24:14](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=5054). Every step authorized, outcome unauthorized — that's a maintainer problem in any project that lets bots comment, label, or close. - Contributors: turn the Snowflake two-phase model into an onboarding path. Phase one identifies your pattern-finders; phase two documents their patterns for contributors who don't want to invent any. - Sponsors: if your project accepts vendor AI credits, publish a token budget and a forecast the way Glean's usage controls do. Transparent ledgers beat surprise bills in both directions. - Ecosystem: coordinate a cross-project drill on agent-initiated disclosure. If an agent opens a PR containing another org's data, whose incident is it? #### WordPress - Core: nothing in this keynote touches WordPress Core directly. The transferable item is the release-criteria question — what's your equivalent of "no GA label on any announcement," and would your release notes pass that test? - Core Performance: Glean's token-reduction argument is a caching argument in different clothes — precompute the index rather than rebuild context per request. Worth a profiling comparison against any AI feature that does live multi-source retrieval per page load. - Core AI: adopt the keynote's evaluation posture, not its conclusions. Build an evaluation set on real WordPress tasks before selecting models, and record the harness alongside the score, because the keynote proves how fast a frontier chart goes stale. - **Core Program Team:** the cross-team dependency worth naming is agent identity. If AI tooling gains distinct identities in .org systems, that spans Meta, Plugins, Themes, and Support at once. - Mobile: no mobile-relevant content in this source. Nearest transferable item is Snap's routing pattern — automate common questions, escalate complex or sensitive ones to a person. - Playground: build a reproducible lab that demonstrates the multi-step-authorized-exfiltration pattern in a sandboxed WordPress install. It's a teachable security module with no real data at risk. - Community: the GM enablement approach — join existing all-hands and staff meetings rather than schedule new training — is directly portable to meetup organizing. Meet the group where it already gathers. - Training: build a lesson on cost-aware model selection. The keynote's CTO-resetting-passwords analogy is the clearest teaching image available for why frontier models on simple tasks is waste. - Marketing: avoid the keynote's habit of stating a superiority claim without a baseline. "Only company that truly understands your enterprise" is the anti-pattern. - Support: DaVita's three conditions (in workflow, reliable every time, adds value) are a usable rubric for evaluating any AI assist in the support forums. - Design: interactive dashboards as a shareable artifact type is a design pattern worth studying — the demo's value came from drill-down and shared next-steps editing, not visualization. - Documentation: document availability status explicitly on every feature page. This keynote announced at least eight capabilities and labeled none GA, beta, or preview; don't reproduce that gap. - Photos: not applicable to this source. No imagery, licensing, or media-taxonomy content appears in the keynote. - TV: caption quality is the direct tie-in. This transcript garbles product names badly enough to make the source hard to cite; a human caption pass is the fix. - Polyglots: not applicable to this source. The keynote is English-only and names no localization or multilingual capability. - Accessibility: audit any AI-generated artifact (the demoed slide deck, the dashboards) for whether generated output carries alt text, heading structure, and table semantics. The keynote never addresses this. - Meta: if .org adopts any MCP surface, the keynote's federated-MCP critique is worth reading as a performance warning — unreconciled search quality across sources, then dedupe and join at runtime. - CLI: the transferable pattern is scoped permissions per command, matching the keynote's rule that an agent should be restricted below the user's own rights to the scope of its task. - Hosting: spend forecasting and per-department caps are the hosting-relevant primitives. Any host offering AI features needs both before the first surprise invoice. - Tide: the keynote's evaluation framework is the analogue — publish scoring methodology alongside scores, which is exactly what Glean's benchmark does not do. - Openverse: not applicable to this source. No content ingestion, media licensing, or moderation topic appears. - Themes: no theme-relevant content. Nearest transferable item is the review-throughput argument implicit in Snap's 90% PR review coverage by an internal AI reviewer. - Plugins: same as Themes — Snap's AI code reviewer covering 90% of pull requests and saving an estimated 30,000 hours is the datapoint to interrogate if automated plugin review comes up. #### The WPOCC - Programs: the Snowflake "AI superheroes" model maps cleanly to a fellowship track — fund the people already inventing patterns, then fund the documentation of those patterns separately. - Projects: a governance clinic on agent identity and permissions would be timely. Arvind Jain's three rules (no data the human can't see, no changes the human couldn't make, further restricted to task scope) are a usable starting charter. - Funding experiments: the keynote's spend-forecasting concept transfers directly to microgrants — publish the trend, not just the cap, and pay out against milestones the way Glean pays tokens against task value. ### FAIR - Technical independence: Glean's "AI sovereignty" framing — enterprises should own and govern models, context, skills, and memories because front doors come and go — is the same architectural argument FAIR makes about metadata and distribution. Worth citing, with the caveat that the sovereignty here is delivered by a single vendor's control plane. - Policies and covenants: the memory-portability announcement is an anti-capture claim. Ask what the export format is, whether it's documented, and whether portability survives contract termination — none of which the keynote addresses. - Productized governance artifacts: agent governance policies (read restrictions, high-risk write restrictions, per-agent access controls) are a template worth adapting into a written operator agreement for any automation granted repository or registry access. ## 🧪 Hypotheses 1. **Auto routing cuts our per-task cost by less than 81% on our corpus.** The 81% figure comes from Glean's query-set composition against one competitor harness. If our task mix skews toward coding — where Eddie showed a steep cost-performance slope — routing has less cheap headroom to exploit, and the saving compresses. 2. **Most of the saving comes from token volume, not model price.** Glean attributes 70% fewer tokens to the precomputed index. If true, organizations with poorly indexed or permission-fragmented content will see much smaller savings, because the index can't precompute what it can't see. 3. **Department-level intelligence profiles will produce quality complaints within one quarter of rollout.** A department set to "cost efficient" gets systematically weaker models. The keynote presents this as budget management; the predicted failure mode is uneven output quality tracked to org chart rather than task difficulty. 4. **Context-aware threat detection will generate more false positives on cross-functional workers than on single-account workers.** The demo's detection logic keys on Alex working only Atlas Labs. Someone who legitimately spans many customers or business units breaks that signal. 