AI Agents in GTM: What They Should Truly Do  🔗 **Source:** [AI Agents in GTM: What They Should Truly Do](https://www.commonroom.io/blog/ai-agents-gtm/) > [!note]+ Document Notes > tags: [ai-agents, go-to-market, revops, sales-automation, crm-data-hygiene] > author: [MISSING] > source: Common Room > reference: https://www.commonroom.io/blog/ai-agents-gtm/ > date created: {{date:YYYY-MM-DD}} {{time:HH:mm}} > > # 🏷️ Title: AI Agents in GTM: What They Should Truly Do > <!-- 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) ### 🔗 References - [72% of time sales reps spend on non-selling work — Salesforce](https://www.salesforce.com/news/stories/sales-research-2023/) - [CRM data decay rate 25-30% per year — Verum](https://veruminc.com/resources/crm-data-decay-rate) - [Common Room — AI GTM strategy](https://www.commonroom.io/blog/ai-go-to-market-strategy/) ## 📝 Summary - **What happened.** Common Room argues that "AI agent" has lost meaning through hype, and offers a plain working definition: a system that takes an input, does a defined job, and produces output a human can act on. It then names three GTM jobs where agents create real leverage — research, personalization, and data hygiene — and pitches its own products against each. - **Why it matters.** Most AI-agent disappointments are deployment failures, not tool failures: agents get pointed at the wrong job or fed fragmented/stale data. The piece reframes the buying question from "which agent" to "which job, on what data foundation." - **Key stakeholders.** RevOps teams (the named audience), SDRs/AEs whose non-selling time the agents target, and Common Room as vendor. - **Risks and tradeoffs.** This is a vendor blog; every problem framing routes to a Common Room product (Context360, message personalization agent, DataAgent). The three-jobs taxonomy is plausible but unproven, and the "foundation first" argument conveniently positions the vendor's unified-data layer as the prerequisite purchase. - **Bottom line.** Solid, non-hype mental model — define the job and clean the data before deploying agents — wrapped in a product pitch. Take the framework; discount the "only our foundation makes it work" implication. ## 📖 Main Content - **Core definition.** An AI agent is deflated from autonomy hype to "input → defined job → actionable output." Useful in GTM only when aimed at the right job with the right inputs. - **Why deployments disappoint.** Teams deploy against their biggest workflow, get more volume at the same close rate, and mistake speed for improvement. The tool works; the deployment (wrong job or wrong inputs) fails. "An AI agent is only as useful as the data it operates on and the job it's been given to do." - **Start with the job.** Jobs worth automating are repetitive, data-intensive, have a clear definition of done, and are currently done inconsistently across the team. Strategy, relationships, and judgment stay human. - **Three jobs.** - *Research agent — context before contact.* Cites Salesforce's 72% non-selling-time figure. Distinguishes information (a Series B six months ago) from context (Series B + three new sales leaders + two pricing-page visits + match to last six closed-won deals). RevOps must give the agent unified first- and third-party data resolved to real people/accounts, and define what a "useful brief" means per motion. - *Personalization agent — written for someone, not at someone.* First-wave AI outreach trained buyers to distrust name/company mail-merge. Real personalization is behavioral, not demographic, and must sit downstream of research ("Research first, personalization second"). - *Data hygiene agent — the one everybody avoids.* CRM accuracy decays ~25-30%/year. A continuous hygiene agent (vs. quarterly cleanup) flags stale/duplicate/role-changed records and preserves relationship history. - **System effect.** The three compound on a shared intelligence foundation; reps arrive to a pre-organized queue and keep only the "irreplaceable" conversation/judgment work. - **What RevOps must get right.** Foundation before agents; define the job before selecting the tool; adoption is a design problem (embed in Slack/CRM/email), not a training problem. - **Uncertainties / contested claims.