The AI Proposal That Lost Us a Client - CourtneyR.dev's 2nd Brain

The AI Proposal That Lost Us a Client ![rw-book-cover](https://d34adp677peecb.cloudfront.net/static/images/article1.be68295a7e40.png) 🔗 **Source:** [The AI Proposal That Lost Us a Client](https://www.youtube.com/watch?v=RJ3cyazAycU) > [!note]+ Document Notes > tags: [ai-proposals, client-trust, fractional-cto, data-governance, small-business-it] > author: Duncan Isaaksen-Loxton > source: Agency Mavericks (YouTube) > reference: https://www.youtube.com/watch?v=RJ3cyazAycU > date created: 2026-09-29 10:30 > > > [!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 Do not quote or attribute any other line to the named author.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: The AI Proposal That Lost Us a Client (with Duncan Isaaksen-Loxton) >  ## 🧭 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--the-ai-proposal-that-lost-us-a-client-with-duncan-isaaksen-loxton) ## 📝 Summary - **What happened.** Agency Mavericks (host Troy Dean) published a 50:13 episode of The Agency Hour on 2026-09-29 with Duncan Isaaksen-Loxton, who runs sixfive.io, a fractional CTO firm for businesses of 1 to 50 seats. The auto-transcript renders the company as "65.io" and "65," the guest as "Duncan Isaacson Lockxton," and the AI assistant the host uses as "Victor" (the description spells it Viktor). Quotes below keep the transcript spelling. - **The claim.** AI cut Duncan's research on eight client websites from about two days to 20 to 30 minutes, and a proposal took "a couple hours." The client still rejected the 13-page result because it "sounds like you did this proposal with AI" and read as disconnected from the discovery call. Duncan says that cost him the job (00:43:15 to 00:46:09). - **The supporting argument.** Two thirds of the episode builds the setup: owners DIY their tech, staff paste business data into personal ChatGPT accounts, and AI "accelerating mess" when foundations (three core tools, offboarding, secrets hygiene, backups) are missing. The proposal story closes the loop. - **Why it matters.** The failure wasn't accuracy. A human read it, tweaked it and sent it under a stated policy, yet the voice mismatch against the call was enough for the buyer to withdraw trust. Review-before-send isn't a sufficient control for client-facing prose. - **Key stakeholders.** Duncan and sixfive.io, Troy Dean and Agency Mavericks (also a Viktor promoter and Duncan's client), Duncan's small-business clients, the unnamed proposal client, and staff and business owners liable for data leaks. - **Risks and tradeoffs.** One lost deal is a single anecdote with a client-stated cause, not a measured effect. Both speakers sell services next to the advice: Duncan sells tech management, Troy promotes Viktor. Troy discloses the client relationship at 00:09:20. Several figures are unsourced: the 30-day Drive recovery window, an Australian Medicare leak "in the past few days," and a September 2026 claim about Google and AI content. - **Bottom line.** The durable lesson is procedural: use AI for research and drafting, keep the human voice and the relationship in the customer-facing layer, and fix foundations before automating. The evidence is one episode of practitioner testimony, useful and unmeasured. ## 📖 Main content **Thread one: foundations before AI.** Duncan frames sixfive.io's "business operating system" as a path from start-up tool choices to compliance, privacy and scale, then to an "optimized" phase where AI fits (00:03:37 to 00:05:02). His stated rule: AI without foundations only accelerates mess. He argues most businesses run on three core tools, plus an industry-specific tool where one exists (00:06:27). **Thread two: DIY tech and its costs.** Troy asks why owners superhero everything themselves. Duncan answers with survival economics: an early owner is the cleaner, salesperson and proposal writer at once, so a quick fix that works today stays forever (00:10:46 to 00:12:14). His example is a free @gmail.com address, which has different privacy terms from paid Google Workspace or Microsoft 365, and staff accounts like "aaron's plumbing04." **Thread three: AI data exposure.** A live client has two staff members putting business data into personal ChatGPT accounts. Duncan's point is liability: the data leaves the business and the owner stays responsible for it (00:12:57). Troy adds a secondhand story of a hospital supervisor giving a government RFP to ChatGPT, which would breach the contract terms (00:16:35), and Troy's own memory of a WooThemes card breach (00:14:22, dated by Troy to roughly 2014 or 2015 and unverified). **Thread four: selling protection without fear.** Duncan concedes it "does feel more like alarmist protectionism sometimes" and reframes protection as self-care and image: a free Gmail address signals you don't look after yourself (00:18:44). He backs up backups with a claim that Google Drive gives about a 30-day window and recommends at least a year of retention (00:20:11). This figure is his statement; check Google's admin documentation before repeating it. **Thread five: secrets and tool consolidation.** Troy explains the Slack Connect exposure: anyone added to a shared channel, even for 3 weeks, saw every key or password shared there. Troy's practice now is direct messages with Viktor and storing secrets in Doppler (00:21:35 to 00:22:18). Duncan argues for one identity provider (Google Workspace, Google Chat, Google Meet) because it shrinks the surface to back up, audit, onboard and offboard (00:23:43 to 00:25:11), with Zoom kept only for webinars. **Thread six: where websites and content fit.** Duncan sorts website people into techies, builders and designers, says sixfive.io is not a marketing agency, and says it supports the tech underneath (00:29:31 to 00:31:36). On AI content, he warns that a blog post made in 20 minutes has no expertise or context, and claims Google is noticing (unsourced). Troy uses the power-drill analogy: AI hands everyone a drill, but the tradesperson knows to find the stud first (00:33:48 to 00:34:30). **Thread seven: boundaries with AI-written client documents.** Troy answers clients who send 30-page AI-written business documents with a policy of replying with a 350-page Claude response they must read before engagement continues, as a deliberate provocation (00:35:14). Duncan takes a gentler route: he encourages experimentation, asks for screen-recorded talk-throughs or a conversation instead, and warns that automating everything erodes the human connection that builds trust (00:37:30 to 00:39:40). **Thread eight: the proposal.** The story runs 00:43:15 to 00:46:09. Discovery call, then AI-assisted research on eight sites, then a proposal blended from the call transcript and the rate card. Duncan's policy is that a human reads, tweaks and sends; nothing goes out automatically. He was tired, missed the "dramatic prose" AI adds, and the voice profile did not catch it. The client wrote back that it sounded AI-made and disconnected from the call. Duncan calls it a battle scar. **Contested or unverifiable.