5. **Memory portability via AI gateway will be partial, not symmetric.** Glean can likely aggregate memory out of other apps more readily than those apps will accept memory in. If so, the gateway reduces one lock-in while creating another at the aggregation layer. 6. **Proactive task management drives adoption more than agents do.** Jain's own diagnosis is that people don't know what AI can do for them. A pre-drafted artifact waiting in the assistant requires zero prompting skill; building an independent agent requires a lot. ## 📊 Methodology - **For hypothesis 1 and 2 (cost).** Sample 200 real tasks from your own logs, stratified by the keynote's own categories (information-seeking versus coding versus artifact creation). Run each through Glean auto routing and through your current default, capturing input tokens, output tokens, model selected, reasoning level, and wall-clock latency. Report cost per task as a distribution with a median and a 90th percentile, not a mean — vendor benchmarks that quote averages hide the expensive tail. Hold the harness constant, since Eddie named harness choice as a conditioning variable. - **For the quality half.** Replicate the preference test blind. Two outputs per task, randomized order, three raters per item, a written rubric covering correct, complete, and usable in my actual work — the three criteria named on stage. Report inter-rater agreement. A 78% preference rate is meaningless without it. - **For hypothesis 3 (profiles).** A/B two comparable departments for one quarter, one on balanced and one on cost efficient. Primary outcome: rework rate (tasks the human redid or discarded). Secondary: self-reported satisfaction. Cost saving is only real net of rework. - **For hypothesis 4 (false positives).** Before enabling enforcement, run threat detection in observe-only mode for 30 days. Classify every alert by whether the worker's role legitimately spans multiple data boundaries. Compute precision separately for single-boundary and multi-boundary roles. - **For hypothesis 5 (portability).** Test the export path directly: build a memory in Glean, export it, import it into Cursor or Copilot, then do the reverse. Record format, fidelity loss, and whether the round trip survives. - **For hypothesis 6 (adoption).** Instrument the 5D30 metric Phil Lutka named, segmented by first-touch feature. Compare 5D30 at 30, 60, and 90 days for users whose first meaningful interaction was a proactive task versus an agent build. - **Data sources available in this source.** The keynote supplies no raw data — no benchmark repository, no query set, no eval harness, no customer measurement methodology. Every number cited here is vendor-reported or customer-reported on a vendor stage, and none of it is independently reproducible from the transcript alone. ## ❓ Questions - What exactly is in the benchmark query set, and how closely does its department mix match ours? Eddie named query-set composition as a conditioning variable without describing it. - Which Claude Co-work configuration was tested, with what connectors and what index state? A harness comparison against an unconfigured competitor measures setup, not architecture. - Is Glean Waldo's routing decision auditable per task? If a task routes to a cheap model and the answer is wrong, can we see why it routed that way? - When spend-aware routing downshifts mid-month, does the user know? Silent quality degradation tied to budget is a trust problem, not a cost feature. - What is the actual availability of each announcement? Not one carried GA, beta, or preview. - What happens to independent agents' memory and identity at contract end? The sovereignty argument cuts both ways. - How does agent governance interact with existing IdP and SCIM provisioning? "Access controls at the individual agent level" implies an identity store nobody described. - Does the 95% sensitive-content detection accuracy figure describe precision, recall, or overall accuracy on a balanced set? Those are three very different guarantees. - Which open-source coding harness underlies Glean Tau, and what's the upstream contribution relationship? The keynote says open-source harnesses often beat closed ones without naming either. - DaVita's population is stated as 17,000 and 70,000 in the same answer. Which is the deployed base, and does the 5.4 million query figure normalize per user? ## 🔮 Future research - **An independent enterprise routing benchmark.** The keynote demonstrates both the need and the gap: every vendor will publish a chart where they win. A neutral, versioned, task-stratified benchmark with a published query set would be more valuable than any single vendor's number, and Eddie's own caveat is effectively an argument for one. - **Cost-per-outcome rather than cost-per-task.** Every figure in this keynote measures inputs (tokens, dollars per task) or proxies (minutes saved). DaVita's CIO named the gap out loud when he traded 75,000 days against five minutes of caregiver time. Nobody has a good method for the second kind of measurement. - **Detection of multi-step authorized exfiltration.** The demo describes a genuinely hard class of problem — a chain of individually permitted actions producing a prohibited outcome. The general question of how to specify, detect, and evaluate this class is open research, not a product feature. - **Memory interoperability format.** If memory becomes the lock-in surface, the missing artifact is a portable, inspectable memory schema. No standard exists; the keynote asserts portability without describing a format. - **Agent identity lifecycle.** Independent agents with distinct identities need provisioning, review, rotation, and deprovisioning. NVIDIA's 400+ Slack agents is a live dataset for studying what happens to agents nobody owns anymore. - **The organizational-change half.** Snowflake's SVP gave the only organizational model in 109 minutes, and it's a two-phase hunch, not a study. Whether "let chaos reign, then rein it in" outperforms structured rollout is testable and untested. - **Whether proactivity changes adoption breadth or just adoption depth.** Jain's claim is that proactive AI makes adoption uniform across an enterprise rather than concentrated in early adopters. That's a falsifiable distributional claim nobody tested on stage. ## 💡 Implications **For practice, if you already use Glean at work.** The immediate operational change is auto routing plus usage controls. Before your admin sets department profiles, get a say in which profile your team lands on, because "cost efficient" is a quality decision wearing a budget label. Ask for observe-only analytics first — the self-serve usage analytics described on stage will tell you whose agents are running unattended, and that report is worth having before caps land. If you build agents, the agent governance policies change your design surface: assume read scopes, write restrictions, and per-agent access controls become mandatory fields rather than optional hardening. **For the artifacts you produce.** Interactive dashboards and Tau-generated outputs are shareable artifacts that will circulate past their original context. Two habits protect you: state the data sources on the face of the artifact, and re-check any AI-assembled document for content that crossed a customer, client, or business-unit boundary. The exfiltration demo exists because that failure is easy and invisible. **For policy.** The memory-portability announcement is the one with contract consequences. If your organization is standardizing on Glean as a control plane across Cursor, Copilot, and ChatGPT, the exit cost moves from any single app to the gateway. Get export format and post-termination data access into the agreement now, while portability is the vendor's own headline claim. **For security.