** The 72% and 25-30% figures are single-source secondary citations (Salesforce PR, a vendor resource page). The clean job taxonomy and "compounding moat" claim are asserted, not evidenced. Product capability claims (Context360, DataAgent) are unverified marketing descriptions. ### 🗒️ Notes - Hype-cycle framing: research concept → product category → "agentic AI" on every homepage → meaning lost. - Working definition strips "autonomy theater"; agents are narrow job-doers. - Failure pattern is specific and repeatable: more sequences, same close rate, output "faster and more voluminous" but not better. - Three mispointing examples: research agent on fragmented contacts = fast research on wrong people; personalization on account-level intent = specific-sounding but not; prioritization on stale CRM = confident ranking of accounts that may not exist. - Job-selection test: repetitive, data-intensive, clear definition of done, inconsistently done today. - Research memorable line: non-selling work is "janitorial work that happens to require a college degree." - Information vs. context is the load-bearing distinction of the research section. - Personalization ordering constraint: personalization agent can't be standalone; must operate on context the research agent already surfaced. - Data hygiene analogy: cleanup project vs. continuous agent = "bailing water" vs. "fixing the leak." - Foundation claim: "Nobody's AI stack can outrun a bad foundation. They just generate bad output faster." - Closing pitch paragraph explicitly names Common Room as "the AI-native GTM platform." - Product mentions (promotional inserts): Context360, message personalization agent, DataAgent — each dropped at the end of its matching section. ## 📚 References - [Sales research: reps spend 72% of time on non-selling work](https://www.salesforce.com/news/stories/sales-research-2023/) — Salesforce - [CRM data decay rate (25-30%/year)](https://veruminc.com/resources/crm-data-decay-rate) — Verum - [Confident-sounding noise at scale: AI GTM strategy](https://www.commonroom.io/blog/ai-go-to-market-strategy/) — Common Room - [Five Signals. Five Plays. Zero Cold Outreach.](https://www.commonroom.io/blog/signal-based-outbound-plays/) — Common Room - [Common Room + Nooks](https://www.commonroom.io/blog/nooks/) — Common Room - [Your SDRs aren't slow. Your system is.](https://www.commonroom.io/blog/ai-sdr-playbook/) — Common Room ## 🗣️ Paraphrase - The term "AI agent" got hyped into vagueness; a useful definition is just a machine that takes an input, does a specific job, and hands back something a person can act on — no autonomy mysticism. - Agents are genuinely useful in go-to-market, but only against the right jobs, and most teams pick the wrong ones. - Deployments disappoint because teams aim agents at their biggest workflow and get more output at the same result; the tool didn't fail, the deployment did — usually wrong job or wrong data. - Before deploying, RevOps should name the exact manual job being replaced and what good output looks like; the best-fit jobs are repetitive, data-heavy, clearly "done"-able, and unevenly done today. - Three jobs create real leverage: research (turn scattered signals into context before outreach), personalization (behavioral, specific, downstream of research), and data hygiene (continuous, not quarterly). - The three compound on a shared, clean data foundation, leaving reps only the human work of conversation and judgment. - Three success conditions: fix the foundation first, define the job before choosing a tool, and treat adoption as an embed-into-workflow design problem. ## 📰 Newsletter - Not applicable to this source — reason: the article is a B2B sales/RevOps vendor post with no WordPress, plugin, or community angle to blurb. ## ✅ Action Items #### DevRel - Developer experience: if building agent-style tooling, publish a "define the job + expected output" template before shipping the feature, mirroring the article's job-first framing. - Developer marketing: note the information-vs-context distinction as a reusable content hook for any AI-feature launch. #### OSPO - Not applicable to this source — reason: no open-source governance, licensing, or SBOM content here. #### Community - Not applicable to this source — reason: no maintainer/contributor/sponsor ecosystem content. #### WordPress - Not applicable to this source — reason: nothing touches WordPress core, themes, plugins, or project teams. ### FAIR - Not applicable to this source — reason: no federated-repository or governance-independence content. ## 🧪 Hypotheses - H1: Teams that write an explicit "job + definition of done" before agent deployment report higher rep adoption than teams that deploy tool-first. - H2: Sequencing personalization downstream of research produces higher reply rates than running a standalone personalization agent. - H3: Continuous data-hygiene agents keep CRM accuracy materially above the ~25-30%/year decay baseline versus quarterly cleanups. - H4: Agent output embedded in reps' existing tools (Slack/CRM/email) is adopted more than equivalent output in a separate dashboard. - H5: "More sequences at the same close rate" is a common measured outcome of naive agent rollouts. ### 📊 Methodology - H1/H4: A/B rollout across matched RevOps teams; measure agent-output usage rate and self-reported trust over 90 days. - H2: Controlled outbound test — research-then-personalize pipeline vs. standalone personalization — measure reply and positive-reply rates on matched account lists. - H3: Longitudinal CRM audit sampling contact accuracy quarterly under continuous-agent vs. periodic-cleanup conditions. - H5: Cohort analysis of close rate and sequence volume pre/post deployment across multiple orgs. ### ❓ Questions - What false-positive rate does a continuous hygiene agent produce when flagging "stale" records, and what's the cost of wrongly archiving a live contact? - How is "useful brief" operationalized differently for outbound-enterprise vs. PLG-expansion motions? - Does the "compounding moat" survive when competitors run agents on the same third-party signal sources? - What close-rate lift (not just volume) has actually been measured, and by whom? ### 🔮 Future Research - Independent (non-vendor) benchmarks on close-rate impact of research/personalization/hygiene agents. - Studies on buyer trust decay as AI-generated outreach saturates inboxes. - Cost/accuracy tradeoffs of continuous CRM hygiene automation at scale. ### 💡 Implications - **Practice:** shifts RevOps buying from feature comparison to job definition and data-foundation readiness. - **Policy:** continuous automated CRM editing raises data-governance questions (audit trails, reversibility, consent). - **Security:** unified first- plus third-party buyer data resolved to real people concentrates PII and intent data — a governance and breach-surface concern. - **Education:** the information-vs-context frame is a teachable heuristic for any AI-output evaluation. - **Funding models:** vendors will pitch the "foundation" (data layer) as the anchor purchase that agents ride on. ## 🧠 Learning Styles Assessment - Reading level estimate: Flesch-Kincaid Grade ~8-9 (short sentences, some fragments for rhetorical effect, minimal jargon). Estimated by sentence-length and syllable inspection, not a computed score. #### 👀 Visual learners - Present: two decorative header images; no data visuals. - Improve: a diagram of the research → personalization → hygiene loop; a before/after "information vs. context" brief mockup. #### 🔊 Aural learners - Present: none. - Improve: a 3-minute audio summary or narrated walkthrough of the three-agent system. #### 🗣️ Verbal learners - Present: strong analogies ("mail merge with better syntax," "bailing water vs. fixing the leak," "janitorial work that happens to require a college degree"), clear definitions. - Improve: a short glossary (agent, first-party vs. third-party signal, intent data, resolution). #### 🧩 Kinesthetic learners - Present: implicit checklist in "What RevOps needs to get right." - Improve: a deploy-readiness worksheet (name the job, define done, audit the data). #### 🧠 Logical learners - Present: clear cause→effect chains (bad data → bad briefs → wrong inbox → bad ranking). - Improve: a decision tree for "should this job be agent-automated?"; measurable KPIs per agent. #### 🧑‍🤝‍🧑 Social learners - Present: none beyond implied team context. - Improve: discussion prompts for RevOps + sales alignment on "what good output looks like." #### 🧘‍♀️ Solitary learners - Present: reflective framing ("what specific manual job is this agent replacing?"). - Improve: a self-audit quiz on current data hygiene and job clarity. #### 🌿 Naturalistic learners - Present: none. - Improve: lifecycle framing of CRM data decay as an ecosystem to maintain. ## 📊 Accessibility Assessment #### Screen reader semantics - Present: proper H2 section headings; images exist. - Improve: header images have empty/missing alt text — add descriptive alt or mark decorative. #### Keyboard navigation - Not applicable to this source — reason: static article, no interactive controls to evaluate from the captured HTML. #### Contrast and readability - Not applicable to this source — reason: rendered site styling not inspectable from extracted HTML content. #### Text resizing and reflow - Not applicable to this source — reason: reflow depends on the live site CSS, not evaluable here. #### Media and captions - Present: no audio/video, so no caption needs. - Improve: if repurposed to audio, provide a transcript. #### Forms and interactions - Not applicable to this source — reason: no forms in the article body. #### Standards reference - Missing/empty image alt text maps to WCAG 2.2 AA 1.1.1 Non-text Content. ## 🌍 DEIB Assessment #### Cultural breadth - Present: none; examples are generic B2B SaaS motions. - Improve: acknowledge non-US/non-English GTM contexts and buyer norms. #### Inclusive language - Present: gender-neutral throughout ("reps," "buyers," "the human"). - Improve: define insider jargon (PLG, outbound, closed-won) for readers outside enterprise sales. #### Neurodiversity support - Present: chunked sections, consistent structure, short paragraphs. - Improve: a TL;DR box up top and icon+text section labels. #### Representation - Not applicable to this source — reason: no people, roles, or imagery of individuals depicted to assess. #### Psychological safety - Present: respectful, non-blaming framing ("the tool didn't fail, the deployment did"). - Improve: none material. ## ✅ Logical Validity #### Strong reasoning - **Causal chain (data hygiene section):** stale records → stale briefs → wrong-inbox messages → invalid ranking is a clean, valid causal argument with a clear mechanism. Strong because each link is independently plausible and the conclusion ("bad output faster") follows. - **Definitional deflation (intro):** reducing "AI agent" to input→job→output is a useful stipulative definition that grounds the rest of the piece. Strong as a framing move; clarify by acknowledging it's stipulative, not the industry consensus. - **Analogical reasoning:** "mail merge with better syntax" and "bailing water vs. fixing the leak" are apt analogies that carry real explanatory weight rather than mere color. #### Fallacies - **Self-serving bias / promotional framing (the dominant weakness).** Every diagnosed problem resolves to a Common Room product. The vendor pitch "Common Room's research agents pull from a unified intelligence foundation that combines first-party customer data with real-world buyer signals, resolve everything through Context360 to real people and accounts, and surface briefs that give reps the context they need before they ever open a compose window." presents proprietary capability as the natural solution to the problem the article just defined. Fix: separate the vendor-neutral framework from product claims, and cite independent evidence that the foundation must be theirs. - **Appeal to (thin) authority / cherry-picked statistics.** The "72%" and "25-30%" figures anchor the argument but come from a single vendor-friendly source each; presented as settled facts. Fix: cite multiple independent sources or give ranges and methodology. - **False dilemma (mild).** "Adoption is a design problem, not a training problem" frames two contributors as mutually exclusive when both can matter. Fix: "primarily a design problem" preserves the point without the exclusivity. - **Hasty generalization.** "The pattern is consistent enough to be predictable" and "Most teams skip it" assert prevalence without data. Fix: qualify or cite. - **Assertion-as-conclusion.** "That's the compounding moat nobody puts on a roadmap slide" is a rhetorical flourish presented as an earned conclusion; the moat is claimed, not demonstrated. Fix: show the compounding mechanism with a concrete before/after. #### Tone of voice - Confident, punchy, contrarian-insider. Uses sentence fragments for emphasis ("Research. Personalization. Data hygiene."), plain-spoken authority, and a mild disdain for hype. Reads as a knowledgeable practitioner — which is also the persuasion vehicle for the underlying pitch. ## 🛡️ DARVO Analysis - **Deny:** No accused-party denial dynamic. Mild deflection of category-level criticism — concedes "personalization has a bad reputation right now, and it deserves it," then reframes the failure as bad inputs rather than a flaw in the vendor's approach. - **Attack:** No attacks on individuals. Rhetorical jabs at unnamed "vendor demos" ("this is the part the vendor demos don't show you") position competitors as the problem while exempting the author's own product. - **Reverse victim/offender:** Not applicable to this source — reason: no accuser/accused dynamic; the piece is persuasive marketing, not a defense against an allegation. - **Emotional cues / rhetoric:** relies on validation ("you were right to distrust that outreach") and relief ("removing everything that was getting in the human's way") rather than fear or grievance. ## 🎯 Conclusion - **Bottom line.** The framework is genuinely useful: deflate "AI agent" to a job-doer, pick jobs that are repetitive/data-intensive/clearly-done/inconsistent, sequence research → personalization, and fix data hygiene continuously. The recurring move — every problem resolves to a Common Room product on a Common Room "foundation" — is the part to read critically. - **Practical next steps for readers.