** The causal claim (lost job because of the proposal) rests on the client's stated reason, which is reasonable but single-source. The three-tools claim, the Drive window, the Google claim and the Medicare leak have no source in the transcript. The description says "The research that used to take two days took 30 minutes"; the transcript says "couple days of work" versus "20 minutes, half an hour." ## 🗒️ Notes #### ⏱️ [00:00:00](https://www.youtube.com/watch?v=RJ3cyazAycU&t=0) Cold open - The episode opens on the client's feedback before any introduction. - The client contrasts the proposal with the discovery call: they left the call feeling understood. > [!quote] [00:00:00] > It sounds like you did this proposal with AI. #### ⏱️ [00:00:25](https://www.youtube.com/watch?v=RJ3cyazAycU&t=25) Intro - Troy Dean announces a new podcast format and cadence in the coming weeks, a fan-mail voice-message feature, and asks for email at the address in the description. - Troy introduces Duncan as a Maverick's Club member of about 3 years running a fractional CTO agency, and says the conversation leans on AI. #### ⏱️ [00:02:58](https://www.youtube.com/watch?v=RJ3cyazAycU&t=178) Who is Duncan and what sixfive.io does - Fractional CTO defined as a monthly slice of time, like an accountant, to fine-tune and move you forward. - A "business operating system" runs from start-up tool choices through compliance, privacy and scale to an optimized phase aligned with AI adoption. - Ideal client: entrepreneurs and small businesses with 1 to 50 seats who can't yet hire a full-time technologist. > [!quote] [00:05:02] > AI without having the foundation in place first cuz otherwise all you're doing is you're accelerating mess. #### ⏱️ [00:07:11](https://www.youtube.com/watch?v=RJ3cyazAycU&t=431) From web developer to fractional CTO - Duncan started in a web development agency while at university, at the beginning of the internet age. - His scope grew from websites to email to what looks like small-business IT support. - Framing: get the foundation right and AI removes drudgery work that businesses used to price by the hour. #### ⏱️ [00:09:33](https://www.youtube.com/watch?v=RJ3cyazAycU&t=573) Why business owners DIY their tech - Troy discloses that the two are each other's clients, then tells the DKIM-fix story that started the conversation. - The half-hour myth: a quick fix becomes an 11 o'clock Tuesday night mess you can't unscramble. - Duncan's cause: an early owner has a "bajillion problems," so a fix only needs to work right now. > [!quote] [00:10:03] > I think the myth is that this will only take half an hour to fix this thing, right? #### ⏱️ [00:12:45](https://www.youtube.com/watch?v=RJ3cyazAycU&t=765) Staff putting business data into personal ChatGPT - Two staff at a current client use personal ChatGPT accounts with business data. - Free Gmail versus paid Workspace or 365 carries different privacy terms. > [!quote] [00:12:57] > and you as the business owner are responsible for that content and for the keeping of it. #### ⏱️ [00:14:14](https://www.youtube.com/watch?v=RJ3cyazAycU&t=854) Data breaches and how AI amplifies bad decisions - Troy recalls WooThemes, with about 35 employees, hacked after storing customer cards in its own database instead of using a payment processor (Troy's recollection, roughly 2014 to 2015). - The founders made a well-meant decision that exposed customers, which Troy uses to argue for slowing down. - A friend's hospital supervisor reportedly gave a government RFP to ChatGPT, breaching the contract terms. Secondhand, unnamed. > [!quote] [00:15:51] > AI amplifies everything. #### ⏱️ [00:17:00](https://www.youtube.com/watch?v=RJ3cyazAycU&t=1020) Selling protection without the fear - Troy compares accessibility and backups: nobody buys them until they're sued or lose data. - Duncan admits it can feel like alarmist protectionism, then frames protection as care for the business and how clients see you. - Troy tells a story of being pulled into Visa and Mastercard over a customer's breach (00:19:28). - Backup pitch: Drive has about a 30-day recovery window per Duncan, so he recommends at least a year of retention. #### ⏱️ [00:21:11](https://www.youtube.com/watch?v=RJ3cyazAycU&t=1271) Slack Connect, API keys and secrets hygiene - A removed contractor still saw 3 weeks of a Slack Connect channel, including any shared keys or passwords. - Troy's fix: direct messages with the AI assistant only, secrets stored in Doppler ("one password for AI agents"). - Google Drive sharing: think about view, comment, edit and owner permissions before you share. > [!quote] [00:22:18] > I only share things in a direct message with Victor because no one else has access to that conversation with Victor. #### ⏱️ [00:23:17](https://www.youtube.com/watch?v=RJ3cyazAycU&t=1397) Onboarding, offboarding and the three-tool stack - Rule: one Google Workspace identity for every login, and no new product if Workspace already covers the job. - Consolidation shrinks backup, audit, onboarding and offboarding work; turning off one account collapses the "sign in with Google" sprawl. - Offboarding still needs decisions: where the data migrates, who inherits it, and where the email redirects. > [!quote] [00:24:28] > There's less for you to need to back up. There's less for you to need to security audit. #### ⏱️ [00:26:09](https://www.youtube.com/watch?v=RJ3cyazAycU&t=1569) Google Workspace vs Microsoft - Duncan prefers Meet over Zoom, with Zoom only for webinars; Gemini transcripts land in Drive. - Adding Zoom pulls in a bridge tool (Fathom or similar) and turns one ecosystem into about four. - sixfive.io is a Google partner; for Microsoft 365 shops it doesn't support the product but checks email configuration and deliverability. #### ⏱️ [00:29:31](https://www.youtube.com/watch?v=RJ3cyazAycU&t=1771) Techies, builders and designers: where websites fit - Duncan sorts website people into techies, builders and designers, and says sixfive.io is not a marketing agency. - Google and Microsoft sell you an account and stop; domain, DKIM and DMARC setup is on you. - Scope: a well-structured, fast, AI-readable website underneath whatever the marketing partner builds. No SEO, ads or content services. #### ⏱️ [00:31:40](https://www.youtube.com/watch?v=RJ3cyazAycU&t=1900) AI content and the power-drill problem - AI makes a blog post in 20 minutes but with no expertise or context; Duncan asserts Google is noticing (unsourced). - Human in the loop: extract the IP, generate in context, review, get client sign-off, then publish. - Duncan calls his own marketing "utter crap" for lack of an expert. > [!quote] [00:34:30] > AI is like the power drill, right? But the tradey who's who's done this before and has put a few holes in walls - A support ticket example: a client pasted ChatGPT code into a snippet plugin, it failed, and Duncan's point was that they already pay him for this. #### ⏱️ [00:35:09](https://www.youtube.com/watch?v=RJ3cyazAycU&t=2109) The 30-page AI document policy - Troy's provocation policy for AI-written client documents. - Why Troy objects: no time to read it, Troy will summarize it with AI anyway, and it lacks a goal, strategy or next tactics. > [!quote] [00:35:14] > my response to them is a 350 page response written by Claude that they have to read word for word before our engagement continues. - Duncan's approach is softer: encourage AI experiments, know your limits, request screen-recorded talk-throughs or a conversation. #### ⏱️ [00:39:40](https://www.youtube.com/watch?v=RJ3cyazAycU&t=2380) Human connection can't be automated - Troy tells of Duncan leaning in to ask how Troy was before recording, a moment Troy says nobody gives a business owner. - Argument: AI takes the computer work, freeing time for relationship work. - The job of an agency owner is deciding what to do next, and battle scars let you veto AI's ideas that already failed. > [!quote] [00:41:05] > You can't replicate that with any kind of technology. #### ⏱️ [00:43:15](https://www.youtube.com/watch?v=RJ3cyazAycU&t=2595) The AI proposal that lost a job - Setup: eight websites, a discovery call, AI research in 20 to 30 minutes instead of a couple of days. - Policy: generate with AI, read it, tweak it, make it your own, send it, and never let AI send. > [!quote] [00:44:43] > if you generate it with AI you read it you tweak it you make it your own you send it - Failure: Duncan was tired, missed AI's dramatic prose, and his voice-profile checker didn't catch it. - The client wrote back that it read like AI and nothing like the call. > [!quote] [00:46:09] > It didn't work and we lost a job because of it. #### ⏱️ [00:46:40](https://www.youtube.com/watch?v=RJ3cyazAycU&t=2800) How to reach Duncan - Best route is the sixfive.io site, which offers a 15-minute booking link; he also refers Microsoft-only prospects to partners. - LinkedIn is the second route. Troy promises links in the show notes and jokes that the name comes from Duncan's height, 6'5". #### ⏱️ [00:49:08](https://www.youtube.com/watch?v=RJ3cyazAycU&t=2948) Outro - Troy thanks Duncan for pushing Troy on Viktor as a product and service. - Troy points to theagencyhour.com, a YouTube playlist of the back catalog (Troy says about 175 episodes), and a coming fan-mail button. ## 📚 References The transcript itself contains no links; the following come from the video description or are primary sources for named entities. - [The AI Proposal That Lost Us a Client (with Duncan Isaaksen-Loxton)](https://www.youtube.com/watch?v=RJ3cyazAycU) — the source episode, 50:13, uploaded 2026-09-29 by Agency Mavericks. - [sixfive.io](https://sixfive.io/) — Duncan's fractional CTO firm; Duncan points listeners here for the 15-minute booking link at 00:46:40. - [Duncan Isaaksen-Loxton on LinkedIn](https://www.linkedin.com/in/loxton/) — the second contact route Duncan names at 00:47:35. - [Doppler](https://doppler.com) — the secrets manager Troy says Troy uses at 00:22:18; listed under "Mentioned in this episode." - [Viktor](https://mvrx.link/get-viktor) — the AI assistant Troy uses in Slack; the transcript spells it "Victor." The link is a tracked short link from the description, so treat it as promotional. - [The Agency Hour](https://theagencyhour.com) — the podcast home Troy points to at 00:49:01. - [Agency Mavericks](https://agencymavericks.com) — the publisher and Troy Dean's business; the transcript renders the email domain as "aggencymavericks" and "agencymicks," while the description gives [email protected]. - Google Workspace, Microsoft 365, Zoom, Slack Connect, Fathom, Cloudflare, Figma — products named in passing; consult vendor docs for the retention, sharing and offboarding behavior discussed. - DKIM and DMARC — the email-authentication standards behind the deliverability problem Duncan says clients hit (00:30:52); the transcript renders DMARC as "demark." ## 🗣️ Paraphrase Duncan runs tech for small businesses and argues that most trouble comes from owners fixing things themselves in the middle of a busy week. A free Gmail address, a personal ChatGPT login, a password pasted into a shared Slack channel: each solves today's problem and creates a liability that shows up months later. He wants foundations first, which for him means about three core tools, one identity for every login, backups with a year of retention, and a checklist for when people join or leave. AI comes after that, because AI on a messy base speeds the mess up. Troy adds war stories about breaches and data exposure, then moves to a cultural point: AI is a power drill, and knowing where the stud is takes experience. Both hosts want people to explore AI but keep a human in the loop and recognize where their expertise ends. The closing story tests that advice against Duncan's own practice. He used AI to compress research on eight websites from days to half an hour and drafted a 13-page proposal in hours, then read and edited it as his policy requires. The client still noticed the AI voice, contrasted it with the human connection from the call, and walked away. Duncan treats the loss as a battle scar. The episode's implicit finding is that a human reviewer who is tired can pass a document that a customer will fail. ## 📰 Newsletter **The proposal was accurate and it still cost the job.** In this Agency Mavericks episode, Duncan Isaaksen-Loxton of sixfive.io tells how an AI-assisted, 13-page proposal for eight websites lost him a client. The research took 20 to 30 minutes instead of a couple of days, and he read and edited the draft before sending. The client wrote back that it read like AI and sounded nothing like the discovery call. Why it matters to WordPress shops: agency and freelance work runs on trust built in the first conversation. A proposal, audit summary or support reply that shifts voice can undo it. Duncan's rule of thumb is to keep AI in research and drafting and keep people in the moments that carry the relationship. The rest of the episode is a practical checklist for small-site owners. Use one identity provider, not a personal Gmail per staff member. Don't paste keys into Slack Connect channels (a removed contractor still saw 3 weeks of history). Keep at least a year of backups, since Duncan says Google Drive's window is about 30 days. Check the claims against vendor docs before you repeat them. Links: [the episode](https://www.youtube.com/watch?v=RJ3cyazAycU), [sixfive.io](https://sixfive.io/), [Doppler](https://doppler.com). ## ✅ Action items #### DevRel - Developer advocacy: write a short talk on "AI drafts, humans own the voice," using the proposal story (00:43:15) as the case and publishing the review checklist alongside it. - Education: run a workshop where participants draft a proposal with AI, then read it aloud next to a call transcript to spot voice drift, the failure Duncan describes at 00:45:26. - Developer experience: add a "send-time check" to any AI-assisted docs or reply workflow: compare the draft against a real prior message from the same author before it goes out. - Developer marketing: don't publish