** The keynote's threat model is correct and worth adopting independent of the product: valid permissions plus an agent that can chain steps equals a new insider-risk class. Two controls follow that don't require buying anything — require a human review step before agent-generated content leaves an account boundary, and log agent action chains, not just individual actions. **For education and enablement.** GM's approach beat its technology: meet two populations separately, join meetings people already attend, and anchor on 5D30 rather than seats. If you're responsible for rollout anywhere, that's the most directly reusable thing in 109 minutes. Snowflake's phase-two framing adds the missing half — most people don't want to discover patterns, they want to be handed them. **For funding models and budgets.** The finance segment's honest admission is that three years of "token max" advice produced exactly the cost problem it now sells the cure for. Whatever your AI budget process, build the forecast before the cap, because a cap that hits your best engineer on the 20th converts a cost problem into a productivity problem. Jain's warning is the one to carry into any budget conversation: right now AI often costs more than the value it delivers, and the goal is to not run out of money before that inverts. ## 🧠 Learning styles assessment - **Reading level estimate:** Flesch-Kincaid Grade 7.3, Flesch Reading Ease 70.1, computed on the cleaned transcript (16,074 words, 1,042 sentences, 1.43 syllables per word, 15.4 words per sentence) with timestamps, speaker markers, and bracketed audio cues stripped. That's plain-spoken conference register. The score understates actual difficulty: the vocabulary load is where the demand sits — Pareto frontier, post-training, reasoning level, federated MCP, ontology, exfiltration, 5D30 — and none of those are common words. A listener without enterprise AI vocabulary will find this much harder than grade 7. #### 👀 Visual learners - Present: the Pareto frontier chart, the Glean Transform dashboard drill-down, the renewal-health dashboard, the threat-detection findings dashboard, and three produced video segments. Screen content carries a large share of the meaning. - Improve: the visuals are unnarrated in the audio track. The Pareto chart in particular is described with directional language only ("lower left quadrant," "top right"), so anyone listening without video loses the entire model-positioning argument. Add spoken axis values, a static published version of the chart, and an architecture diagram for the AI gateway showing where LLMs, MCPs, memories, and skills sit relative to front doors. #### 🔊 Aural learners - Present: this is a spoken keynote with a livestream, and the conversational fireside format carries well without visuals for the GM, Snowflake, and DaVita segments. - Improve: audio quality in the transcript suggests mic and crosstalk problems (repeated stutters, dropped names). Publish a 5-minute audio recap of the announcements and a separate 3-minute recap of the numbers, since the announcements are scattered across seven speakers over 109 minutes. #### 🗣️ Verbal learners - Present: strong analogy use — the phone's low-power mode for spend-aware routing, the CTO resetting passwords for frontier-model misuse, the credit card with no spending limit for unconstrained tokens, the 11th team member for an independent agent, Andy Grove's chaos line for adoption phasing. The on-stage definition of Pareto optimal is genuinely well-constructed. - Improve: key terms go undefined — 5D30 is used as though universal, "harness" is never defined despite carrying the open-source argument, and "token yield" is introduced as a coined term without a formula. A glossary card on screen at first use would fix all three. #### 🧩 Kinesthetic learners - Present: five live demos with visible clicks — dashboard drill-down, team chat thread building, Transform blueprint to one-click agent, threat-detection investigation, Tau prospecting run. - Improve: no sandbox, no follow-along path, no sample data. Every demo uses fictional companies (Frasier Automotive, Atlas Labs, Northstar Health, Pioneer Bank) that viewers can't reproduce. A published demo environment with those same fixtures would convert five watched demos into five completed exercises. #### 🧠 Logical learners - Present: the keynote has a clean three-problem structure with each segment mapping to one problem. The Pareto definition, the task-slope comparison (flat for information-seeking, steep for coding), and the cost decomposition (cheaper models plus 70% fewer tokens) are all properly reasoned. - Improve: no decision tree for when auto routing helps versus when it doesn't, no error budget or failure-rate discussion anywhere, and no methodology table behind the headline benchmark. The single most useful addition would be a table of task type against recommended profile against expected cost delta. #### 🧑‍🤝‍🧑 Social learners - Present: the whole team chat announcement is about multiplayer AI, and three customer segments describe peer-sharing structures — Snap's monthly agent-sharing program, DaVita's innovation marketplace, NVIDIA's 400+ employee-built Slack agents. - Improve: no moderated Q&A in the keynote itself, and the panel took no audience questions despite Alexander running a live hand-raise. Breakouts are referenced but not linked. A cohort guide for the agent-sharing pattern would turn three anecdotes into a reusable program. #### 🧘‍♀️ Solitary learners - Present: the segment structure supports self-paced viewing, and the token-yield block at 59:01 stands alone as a complete argument. - Improve: no chapter markers, no reflection prompts, no self-check. A published index of announcements with timestamps would cut the re-find cost dramatically for a 109-minute video with eight announcements. #### 🌿 Naturalistic learners - Present: almost nothing. The cost-efficiency argument has an implicit energy story — 70% fewer tokens and smaller models mean less compute — but nobody makes it. The phone battery analogy is the closest the keynote gets to resource thinking. - Improve: state the environmental corollary directly. Routing simple tasks to small models is the same argument as right-sizing infrastructure, and the compute-per-task figure is already being measured. Add lifecycle thinking for agents too: NVIDIA's 400+ agents and the "whose agents are stuck running" analytics question both point at abandoned compute nobody reclaims. ## 📊 Accessibility assessment Scope note: this assesses the published keynote video and its auto-generated transcript, which is what a viewer actually receives, plus what the demoed product UI reveals on screen. #### Screen reader semantics - Present: nothing assessable. The deliverable is a video; the product UI shown (dashboards, findings lists, agent configuration panels) is not inspectable for heading structure, landmarks, or accessible names from the recording. - Improve: publish an accessibility conformance report for the demoed surfaces — interactive dashboards in particular, since they are announced as a shareable artifact type that will be consumed by people who didn't build them. Maps to WCAG 2.2 AA **1.3.1 Info and Relationships**, **4.1.2 Name, Role, Value**, and **2.4.6 Headings and Labels**. #### Keyboard navigation - Present: not demonstrated. Every demo is driven by mouse clicks described aloud as "one click." - Improve: confirm that dashboard drill-down, team chat thread