** Before buying any GTM agent: (1) write the exact manual job and its definition of done; (2) audit whether your data foundation can support it; (3) pilot on your data and your motion, measuring close-rate lift, not sequence volume; (4) embed output where reps already work; (5) treat the "you need our unified foundation" claim as a hypothesis to test, not a premise to accept. ## 🔄 Refactor Appendix — AI Agents in GTM: What They Should Truly Do > Density: 2–5 Literature atoms (H3), 0–3 Permanent candidates (H4). > Defaults: 3 Literature, 1 Permanent. Resonance filter: if you can't > articulate in one line why a claim matters, drop it. ### 📚 Literature Atoms (H3) #### [H1] Agent value equals data quality times job fit > [!quote] ID:rw_01kx13dy4h1psaeyj10s7rv1s0 > An AI agent is only as useful as the data it operates on and the job it's been given to do. **Context & use**: - "Why most AI agent deployments disappoint" section. - The article's central claim; reframes agent buying from tool selection to job + data readiness. **Tags**: #xref **Highlight**: [view highlight](https://www.commonroom.io/blog/ai-agents-gtm/) <!-- atomicity: 1 idea ✓ --> #### [H2] Deployment fails, not the tool > [!quote] ID:rw_01kx13dy4h1psaeyj10s7rv1s0b > The tool didn't fail. The deployment did. **Context & use**: - Same disappointment section. - Relocates blame for failed AI rollouts from vendor tech to job/input choices — useful diagnostic lens. **Tags**: #xref **Highlight**: [view highlight](https://www.commonroom.io/blog/ai-agents-gtm/) <!-- atomicity: 1 idea ✓ --> #### [H3] Bad foundations just fail faster > [!quote] ID:rw_01kx13dy4h1psaeyj10s7rv1s0c > Nobody's AI stack can outrun a bad foundation. They just generate bad output faster. **Context & use**: - Data hygiene section. - Makes the case that data quality gates all downstream agent value; also the article's pivot into its foundation pitch. **Tags**: #xref #aphorisms **Highlight**: [view highlight](https://www.commonroom.io/blog/ai-agents-gtm/) <!-- atomicity: 1 idea ✓ --> #### [H4] Research produces context, not just information > [!quote] ID:rw_01kx13dy4h1psaeyj10s7rv1s0d > Context is what reps need before they pick up the phone. **Context & use**: - Research agent section. - The information-vs-context distinction is the piece's most transferable evaluative heuristic for AI output. **Tags**: #xref **Highlight**: [view highlight](https://www.commonroom.io/blog/ai-agents-gtm/) <!-- atomicity: 1 idea ✓ --> ### 🧠 Permanent Candidates (H4) #### [P1] Define the job before selecting the automation **Claim** (one sentence, Courtney's voice): An automation is only worth deploying once you can state the exact manual job it replaces and what good output looks like. **Evidence pointer**: [[#H1]] **Scope & conditions**: Holds for narrow, repeatable, data-intensive jobs with a clear definition of done; breaks for judgment-heavy or relationship work where "good output" resists specification. **Links**: - [[AI tooling evaluation]] — because extends - [[Vendor-demo skepticism]] — because applies-in agent procurement <!-- atomicity: 1 idea ✓ --> ### ✏️ Fleeting Queue - [F1] Test the information-vs-context heuristic against my own AI-output reviews — https://www.commonroom.io/blog/ai-agents-gtm/ #fleeting #inbox - [F2] Note the "adoption is a design problem, not a training problem" line as a reusable framing (with the false-dilemma caveat) #fleeting #inbox ### → Things 3 (max 2 actions) - [ ] Draft a one-page "define the job + definition of done" agent-readiness checklist — content backlog #things <!-- deep-read:sections end --> 📄 **Full text:** [[Full Document Contents/Articles/AI Agents in GTM What They Should Truly Do.md|AI Agents in GTM: What They Should Truly Do by commonroom.io]] ## Highlights - The tool didn't fail. The deployment did. ([View Highlight](https://read.readwise.io/read/01kx43fhhhcdzra26pchfab2cc)) - Note: Relocates blame from vendor tech to job/input choice. - Tags: atom/literature/h2 - An AI agent is only as useful as the data it operates on and the job it's been given to do. ([View Highlight](https://read.readwise.io/read/01kx43fbrcqrh2tphrmpv8f5a1)) - Note: Central thesis: agent value = data quality × job fit. - Tags: atom/literature/h1 - Context is what reps need before they pick up the phone. ([View Highlight](https://read.readwise.io/read/01kx43fsy5xwheszyqm4rv9hfw)) - Note: Information-vs-context — the most transferable heuristic. - Tags: atom/literature/h4 - Nobody's AI stack can outrun a bad foundation. They just generate bad output faster. ([View Highlight](https://read.readwise.io/read/01kx43fp73wghezkyrqyt5sg6p)) - Note: Data quality gates all downstream value (also the pitch pivot). - Tags: atom/literature/h3 - That's the compounding moat nobody puts on a roadmap slide. ([View Highlight](https://read.readwise.io/read/01kx43fy6zgmvsdt3n8pr2wrqs)) - Note: Fallacy — self-serving: moat asserted, not demonstrated (vendor framing). - Tags: fallacy/self-serving