AI-drafted outreach without a named human edit pass; the episode's client caught one at first read. - Community management: state a norm for AI-written posts in community channels (disclose, keep it short), echoing the 30-page document complaint at 00:35:14. - Internal/external/cross-project: no direct action from this source; nearest relevance: agree on who owns tone review when a partner and a vendor co-write material. #### OSPO - Policy and governance: add a rule that contributors and staff use organization-managed AI accounts, not personal ones, for any project or business data (00:12:57). - License compliance: no direct action from this source; nearest relevance: check the terms of any AI account before contract or customer text goes into it. - Security management: adopt a secrets standard: no keys or passwords in shared chat, including cross-organization channels (00:21:35), with a secrets manager as the alternative. - Software inventory: list AI tools in use, including personal accounts staff have reported, so data flows are known. - Developer support: publish a one-page offboarding checklist covering account shutdown, data inheritance and email redirect (00:24:28). - Community engagement: no direct action from this source; nearest relevance: ask sponsors and partners which AI tools touch shared data. - Strategic alignment: track two figures the episode implies: hours saved on research and win rate on AI-assisted proposals, measured against the human-only baseline. #### Community - Maintainers: no direct action from this source; nearest relevance: collaborators removed from a shared channel keep what they saw, so rotate any credentials shared there. - Contributors: no direct action from this source; nearest relevance: a note on how to request a call or recorded walkthrough instead of sending a long AI-written document. - Sponsors: no direct action from this source; nearest relevance: sponsor reports should carry a human summary, not raw AI output. - Ecosystem: no direct action from this source; nearest relevance: share a common template for reporting data exposure caused by staff use of personal AI accounts. #### WordPress - Core: no direct action from this source; nearest relevance: none beyond general security hygiene. - Core Performance: no direct action from this source; nearest relevance: none. - Core AI: document that AI-generated content needs a human review step and a voice check before it reaches users, using the proposal failure as a worked example. - **Core Program Team:** no direct action from this source; nearest relevance: coordinate one policy on personal AI accounts across teams. - Mobile: no direct action from this source; nearest relevance: none. - Playground: no direct action from this source; nearest relevance: none. - Community: no direct action from this source; nearest relevance: meetup talk on staying human while using AI. - Training: add a lesson on keeping a personal voice when using AI drafting tools and on reading drafts against a source recording. - Marketing: no direct action from this source; nearest relevance: flag AI drafts for a human tone pass before they publish. - Support: check AI-suggested code snippets for safety before customers paste them into snippet plugins, as in the ticket at 00:34:30. - Design: no direct action from this source; nearest relevance: none. - Documentation: add a note that AI-generated code snippets should be tested in staging, not pasted into a production snippet plugin. - Photos: no direct action from this source; nearest relevance: none. - TV: no direct action from this source; nearest relevance: none. - Polyglots: no direct action from this source; nearest relevance: none. - Accessibility: no direct action from this source; nearest relevance: the host mentions accessibility lawsuits as an example of a risk nobody buys protection for until it lands (00:17:19). - Meta: no direct action from this source; nearest relevance: ensure .org accounts use organizational identity, not personal addresses, where roles need it. - CLI: no direct action from this source; nearest relevance: none. - Hosting: add a cross-host checklist item for email authentication (SPF, DKIM, DMARC), the deliverability problem Duncan describes at 00:30:52. - Tide: no direct action from this source; nearest relevance: none. - Openverse: no direct action from this source; nearest relevance: none. - Themes: no direct action from this source; nearest relevance: none. - Plugins: no direct action from this source; nearest relevance: none. #### The WPOCC - Programs: no direct action from this source; nearest relevance: The WP Open Community Collective could offer small-business owners a short guide on AI-account hygiene. - Projects: no direct action from this source; nearest relevance: none. - Funding experiments: no direct action from this source; nearest relevance: none. ### FAIR - Technical independence: no direct action from this source; nearest relevance: prefer a single organizational identity for tooling to reduce dependencies. - Policies and covenants: no direct action from this source; nearest relevance: a covenant clause on how contributors may use AI with project data. - Productized governance artifacts: no direct action from this source; nearest relevance: an onboarding and offboarding checklist as a governance artifact. ## 🧪 Hypotheses 1. AI-drafted proposals get lower win rates than human-drafted ones when the buyer had a prior live conversation with the seller, because voice drift against the call breaks trust. 2. A voice-profile check catches under half of the stylistic tells (dramatic prose, over-elaboration) that a client notices on first read. 3. Consolidating small businesses to one identity provider cuts offboarding time and orphaned-account counts by at least a third. 4. Staff who use personal AI accounts for business data outnumber staff who use business-managed ones in companies of 1 to 50 seats where no AI policy exists. 5. Sending a short recorded walkthrough in place of a long AI-written document raises the rate of a follow-up meeting within a week. 6. Time saved on research is spent on strategy only when a policy names where the freed hours go; otherwise it goes to volume. ## 📊 Methodology - **H1 and H5.** Run a paired experiment across a set of proposals: for matched prospects, send some drafted with AI and edited by the owner and others written from scratch, blind to the buyer. Track win rate and log any buyer comment that mentions tone. Pre-register the definition of "AI-drafted." Duncan's single case can't support this alone. - **H2.** Collect 30 AI drafts run through a voice-profile checker, then have five reviewers who know the author mark tells. Compare the checker's flags to reviewer flags, reporting recall. - **H3.** Pick 10 firms before and after consolidation; measure minutes per offboarding, count of accounts found after departure, and SaaS seat count. - **H4.** Survey 100 staff across firms of 1 to 50 seats, asking which accounts they use for work AI. Cross-check with expense and SSO logs where available to reduce self-report bias. - **H6.