participation, and agent configuration are fully keyboard operable with visible focus. Maps to **2.1.1 Keyboard**, **2.1.2 No Keyboard Trap**, **2.4.7 Focus Visible**, and **2.4.11 Focus Not Obscured (Minimum)**, which is new in 2.2 and commonly failed by dashboard overlays and drill-down panels. #### Contrast and readability - Present: unverifiable from the recording. Stage lighting and video compression make on-screen UI contrast impossible to measure. - Improve: the Pareto frontier chart is the specific risk — a scatter plot distinguishing 40+ models by color and position, presented at distance on a conference screen. Maps to **1.4.3 Contrast (Minimum)**, **1.4.11 Non-text Contrast**, and **1.4.1 Use of Color**, since the chart appears to encode model family by color alone. #### Text resizing and reflow - Present: not assessable from video. - Improve: interactive dashboards combining CRM, telemetry, and unstructured data are dense by design; test at 200% zoom and 320 CSS pixels wide. Maps to **1.4.4 Resize Text**, **1.4.10 Reflow**, and **1.4.12 Text Spacing**. #### Media and captions - Present: YouTube auto-captions exist and a transcript is retrievable. This is the category with concrete, verifiable failures. - Improve: the auto-transcript is materially inaccurate on exactly the words that matter — the company's own name renders as "Gleam," "Gleen," "Clean," and "Le"; Glean Tau renders four different ways including "Glean Towels"; DaVita becomes "Dvita" and "Dvidita"; Nemotron becomes "Neotron" and "Neatron"; a dollar figure loses its "$0." prefix, changing $0.58 into 58. Long stretches of the pre-roll are captioned as "Heat. Heat. Heat." A product keynote whose product names are unreadable in captions fails its deaf and hard-of-hearing audience specifically on its core content. Human-corrected captions, a published transcript, and audio descriptions for the unnarrated Pareto chart would fix it. Maps to **1.2.2 Captions (Prerecorded)**, **1.2.3 Audio Description or Media Alternative**, and **1.2.5 Audio Description (Prerecorded)**. #### Forms and interactions - Present: the demos show form-driven configuration (agent setup with instructions, tools, knowledge, config; admin profile selection; budget setting) but no labels, error states, or validation behavior are visible. - Improve: budget and cap configuration is the highest-stakes form shown — a mis-set cap cuts off a team. Test for accessible labels, error identification, error suggestion, and confirmation before destructive change. Maps to **3.3.1 Error Identification**, **3.3.2 Labels or Instructions**, **3.3.3 Error Suggestion**, **3.3.4 Error Prevention**, and **2.5.8 Target Size (Minimum)**, new in 2.2. #### Standards reference - Primary gaps against WCAG 2.2 AA, in priority order: **1.2.2 Captions** (confirmed failure — auto-captions garble the core subject matter), **1.2.5 Audio Description** (confirmed gap — the Pareto chart argument is inaccessible without vision), **1.4.1 Use of Color** (likely, on the frontier chart), **2.4.11 Focus Not Obscured** and **2.5.8 Target Size** (unverified, both new in 2.2 and both commonly missed in dense dashboard UI). The two confirmed failures are in the published artifact and are fixable today without touching the product. ## 🌍 DEIB assessment #### Cultural breadth - Present: speakers join from London (EMEA lead) and the customer set spans automotive, healthcare, construction, semiconductors, social media, and data infrastructure. NVIDIA's sovereign AI examples name builders in Korea, India, and Germany, and the data-sovereignty and retention-requirement framing acknowledges that regulatory context differs by region. - Improve: every case study is a large US-headquartered enterprise, and every example workflow is US-shaped (Zendesk tickets, R&D tax credits, quarterly performance reviews). No small or mid-size organization appears, and no non-English deployment is described despite the EMEA lead opening the show. Add a European or Asian customer speaking about their own regulatory constraints, and one deployment under 1,000 seats. #### Inclusive language - Present: gender-neutral throughout. Roles are described by function rather than assumption, and the healthcare segment is careful and person-first about patients ("people who have kidney disease," "our caregivers"). - Improve: heavy undefined jargon is the real exclusion mechanism here — 5D30, harness, token yield, Pareto frontier, post-trained, federated MCP, RAG, ACLs. A newcomer to enterprise AI is locked out of the cost argument specifically, which is the segment with the widest audience. Also worth flagging: "bot sitting," "slop," and "AI superheroes" are in-group coinages that read as fluent to insiders and opaque to everyone else. #### Neurodiversity support - Present: consistent segment structure (announcement, then demo, then customer proof), short segments averaging under 10 minutes, and repeated verbal signposting — the three levers are named before they're explained and restated after. - Improve: 109 minutes with no published chapter markers, eight announcements scattered across seven speakers, and four separate music-and-video interludes that break attention without summarizing. Loud transition stingers with no warning are a sensory issue. Publish chapters, put a running announcement tally on screen, and offer a low-stimulation cut without the video interludes. #### Representation - Present: the speaker lineup spans founder, COO, product, engineering, finance, field marketing, and solutions engineering, with both technical and business voices, and mixes vendor voices with five external customer and partner voices who are given real airtime rather than testimonial slots. Women appear in host, demo, moderator, and partner-expert roles, including the NVIDIA technical segment and the CNBC panel moderation. - Improve: the balance of quoting skews hard toward executives — every external speaker is a CIO, SVP, executive director, or CEO. Nobody who actually uses these tools daily speaks: no support engineer, no design release engineer among GM's 6,000, no DaVita clinic worker. The people whose work is being transformed are described in the third person throughout. One practitioner voice per customer segment would change the evidentiary quality, not just the optics. #### Psychological safety - Present: two genuinely good moments. Snowflake's SVP names the emotional reality directly — people who spent 20 years mastering a craft now watch an agent do parts of it, and he describes staff walking into his office to say they went through their stages of acceptance that day. The finance segment explicitly defends CFOs as "not the villain here," defusing a common internal antagonism. DaVita's "healthcare has no return button" is an honest statement of stakes rather than a hedge. - Improve: no discussion anywhere of job displacement, workload change, or what happens to the 11th-team-member framing when the team was previously ten humans. Jain says plainly that "we're not going to be doing most of the things that we do today anymore" and the keynote moves on within one sentence. Alexander's high-agency framing — the most productive individuals could be running unicorn companies while less productive folks stay where they are — is stated as neutral observation and lands as a warning without any accompanying support model. A vendor keynote isn't obligated to solve this, but naming it and pointing at enablement resources would cost 60 seconds. ## ✅ Logical validity #### Strong reasoning - **The architecture-choice argument (causal, [00:21:27](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=1287)).