** Track how five agencies allocate hours saved from AI research over 8 weeks, comparing agencies with and without a written allocation policy. - **Controls throughout.** Record the model used, the prompt, the human edit time and the buyer's prior contact with the author. ## ❓ Questions - Beyond the client's remark, what specific passages read as AI, and would a line-edit have fixed them or was the tone the whole problem? - What did the voice-profile check flag on that proposal, if anything, and why did it miss the dramatic prose? - How often does the "read, tweak, send" policy fail when the reviewer is tired, and what forcing step would help? - Does disclosing AI use up front change how buyers read the document? - Is the roughly 30-day Google Drive recovery window accurate for the plan tiers Duncan's clients use? - Which of the "three tools" does he count, and how does the count change for an ecommerce or regulated business? - What share of small-business data exposure comes from staff accounts versus misconfigured sharing, given the examples in the episode? - What remedies exist when a client's staff already put data into a personal ChatGPT account? ## 🔮 Future research - **Voice drift as a trust signal.** No study cited here measures how much stylistic mismatch between a live call and a written follow-up moves buyer trust. A controlled test in service sales would replace one anecdote with data. - **Automation of research versus writing.** The episode treats research (2 days to 30 minutes) and drafting differently in effect. Separate measurement would show whether the win came from time saved or the loss came from tone. - **Personal-account AI exposure rates.** The two-staff example is one client. A prevalence estimate for small firms would size the problem. - **Cost of DIY tech.** The half-hour myth is asserted, not quantified. Track hours spent by owners on self-fixed tech problems against the fee for managed service. - **Search engines and AI content.** The claim that Google is penalizing AI-generated content needs a source and a date. ## 💡 Implications - **Practice.** Add a call-to-document check to client-facing writing: read the draft against the call recording or transcript before sending, and rewrite any section that doesn't sound like the speaker. - **Policy.** Businesses need a written rule on personal AI accounts and business data. The liability stays with the owner regardless of which account the staff member used (00:12:57). - **Security.** Shared channels that cross organizations leak history to anyone added later, so secrets belong in a manager, not in chat (00:21:35 to 00:22:18). - **Education.** The power-drill analogy (00:34:30) is a usable teaching device for why expertise still matters, and the proposal loss is a cheap case study for classrooms. - **Funding.** Owners who can't yet afford an expert use AI as a stopgap (00:38:13); funders of small-business programs could target foundation setup before AI adoption. - **Sales.** Time saved on research doesn't transfer to relationship trust. Use it to add human touch, not to reduce it. ## 🧠 Learning styles assessment - **Reading level estimate.** Flesch-Kincaid Grade Level about 5.6 and Flesch Reading Ease about 81.8, computed by script on the timestamp-stripped transcript (9,795 words, 651 sentence-ending marks). Spoken conversational English scores lower than written prose on both formulas, so treat it as a listening-level estimate. Domain terms (DKIM, fractional CTO, Slack Connect, API keys) raise the comprehension load for non-technical listeners. #### 👀 Visual learners - Present: none identifiable from the transcript; the Riverside recording is likely a two-person video call. - Improve: add a one-slide diagram of the "foundation, operations, optimized" path at 00:03:37 and an on-screen timeline of the proposal story. #### 🔊 Aural learners - Present: the whole episode is conversational audio with an introduction, chapters and an outro recap of where to find things. - Improve: add a 90-second audio summary of the five rules (one identity, no secrets in chat, backups, offboarding, human review) for people who won't listen for 50 minutes. #### 🗣️ Verbal learners - Present: analogies (power drill, accountant, superhero), stories (WooThemes, the hospital, the proposal), and plain definitions of a fractional CTO. - Improve: define DKIM, DMARC and Slack Connect in one sentence each on first use. #### 🧩 Kinesthetic learners - Present: concrete practices such as moving secrets into a manager, a screen-recorded walkthrough in place of a document, and a form for offboarding. - Improve: link a checklist or template for offboarding and for the proposal read-back so listeners can do it the same day. #### 🧠 Logical learners - Present: a clear chain: foundations, then AI, or mess accelerates; fewer tools mean less to back up, audit, onboard and offboard (00:24:28). - Improve: quantify the claims (30-day window, three tools, half-hour myth) with sources so the chain can be checked. #### 👥 Social learners - Present: two hosts in dialogue, with disclosed relationships, plus an invitation to email the host and a coming fan-mail feature. - Improve: add a prompt for listeners to share their own AI-proposal experiences in the community. #### 🧘 Solitary learners - Present: chapters let a listener revisit a single topic (17 chapters). - Improve: add a reflection question at the end, such as "Which document have you sent this month that you'd fail to recognize as your own voice?" #### 🌿 Naturalistic learners - Present: minimal; the tradesperson and stud analogy draws on physical work. - Improve: no strong connection to natural systems; a note on the energy cost of AI use would be a stretch and isn't supported by the source. ## 📊 Accessibility assessment #### Screen reader semantics - Present: 17 chapters with titles (published in the description and as chapter markers) give a navigable structure. - Improve: the description carries bullets and links as plain text; make sure link text is descriptive rather than raw URLs. Maps to WCAG 2.2 AA 2.4.6 (Headings and Labels) and 2.4.4 (Link Purpose). #### Keyboard navigation - Present: the player and chapter list on YouTube are keyboard operable. - Improve: no content-side change needed beyond the chapters; the platform controls 2.1.1 and 2.4.7. #### Contrast and readability - Present: unverifiable from a transcript. - Improve: if the video shows on-screen text or a lower-third, verify 4.5:1 contrast (1.4.3). #### Text resizing and reflow - Present: platform captions can be resized by the viewer. - Improve: any burned-in text can't reflow (1.4.4, 1.4.10); avoid it or repeat it in the description. #### Media and captions - Present: auto-captions exist and chapters break the audio up. - Improve: the auto-captions garble names and terms: "65.io" for sixfive.io, "Isaacson Lockxton" for Isaaksen-Loxton, "Victor" for Viktor, "aggencymavericks" for agencymavericks, "demark" for DMARC, and a final "Thanks, Mark" that appears to address the host. Human-corrected captions would fix 1.2.2 (Captions, Prerecorded). No audio description is needed for a talk, but 1.2.3 applies if visuals carry information. #### Forms and interactions - Present: the description links a fan-mail form and gives an email address as an alternative. - Improve: keep a text alternative to the voice-message feature, and make sure the form has labels and error handling (3.3.2, 3.3.1). The email route already covers this. #### Standards reference - Main gaps against WCAG 2.2 AA: 1.2.2 (inaccurate captions on proper nouns and product names), 2.4.4 (bare URLs in the description) and 1.4.4 or 1.4.10 if burned-in text exists. The caption fix has the most impact for search, since the guest's name is misspelled throughout. ## 🌍 DEIB assessment #### Cultural breadth - Present: speakers in Spain and Australia, references to Europe (WordCamp Europe), the US (accessibility lawsuits) and Australia (Medicare, Melbourne). - Improve: examples are mostly English-speaking, Western small businesses; add a non-English-language or non-Google-ecosystem case, since the advice assumes Google Workspace. #### Inclusive language - Present: plain, direct speech and few unexplained acronyms. - Improve: "some idiot invented this thing called AI" (00:07:53) and "every dad thought he was a trady" (00:34:30) use an insult and a gendered stereotype; "tradies" is regional slang. Swap for neutral phrasing. #### Neurodiversity support - Present: 17 chapters and a stable structure of story, point and takeaway support scanning. - Improve: the episode wanders (Teams rant, Riverside audio problem, height joke); a short recap at the end of each chapter would help listeners who lose the thread. #### Representation - Present: two male voices and one unnamed female client (the hospital friend) in anecdotes. - Improve: no guest from a non-Google or larger-scale environment, and no client voice. The proposal client speaks only through Duncan's retelling. #### Psychological safety - Present: Duncan admits his own failure ("fair cop") and calls it a battle scar; Troy frames the host's own DIY habits with humor. - Improve: the 350-page reply policy reads as punitive toward clients; a note that it's a provocation, not a template, would lower the risk that listeners copy it. ## ✅ Logical validity #### Strong reasoning - **Causal mechanism, 00:12:57.** Personal accounts move data out of business control, and the owner stays responsible for it. The chain (account, data, liability) is complete and testable. - **Mechanism-based argument, 00:24:28.** One identity provider means fewer tools to back up, audit, onboard and offboard. Each item follows from having fewer systems. - **Exposure model, 00:21:35.** A removed contractor keeps 3 weeks of shared history including keys, which is checkable against how Slack Connect works. - **Analogy, 00:34:30.** The power-drill comparison fits the claim that tools don't supply the judgment to find the stud, and the support-ticket example (ChatGPT code pasted into a snippet plugin) is a concrete instance. - **Case evidence, 00:45:26 to 00:46:09.** The proposal story carries a client-stated reason, a policy that was followed, and a self-critical read-back, which raises its credibility above an assertion. - **Self-limiting statements.** Duncan says his own marketing is "utter crap," and Troy says AI "has problems." Both concessions add credibility. #### Fallacies - **Hasty generalization, 00:46:09.** "It didn't work and we lost a job because of it." One client, one proposal, one reviewer who was tired. *Why weak:* it can't show AI proposals generally lose trust; it shows this one did. *Fix:* frame it as a case and collect a sample before setting policy. - **Appeal to fear, 00:17:19 and 00:18:02.** Accessibility suits and lost backups are used to motivate spending; Duncan grants it can feel like "alarmist protectionism" at 00:18:44. *Why weak:* fear without probabilities inflates risk. *Fix:* give incident rates or cost ranges for each risk. - **Unsupported assertion, 00:32:20.** "Google are starting to realize that the engine is just feeding itself with crap." No source, metric or date. *Fix:* cite the Google guidance or a study. - **Sweeping generalization, 00:06:27.** "almost every business can operate with three bits of technology." No survey or definition of "bit." *Fix:* name the three and state which businesses fit. - **Universal claim, 00:22:18.** "We all have" shared a key in Slack. Plausible but unmeasured and used to cover every listener. *Fix:* say "most teams I see" and cite a count if one exists. - **Anecdote as evidence, 00:16:35.** A hospital supervisor giving a government RFP to ChatGPT is secondhand, unnamed and unverifiable. *Fix:* label it hearsay or drop it. - **Ad hominem framing, 00:07:53.** "some idiot invented this thing called AI" is a joke, but it dismisses the field's makers instead of the argument. *Fix:* state the concern (confusion, hype) directly. - **Commercial interest, throughout.** Duncan sells tech management, so advice to stop DIY favors him, and Troy promotes Viktor (linked in the description). Troy discloses the client relationship at 00:09:20. *Fix:* a one-line note that both speakers benefit from the recommendations. #### Tone of voice - Conversational, candid and self-deprecating, with a mentoring register between two peers. Duncan is measured and framework-driven ("your foundations is that first layer"), and Troy is expressive, using profanity-adjacent bluntness ("It's freaking horrible") and strong opinions on Teams. Both use humor to soften warnings. The tone shifts to sober at 00:45:26 when Duncan reads the client's feedback and says "Oh shit" and "fair cop," and it stays plain through the loss. ## 🛡️ DARVO analysis Not applicable to this source. DARVO (Deny, Attack, Reverse Victim and Offender) describes a defensive reaction to accusation. Here nobody is accused; the closest thing to criticism is the client's feedback on the proposal, which Duncan accepts. - **Deny:** none. Duncan agrees with the client ("fair cop") at 00:45:26 and takes ownership of the miss. - **Attack:** none. He thanks the client for the feedback. The only target of sharp language is Microsoft Teams (00:27:22 to 00:28:04) and, mildly, clients who send long AI documents (00:35:14), neither of which is a response to an accusation. - **Reverse victim and offender:** none. Duncan doesn't cast himself as the wronged party; he frames the loss as a lesson. ## 🎯 Conclusion The episode's central claim is that AI works when the foundations exist and a human stays in the relationship layer. The best evidence is Duncan's own proposal: research compressed by more than an order of magnitude, a human review under a written policy, and still a lost client because the voice no longer matched the person from the call. The supporting advice (one identity, no secrets in chat, backups, offboarding, no personal AI accounts for business data) follows from clear mechanisms, though several figures (the 30-day window, three tools, Google's