** "Most systems still point the most expensive models at the simplest of the tasks. That's not really a limitation of AI. It's an architecture choice." This is strong because it identifies a specific, checkable mechanism and locates the cause in a design decision rather than a technology limit. Anyone can verify it against their own logs. To clarify further: state what fraction of observed enterprise tasks fall in the simple bucket, which would turn a mechanism into a magnitude. - **The single-provider argument (inductive, [01:03:14](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=3794)).** Eddie notes the chart he just presented would have looked different a month ago because of OpenAI's price cut and because Kimi K3 didn't exist, and then concludes you can't bet on one provider. The reasoning is strong precisely because it undercuts his own artifact — the evidence for the conclusion is the obsolescence of the evidence. Self-defeating slides are rarely offered voluntarily and this one earns credibility. - **The token-decomposition argument (causal, [01:05:58](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=3958)).** Splitting the 81% saving into cheaper model selection and 70% fewer tokens, then attributing the token reduction to a precomputed index versus live federated search, is a genuine mechanism rather than a correlation. It's also falsifiable: anyone can measure token counts on identical tasks across both architectures. - **The exfiltration argument (deductive, [01:24:14](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=5054)).** Every action was permitted; the outcome was prohibited; therefore per-action permission checking is insufficient. This is valid and the demo makes the premises concrete. It's the most rigorous five minutes in the keynote. - **The agent-permission rules (deductive, [00:51:11](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=3071)).** Jain's three rules follow from a single principle — an agent acting on your behalf inherits a subset of your authority, never a superset, and should be narrowed further to its task scope. Clean, implementable, and stated without product attachment. - **Adoption method as cause (analogical, [00:40:12](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=2412)).** Phil Lutka attributes GM's adoption to enablement effort rather than product quality, explicitly rejecting "we'll just turn it on." Attributing your own success to work you did rather than to the vendor on stage with you is the opposite of the expected move and makes the account more credible. #### Fallacies - **Unsupported uniqueness claim — [00:17:18](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=1038).** "Glean is the only context and intelligence company that truly understands your enterprise." No definition of "truly understands," no competitor evaluated, no test proposed. The immediately preceding sentence dismisses "the industry's typical approach of delivering shallow contexts via MCP" without naming a vendor or measuring depth. **Fix:** replace with a measurable claim — for example, retrieval precision at k on a named enterprise task set against two named alternatives — or drop the word "only." - **Self-serving benchmark presented as general result — [00:22:17](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=1337) and [01:04:27](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=3867).** The 81% and 78% figures are stated in the CEO's keynote as flat facts with no qualifier, then re-stated at the announcement. The methodology caveat arrives only at [01:02:01](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=3721), from a different speaker, about a different artifact (the Pareto chart), and is never reattached to the headline number. **Fix:** put the qualifier in the same breath as the number every time, and publish the query set, harness configuration, and competitor setup. Worth crediting: Eddie did state the caveat aloud, which is more than most vendor benchmarks manage — the fallacy is in placement, not in concealment. - **Missing baseline on the comparison target — [01:04:27](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=3867).** $0.58 versus $2.98 "on a similar task" tells you nothing about how Claude Co-work was configured. An unconfigured competitor is a strawman even when the comparison is made in good faith. **Fix:** publish the competitor's connector set and index state, or compare against two competitors so configuration error is visible. - **Faulty arithmetic presented as a result — [01:35:15](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=5715).** "our NPS with our internal engineers has gone from 15 satisfied 60 dissatisfied to 60 satisfied 15 dissatisfied. So 16x swing in 18 months." The NPS moves from −45 to +45, a 90-point swing. There is no operation on those numbers that yields 16. **Fix:** say "a 90-point NPS swing, from −45 to +45," which is a stronger and true claim. (This one is a customer's number, not Glean's, and nobody on stage corrected it.) - **Anecdote as generalizable proof — [01:17:52](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=4672) and [01:41:25](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=6085).** Snap's 62% ticket reduction and 30,000 saved review hours, and DaVita's 108 million minutes, are presented as what Glean delivers rather than what one organization measured under conditions nobody described. Snap's support-agent figure in particular covers a consumer support agent answering Snapchatter questions — that is not a Glean deployment metric at all, and it sits in a Glean keynote adjacent to Glean metrics. **Fix:** separate the slides. Label which numbers are Glean-attributable and which describe the customer's broader AI program. - **Internally inconsistent figure left standing — [01:42:23](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=6143).** DaVita's deployment is stated as "all 17,000 people" and then as "of the 70,000 uh users almost 40,000 engage" within one answer, and the interviewer echoes 70,000. The 5.4 million queries and the 19–20 minute saving are anchored to an unresolved denominator. **Fix:** state the deployed population once and derive everything from it. Per-user query rate is the number a buyer actually needs. - **Composition error in the savings claim — [01:47:05](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=6425).** 5.4 million queries times 20 minutes equals 108 million minutes, presented as 75,000 days. This treats every query's saving as fully additive and fully realized, which assumes no query was redundant, no saving was partial, and all recovered time converted to other work. To DaVita's CIO's credit, he immediately undercuts the arithmetic himself — the real ROI is five extra minutes with a patient. **Fix:** report a sampled, self-reported saving with a confidence interval instead of multiplying a mean by a total. - **Equivocation on open source — [00:23:08](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=1388).** "Open-source harnesses often beat closed ones in benchmarks" names no harness and no benchmark, and sits next to Glean Tau, whose underlying harness is also unnamed. **Fix:** name the harness and cite one benchmark. - **Tension worth flagging rather than a fallacy.** Glean disparages "shallow contexts via MCP" at [00:17:18](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=1038), reports that 78% of customers reach Glean via APIs and MCP at [00:19:42](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=1182), and announces a control plane "for LLMs and MCPs" at [01:27:11](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=5231). These are reconcilable — the critique targets federated live retrieval, not the protocol — but the keynote never does the reconciling, and a listener reasonably hears a contradiction. **What is not fallacious, checked against the transcript.** The Pareto chart is explicitly caveated on stage. The model positions are described as conditional on harness and evals. GM's adoption numbers are attributed to GM's own enablement work, not to product magic. NVIDIA's usage figures are given by NVIDIA about NVIDIA with a stated employee denominator. Jain volunteers that AI often costs more than the value it delivers — a statement against interest in a keynote selling AI. Several of these are better disclosure practice than vendor keynotes typically offer, and flagging them as spin would be inaccurate. #### Tone of voice - **Dominant register: confident practitioner, not hype.** The cost segment in particular reads as unusually self-aware — "Our business model gets better when your tokens do more work, not when you burn more of them" is an alignment claim offered for inspection rather than a superlative. - **Self-deprecating humor as a credibility device.** The opening wedding analogy, "not potato frontier," the CTO-password sketch, and Tony admitting "we made this slide, right? And it's already obsolete" all lower the temperature. The humor consistently arrives right before or after a hard number, which softens scrutiny — worth noticing as a rhetorical pattern, not a deception. - **Urgency without alarm.** "A year ago AI wasn't accessing your data... Now, agents are everywhere" and "AI as an insider threat" are threat-framing in a security segment, but paired with a concrete demo rather than left abstract, which keeps it on the defensible side of fear marketing. - **Deference in the customer segments.** The Glean interviewers consistently hand credit to the customer ("we are so grateful for your partnership," "I appreciate you trusting Glean with that"). This reads as genuine and also means nobody on stage ever pushes back on a customer's number — which is how the 16x and the 17,000/70,000 discrepancy both survive. - **One notable tonal outlier.** Alexander's closing advice — high-agency people running unicorn companies while others stay where they are — is delivered flatly as observation. In a room of enterprise buyers it reads as strategic; to anyone hearing it as an employee it reads as a warning. The keynote does not acknowledge the second reading. ## 🛡️ DARVO analysis DARVO (Deny, Attack, Reverse Victim and Offender) describes a response pattern to accusation or accountability. This source is a vendor product keynote with no accusation, accuser, or misconduct allegation in play, so the pattern does not apply in its proper sense. The honest finding is that there is no DARVO here. What follows are the nearest structural analogues, offered because the rhetorical moves are worth recognizing, not because they constitute the pattern. - **Deny — closest analogue: preemptive category denial, not denial of a charge.** At [01:11:34](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=4294) the finance segment states "this isn't less use of AI. It's just more responsible use of AI." That negates an objection nobody made on stage. It's framing management rather than denial, and it's transparent about doing it. - **Attack — closest analogue: competitor and category criticism.** Named or implied targets are criticized: "the industry's typical approach of delivering shallow contexts via MCP" at [00:17:18](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=1038), "most vendors want you inside their walls" at [01:27:57](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=5277), and Claude Co-work as the benchmark loser. This is ordinary competitive positioning aimed at products, not at people, and it names no individual's credibility. It isn't the attack component of DARVO. - **Reverse victim and offender — closest analogue: two reframings.** First, "speaking for CFOs, they aren't the villain here" at [01:11:34](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=4294) recasts the party imposing cost limits as a reasonable actor. Second, and more interesting, "AI as an insider threat" at [01:22:24](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=4944) reassigns the offender role from a human attacker to the AI system itself, in a keynote selling AI. That's a genuine role reversal, but it runs against the vendor's own interest rather than deflecting from it, which is the opposite of the DARVO dynamic. - **Emotional cues and rhetoric present.** Urgency ("agents are everywhere," "we've all seen the headlines"), in-group belonging (the opening wedding framing, repeated "we're building this together"), and self-deprecation used to lower scrutiny before a number. None of these operate to discredit a complainant, because there is no complainant. - **Related fallacies that do appear** are catalogued in the Logical validity section above: unsupported uniqueness, self-serving benchmark, missing baseline, anecdote as proof, and one arithmetic error. Those are the real weaknesses in this source, and calling them DARVO would overstate both their nature and their severity. ## 🎯 Conclusion Glean announced a lot in 109 minutes — proactive AI (task management, email triage, meeting coach, interactive dashboards, independent agents, self-healing memory and skills, team chat), Glean Transform, auto routing and Glean Waldo under the Glean Intelligence umbrella, AI usage controls with spend forecasting and per-department profiles, agent governance policies, context-aware AI threat detection, the AI gateway with portable memories and skills, and Glean Tau. Not one carried a GA, beta, preview, or date label. Record every one of them as availability unstated until your account team says otherwise in writing. The strongest technical content is the token-yield segment and the exfiltration demo. Both describe mechanisms you can verify rather than outcomes you have to trust: precomputed index versus live federated retrieval, and multi-step authorized actions producing an unauthorized outcome. The weakest content is the benchmark framing — 81% cheaper, 78% preferred, 70% fewer tokens, against a single competitor, on a self-built query set, with the methodology caveat delivered once by a different speaker about a different slide. **Practical next steps, in order.** 1. Ask your Glean account team for written availability status and target dates on auto routing, AI gateway, agent governance policies, and context-aware threat detection. Announcement is not shipment. 2. Before department intelligence profiles get set, get your team's task mix on the table. "Cost efficient" is a quality decision, and Eddie's own chart shows coding tasks have a steep cost-performance slope where cheap routing hurts. 3. Run threat detection in observe-only mode for 30 days before enforcement, and compute precision separately for people whose roles legitimately cross customer or business-unit boundaries. 4. Replicate the benchmark on 200 of your own tasks with a blind preference test, three raters, and a published rubric. Report the median and the 90th percentile, not the mean. 5. If memory portability matters to your exit strategy, get the export format and post-termination access into the contract now, while it's the vendor's own headline claim. 