stance on AI content) have no source here. Both speakers sell related services, and Troy discloses the client relationship. **Next steps for readers:** before any AI-drafted client document goes out, read it side by side with the call notes and rewrite sections that don't sound like you. Audit staff use of personal AI accounts and move business use onto managed ones. Check the Drive retention window and your offboarding steps against your own plan's documentation. Treat the one-job loss as a prompt to measure your own win rate, not as proof. ## 🔄 Refactor Appendix — The AI Proposal That Lost Us a Client (with Duncan Isaaksen-Loxton) > Density: 2–5 Literature atoms (H3 pad), 0–3 Permanent candidates (H4 pad). > Emitted: 5 Literature, 2 Permanent. Resonance filter applied: the Woo themes, > hospital and Visa stories were dropped as secondhand or unverified. > > 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 #### [H1] Clients detect AI proposals that ignore call rapport > [!quote] [00:45:26] > It sounds like you did this proposal with AI. Um, because there's a few things that I think you missed. **Context & use**: - 00:45:26, the client's written feedback on the 13-page, eight-website proposal, after a discovery call the client says built a human connection. - The failure isn't accuracy. A human read and edited the draft under a "read, tweak, send" policy, and the voice mismatch against the call was still enough to cost the job. **Tags**: #xref #articleseed  #### [H2] AI on weak foundations accelerates mess > [!quote] [00:05:02] > AI without having the foundation in place first cuz otherwise all you're doing is you're accelerating mess. **Context & use**: - 00:05:02, closing Duncan's description of the business operating system (foundation, operations, optimized). - The order matters: identity, backups and policy before automation. Test it on any AI rollout at a small firm. **Tags**: #xref  #### [H3] Tools don't supply the judgment to use them > [!quote] [00:34:30] > AI is like the power drill, right? But the tradey who's who's done this before and has put a few holes in walls **Context & use**: - 00:34:30, the analogy for why owners who use AI without expertise produce work that fails later (the picture that falls off the wall). - The quote is cut at the clause end; the full sentence continues "he knows to find a stud first." The snippet-plugin support ticket that follows is the concrete case. **Tags**: #xref #aphorisms  #### [H4] Owners answer for data staff paste into personal AI > [!quote] [00:12:57] > and you as the business owner are responsible for that content and for the keeping of it. **Context & use**: - 00:12:57, about two staff at a current client using personal ChatGPT accounts with business data. - The liability stays with the owner, so the control belongs in policy (managed accounts), not in trust. **Tags**: #xref #share  #### [H5] Read, tweak, send is a weak final gate > [!quote] [00:44:43] > if you generate it with AI you read it you tweak it you make it your own you send it **Context & use**: - 00:44:43, Duncan's stated policy before the failure. He says he was tired and missed AI's dramatic prose, and the voice-profile checker didn't catch it either. - Counterpoint for policy design: a review step depends on the reviewer's state, so pair it with a comparison against the call transcript. **Tags**: #xref #articleseed  ### 🧠 Atomic Split Pad #### [P1] Review steps fail when the reviewer is depleted **Claim** (one sentence, Courtney's voice): A human review step protects quality only when the reviewer has the capacity and a reference to compare against, so a tired reader will pass drafts a customer will fail. **Evidence pointer**: [[#H1]] · [[#H5]] **Scope & conditions**: Holds for client-facing prose where voice is part of the value. Doesn't apply to factual checks with a checklist, and one case (a single lost job) is weak evidence on its own. Strengthen by testing a call-transcript comparison as the gate. **Links**: - [[Human in the loop]] — because extends - [[AI drafting workflows]] — because applies-in agency proposals - [[Automation and trust]] — because contrasts on where the failure sits — in the reviewer rather than the model  #### [P2] Foundations come before AI **Claim** (one sentence, Courtney's voice): AI on top of missing identity, backup and policy basics speeds up existing mistakes instead of fixing them. **Evidence pointer**: [[#H2]] · [[#H4]] **Scope & conditions**: Holds for small organizations with ad hoc tools and personal accounts. Breaks if the AI use is isolated from business data or if the foundation work costs more than the risk it removes. **Links**: - [[Small business security basics]] — because extends - [[Personal accounts and data liability]] — because evidence-for - [[AI adoption journey]] — because applies-in sequencing decisions  ### ✏️ Fleeting Queue - [F1] Check Google's admin documentation for the Drive deleted-file recovery window Duncan calls "about a 30-day window" — https://www.youtube.com/watch?v=RJ3cyazAycU #fleeting #inbox - [F2] Confirm the correct spellings: the transcript's "65.io" is sixfive.io, "Isaacson Lockxton" is Isaaksen-Loxton, "Victor" is Viktor — check the video description before citing #fleeting #inbox - [F3] Open question: does a call-transcript comparison catch voice drift that a voice-profile checker misses? The proposal at 00:45:26 is the case to test #fleeting #inbox - [F4] Article seed: "The proposal was accurate and still lost the job" on how review policies fail when the reviewer is tired #fleeting #inbox #articleseed - [F5] Find a source for the claim that Google is downranking AI-generated content, stated at 00:32:20 without a citation #fleeting #inbox ### → Things 3 (max 2 actions) - [ ] Draft a read-against-the-call checklist for AI-assisted client documents — from the proposal story at 00:43:15 #things - [ ] Verify the Google Drive deleted-file retention window in Google's admin docs before repeating the 30-day figure — fact check, 10 min #things  📄 **Full text:** [[Full Document Contents/Articles/The AI Proposal That Lost Us a Client.md|The AI Proposal That Lost Us a Client by Duncan Isaaksen-Loxton]] ## Highlights - [00:00:00] It sounds like you did this proposal with AI. Um, because there's a few things that I think you missed. Um, but also it's so far away from the way that you talked to me on the call, which I came out of feeling like you understood what was going on. We had a human connection and then you sent me this proposal that just felt so so very disconnected. [music] >> Hey, welcome to the Agency Hour podcast. My name is Troy Dean. I'm your host here and we're back. We haven't uh published an episode of the podcast for a while and that's because we've been extremely freaking busy and uh completely snowed in here, which is a good thing. Uh however, we do have a new format to the [00:00:43] podcast. Uh not this week, but in the coming weeks, I will announce a slightly new format to the podcast and a new rhythm, a new cadence that we'll be publishing