6. Steal GM's rollout method regardless of vendor: two populations, existing meetings, and 5D30 as the metric instead of seat count. ## 🔄 Refactor Appendix — Glean:GO 2026 Keynote > Density: 2–5 Literature atoms, 0–3 Permanent candidates. > Emitted here: 5 Literature atoms (override — dense source, 8+ distinct product announcements > plus a benchmark claim), 3 Permanent candidates (override — claims reusable across unrelated > projects), 7 Fleeting items, 2 Things 3 actions. > > Konik claim-header rules: present tense, declarative, ≤10 words, searchable, unique. > Atomicity: one declarative sentence per atom. "and"/"also" → split. > > Tag pool: #bmf · #addendum · #xref · #articleseed · #storystem · #share · #aphorisms ### 📚 Literature Split Pad (H3) #### [H1] Vendor benchmarks state conditions once, then drop them > [!quote] ID:rw_vGmmt7p2Hyk_3721 > So it's important to note you know these are our results on our benchmark. **Context & use**: - [01:02:01](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=3721), Glean:GO 2026 keynote, Eddie on the Pareto frontier chart - The caveat is spoken once, by a different speaker than the one who announced 81%/78%, about a different artifact, and is never reattached to the headline number. The pattern to remember is placement, not concealment — the disclosure existed and still failed. **Tags**: #xref #articleseed <!-- atomicity: 1 idea ✓ --> #### [H2] Precomputed index beats live federated retrieval on tokens > [!quote] ID:rw_vGmmt7p2Hyk_3958 > We actually use 70% fewer tokens. **Context & use**: - [01:05:58](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=3958), Glean:GO 2026 keynote, Eddie explaining the cost decomposition - Mechanism given: a central index and enterprise graph precompute what federated MCP setups rebuild at runtime by searching live across sources of unreconciled quality, then deduping, filtering, and joining while tokens burn. This is a caching argument and it's independently testable. **Tags**: #xref <!-- atomicity: 1 idea ✓ --> #### [H3] Authorized steps chain into unauthorized outcomes > [!quote] ID:rw_vGmmt7p2Hyk_4998 > That's why today I am super excited to announce contextaware AI threat detection. **Context & use**: - [01:23:18](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=4998) announcement, demonstrated at [01:24:14](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=5054) - The demo is the artifact worth keeping: an engineer's assistant pulls two other customers' RCAs into a reply to a third customer. Every permission valid, no malicious intent, prohibited outcome. Reusable as security training material independent of Glean. **Tags**: #xref #share <!-- atomicity: 1 idea ✓ --> #### [H4] Memory is the new enterprise AI lock-in surface > [!quote] ID:rw_vGmmt7p2Hyk_5231 > That's why we are announcing Glean's AI gateway, a unified control plane for LLMs and MCPs **Context & use**: - [01:27:11](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=5231), Glean:GO 2026 keynote, AI gateway announcement - The second half of the announcement matters more than the first: memories and skills become portable across Cursor, Copilot, and ChatGPT because memory built inside one app has become a data silo. No export format is described, so the portability claim is unverified. **Tags**: #xref #articleseed <!-- atomicity: 1 idea ✓ --> #### [H5] Adoption metrics should measure behavior, not seats > [!quote] ID:rw_vGmmt7p2Hyk_2487 > We we sort of look at the typical metrics, but 5D30 is where we anchor. **Context & use**: - [00:41:27](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=2487), GM fireside, Phil Lutka on 60,000 users in 90 days - 5D30 = five active days in thirty, reported above 80% organization-wide. Lutka attributes the result to enablement work (two populations, existing meetings) rather than to product quality, which is the credibility-earning move in the segment. **Tags**: #xref #share <!-- atomicity: 1 idea ✓ --> #### 🧠 Atomic Split Pad (H4) ##### [P1] A disclosed caveat that travels separately from the number is not a disclosure **Claim** (one sentence, my voice): A methodology caveat only functions if it occupies the same sentence as the figure it qualifies. **Evidence pointer**: [[#[H1] Vendor benchmarks state conditions once, then drop them]] **Scope & conditions**: Holds wherever a number will be re-quoted out of its original context — vendor benchmarks, internal dashboards, research abstracts, release notes. Does not hold for numbers that only ever appear inside a full methods document. **Links**: - [[Evidence standards for vendor claims]] — because extends - [[Progressive summarization loses qualifiers]] — because contrasts on where information degrades - [[Writing release notes with availability status]] — because applies-in technical documentation <!-- atomicity: 1 idea ✓ --> ##### [P2] Permission checks per action cannot catch prohibited outcomes across actions **Claim** (one sentence, my voice): Authorization models that evaluate one action at a time miss violations that only exist in the sequence. **Evidence pointer**: [[#[H3] Authorized steps chain into unauthorized outcomes]] **Scope & conditions**: Holds for any system where an agent or user can compose reads and writes across trust boundaries — support tooling, CI bots, repository automation, aggregation dashboards. Does not hold where every action is already scoped to a single boundary with no cross-boundary read possible. **Links**: - [[Least privilege]] — because extends - [[Agent identity and permissions]] — because applies-in AI governance - [[Per-request authorization]] — because contrasts on evaluation granularity **Optional**: `proof_state: working` <!-- atomicity: 1 idea ✓ --> ##### [P3] Cost-efficiency routing is a quality decision wearing a budget label **Claim** (one sentence, my voice): Assigning a cheaper model tier to a team sets that team's output quality ceiling, so it belongs to whoever owns quality, not whoever owns spend. **Evidence pointer**: [[#[H2] Precomputed index beats live federated retrieval on tokens]] **Scope & conditions**: Holds where task difficulty varies within the group being profiled — Glean's own chart shows coding has a steep cost-performance slope while information-seeking is flat. Weakens where a department's task mix is genuinely uniform and simple. **Links**: - [[Cost as a hidden quality lever]] — because extends - [[Error budgets]] — because contrasts on which constraint is made explicit - [[Model selection for enterprise tasks]] — because applies-in AI operations <!-- atomicity: 1 idea ✓ --> ### ✏️ Fleeting Queue - [F1] Verify whether Glean published the token-cost benchmark said on stage to be "released this morning" (2026-09-21) — no URL given in the keynote #fleeting #inbox - [F2] Identify which open-source coding harness underlies Glean Tau; transcript renders it "Cody Harness" once — https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=5328 #fleeting #inbox - [F3] Confirm the Snowflake SVP's name — transcript gives only "VC" — and whether the "16x swing" appears in any written Snowflake material #fleeting #inbox - [F4] Resolve DaVita's deployed population: 17,000 or 70,000? Both appear in the same answer at [01:42:23](https://www.youtube.com/watch?v=vGmmt7p2Hyk&t=6143) #fleeting #inbox - [F5] Open question: does spend-aware downshifting notify the user when it happens, or degrade silently? Not addressed anywhere in the keynote #fleeting #inbox - [F6] Article idea: "The caveat that didn't travel" — using this keynote's 81%/78% as the worked example of disclosure-by-placement failure #fleeting #inbox #articleseed - [F7] Steal the exfiltration demo as a WordPress Playground security lab: individually authorized actions, prohibited aggregate outcome #fleeting #inbox #share ### → Things 3 (max 2 actions) - [ ] Ask Glean account team for written availability status and dates on auto routing, AI gateway, agent governance policies, and context-aware threat detection — Glean:GO 2026 keynote follow-up #things - [ ] Draft the 200-task benchmark replication plan (blind preference test, 3 raters, median and p90 reported) before any auto-routing rollout — work #things <!-- deep-read:sections end --> 📄 **Full text:** [[Full Document Contents/Articles/GleanGO 2026 Keynote.md|Glean:GO 2026 Keynote by Glean]] ## Highlights - [00:40:50] What's the friction that exists in everyone's daily life? How do we, you know, help people find the person that knows how our, you know, battery chemistry works or a 10-speed transmission works? How do we help them find a relevant like current document and all of those sort of scenarios? And we also wanted to bring it to folks in the sense of we had a lot of employees, thousands that had really not used AI before. And we had another huge population that had and was ready to get cooking, building agents and all that fun stuff. And so we we really met people where they were in that journey and just spent a tremendous amount of [00:41:27] energy and calories on enablement training, joining all people meetings, staff meetings and just again like leading with curiosity like what problems can we help you solve? How can we help you get comfortable through that? On top of just the reality that Glean is super easy to plum, we had it integrated into all of our systems. We made it so that like that first experience on day one through day 90 was great. And I'll say too, you know, when we talk about 60,000 users adopted, like we're pretty serious about that. We we sort of look at the typical metrics, but 5D30 is where we anchor. And like I looked this morning, we're still at over ([View Highlight](https://read.readwise.io/read/01m48t7711p6k565h3abjwg8q1)) - Note: [H5] Adoption metrics should measure behavior, not seats - Tags: atom/literature/h3 - [01:01:21] frontier. I thought that uh the first time I heard it. Pareto frontier. So what is this? This uh this this is a model that shows us price performance. So a a a model is Pareto optimal um if nothing else beats it on performance at the same cost or matches its performance for less. So we see cheaper less capable models in the lower left quadrant and we see the more expensive more capable models in the right quadrant. And when we talk about the frontier that's that curve that traces the best option at every cost point in between them. >> Yeah. So this chart you guys are looking at we built on enterprise data capturing [01:02:01] the variety of tasks and different works across departments that are in the enterprise today. So it's important to note you know these are our results on our benchmark. So they involve you know our query set composition. They do convey these models across the frontier but they're conditioned on the choice of harness as well as the evals in the environment. So if you guys look closely in this setting GPT 5.6 6 Luna does dominate a big part of the frontier given its incredible bang for buck. Um in the middle we see an appearance from open models like Kimmy K3 GLM52 and actually as of this morning you know we haven't ([View Highlight](https://read.readwise.io/read/01m48t6xs7rabmknv7e3h6dwbf)) - Note: [H1] Vendor benchmarks state conditions once, then drop them - Tags: atom/literature/h1 - [01:05:20] and visual artifact creation. So for those kinds of tasks we route to it there. Um but it's not just choosing the model right it's also deciding how hard the model should think. So for simpler tasks we might use 56 Luna but on low reasoning rather than high reasoning. Um and to make that all happen at runtime we've postrained our own specialist model Glean Waldo. um that's post- trained on strong open- source base models such as the Neatron family uh and that model's capable of routing to the right uh reasoning level on demand. Yeah. And you know as an engineer as a product builder the part I really love is we didn't we didn't win that [01:05:58] benchmark on on just this this uh choice the price per token alone. We actually use 70% fewer tokens. Can you tell us why? >> Yeah, totally. All these things kind of compound for token efficiency, right? So it comes back to our roots in search. Um with search we have a very purposeful choice to premputee a central index build this enterprise graph that has all these signals and that way the model can much more efficiently find that relevant context at runtime instead of sort of building it up incrementally wasting a lot of loops collecting it. Right? So cloud co-work and other harnesses other setups might take a different approach ([View Highlight](https://read.readwise.io/read/01m48t704nw23saa40m1ds9ez8)) - Note: [H2] Precomputed index beats live federated retrieval on tokens - Tags: atom/literature/h2 - [01:23:18] runtime. And today we are adding agent governance policies, controlling what agents can read, restricting high-risk rights, and setting access controls right at the individual agent level. But permissions and policies can only go so far. Something risky can look safe in isolation. And it's only when you see the full picture you understand what's really going on. That's why today I am super excited to announce contextaware AI threat detection. It spots high-risk agentic behavior across sessions and uses enterprise context to distinguish between legitimate activity and potential attack. Contextaware threat detection is how we fight what is [01:24:14] becoming the largest insider risk facing enterprises AI itself to see this in action. Let's welcome Blair. >> Hey everybody, it's great to be here. Looks like you've got the uh dashboard up. Let's walk through it. >> Yes. So if we take a look at this dashboard, we can see that there are four issues detected around about two days ago. And if we click here, we can take a closer look at these findings. You know, one of these items stands out. We just detected a sensitive data Xfill request. Let's open it and uh follow the activity. Here we see Alex was working on a Zenesk ticket for a customer Atlas Labs. He asked Glean assistant to reconstruct the ([View Highlight](https://read.readwise.io/read/01m48t72dcs3vc733fafjr7m2p)) - Note: [H3] Authorized steps chain into unauthorized outcomes - Tags: atom/literature/h3 - [01:26:22] information crosses into another's boundary. And not only that, but it understands the work that Alex does. And so it knows that he should only be sharing data that's related to Atlas Labs, not any other customer that he doesn't work with. Every step appeared business as usual. The assistant and Alex had valid permissions. There was no malicious intent here. But when we connected these individual actions to the overall context of the situation, we accurately detected a high confidence exfiltration. Thanks for the demo, Blair. [applause] Now, we've talked a lot about how the AI landscape has changed with more models, [01:27:11] more front doors, and employees reaching for their favorite AI apps. That's why we are announcing Glean's AI gateway, a unified control plane for LLMs and MCPs that secures and governs models, context, skills, memories across every front door. With it, your apps also inherit the efficiencies and cost controls that Tony walked through earlier. And today, we are announcing that AI gateway also makes memories and skills portable across apps. Glean is aggregating memory from across all your contexts, chats, agents, other AI apps to make it sharable and usable everywhere you work. Now, that's a big deal because right ([View Highlight](https://read.readwise.io/read/01m48t74payn3wgqfxth1gr7ww)) - Note: [H4] Memory is the new enterprise AI lock-in surface - Tags: atom/literature/h3