the podcast. And also, I want your feedback. So, uh, you can either send me an email like legit [email protected]. It's my email. I read every email. There's no AI. There's no VA here, right? It's me reading every single email. So, send me an email troy@ag agencymicks.com. In the coming weeks, we're also going to launch a fan mail feature where you will be able to go to um the agency hour.com and you'll be able to click a button and leave us a voice message with a question that I will answer uh on the podcast. And if there's anything specifically ([View Highlight](https://read.readwise.io/read/01m3pske5ddtbw51pagzg4adxv)) - Note: [H1] Clients detect AI proposals that ignore call rapport - Tags: atom/literature/h1 - [00:04:20] operational phase, and then pushing you through into that kind of optimized phase where you can align it to this sort of AI adoption journey. You can think of it as I've just jumped onto chat GPT and asked a quick question got for a generic answer >> through to you know using something like Victor which is Victor now understands my business and how I talk and all of my SOPs and my pricing and my products and my ideal client >> and it can produce a draft email. It can even help me communicate with some of my clients. It can even help me execute some of what I do with my clients as well. So it's it's that kind of >> progressive curve and that journey. But you can't have something like Victor doing all of that cool for you in [00:05:02] AI without having the foundation in place first cuz otherwise all you're doing is you're accelerating mess. >> Yep. >> Does that make sense? >> Yep. Totally. Uh who who's your ideal client? So we mostly work with entrepreneurs uh and small business owners who are kind of one to 50 odd seats uh as in humans in the business uh who are you know going through that journey of well I started a business because I'm really good at this and I want to spend my time doing that but I don't yet have the revenue or um the the capacity to hire someone full-time to look after this for ([View Highlight](https://read.readwise.io/read/01m3pskgxvrxggzzxf1fb22kvn)) - Note: [H2] AI on weak foundations accelerates mess - Tags: atom/literature/h2 - [00:12:57] So, we're we're dealing with a client at the moment who uh has, you know, quite a few um two of the staff have signed up for chat GPT with their own personal email addresses and they're taking and it's a personal chat GBT account, so they're also separate and they're taking business data and putting it into that personal chat GPT account. >> Yeah. >> Right. So there's a few problems with that is is that data is now being exfiltrated from the business >> and you as the business owner are responsible for that content and for the keeping of it. And you know, imagine how you would feel and you probably have done at some point in the in the past that you've discovered your email address and probably your your physical address was [00:13:39] >> in a data leak somewhere, right? In a data dump and it got sold >> and that's not a pleasant thing to know especially when uh I think just in the past few days there's been uh you know data leaks out of all sorts of places and I think Medicare in Australia is one right. >> That's right. Yeah. You know, nobody wants as a business owner, you do not want to be on the end of that phone call >> when someone says, "Hey, your my my stuff has been >> leaked." >> So, you know, the the temptation to kind of go and sign up for the latest shiny latest and greatest and just again solve the problem. >> I get it. >> But there's got to be some discipline around that for multiple reasons. >> Yeah. Yeah. Well, if you think I mean I ([View Highlight](https://read.readwise.io/read/01m3pskqa3eqxvpt8v3t1pc31k)) - Note: [H4] Owners answer for data staff paste into personal AI - Tags: atom/literature/h3 - [00:34:30] Um AI is like the power drill, right? But the tradey who's who's done this before and has put a few holes in walls, he knows to find a stud first, right? And >> and I I'll give you an example. We have a literally got a support ticket I think yesterday or the day before from a client going, "Look, I've found this particular problem on our website. I went to chat GPT. It gave me this snippet of code. I put it into the code snippet editor plugin on my website and it didn't work." [gasps] And I did exactly that. I went, "Oh my god, you you know that you already pay us to do those things for you, right? You don't need to go do that yourself." >> Yeah. Yeah. Do you have a policy about I have a policy that uh people send me people [00:35:14] send me 30page documents written by AI that are their all their thinking all their thinking on the business right and then they want me to you know help them implement this stuff and so I have a policy now that if anyone sends me a long document that's clearly written by AI my response to them is a 350 page response written by Claude that they have to read word for word before our engagement continues. And what I'm trying to do there is I'm trying to say, look, just because you can doesn't mean you should stop doing that because a 30page document on your thinking about the business is might be useful if you're doing a, you know, strategic offsite once a year with your team, but it's not useful for me. One, I haven't ([View Highlight](https://read.readwise.io/read/01m3pskm1pn9qr4h3gqadf4ytk)) - Note: [H3] Tools don't supply the judgment to use them - Tags: atom/literature/h3 - [00:44:43] transcript from that. Uh let's start to kind of meld those two things together with our, you know, rate card essentially like what do we do? How do we do it for people, etc. Okay, proposal. Tada. uh couple hours later right now uh we don't what our policy is if you generate it with AI you read it you tweak it you make it your own you send it okay nothing gets sent by AI automatically so same thing read through the proposal and I was probably knackered or not feeling all that great I missed some of the uh you know the dramatic pros that AI tends to put into things the the overeing of stuff right um we have a fairly good voice profile [00:45:26] that does a bunch of checking for those things and tries to make it more natural, you know, sounding for us and the way that we would write. >> The feedback that I got from the client, so number one, amazing. So, so very grateful I actually got some feedback was it sounds like you did this proposal with AI. Um, because there's a few things that I think you missed. Um, but also it's so far away from the way that you talked to me on the call, which I came out of feeling like you understood what was going on. We had a human connection and then you sent me this proposal that just felt so so very disconnected. And I was like, "Oh shit." [snorts] And read it back and was like, "Yeah, actually, fair cop. You know what? Screwed that up massively." Um, ([View Highlight](https://read.readwise.io/read/01m3pskt7tw772ftgpgp81cg5s)) - Note: [H5] Read, tweak, send is a weak final gate - Tags: atom/literature/h3