Every GTM platform now ships an AI agent. The vendor demos are genuinely impressive, the board is asking where AI shows up in the roadmap, and there's suddenly budget for exactly this. The pressure on anyone who runs a revenue-technology stack is real, and it all points the same direction: turn it on.

Here's the uncomfortable part, from where I sit: most GTM AI initiatives don't fail at the model layer. The models are good — better than the data they're about to be pointed at. They fail at the data layer, quietly, a few weeks after the demo.

What an agent actually consumes

Strip away the interface and a GTM agent consumes three things: your CRM records, your activity history, and your institutional knowledge. Account and opportunity data. Emails, calls, meetings, campaign touches. Whatever documentation and process knowledge you've given it access to.

That inventory should worry you more than any question about the model, because of one property that makes AI different from every reporting tool you've deployed before:

An agent doesn't average your bad data. It speaks it fluently.

A dashboard with duplicate accounts produces a number someone squints at and distrusts. An agent with duplicate accounts produces a confident, well-written, plausible answer assembled from three conflicting versions of the truth — and it delivers that answer directly to a seller who has no way to see the conflict underneath.

The demo works because demo data is clean. Production data is where the account exists three times with two owners and an ambiguous status, where activity capture has a six-month gap from that one integration migration, and where the knowledge base still describes the process two reorgs ago. None of that blocked the demo. All of it surfaces in week three.

Play the failure forward. A seller asks the agent for an account summary before a renewal call. The agent finds the churned-looking duplicate, blends it with the active record, and produces a crisp paragraph implying the customer has been quiet for two quarters — when the real activity lives on the other record. The seller either catches it and stops trusting the tool, or doesn't catch it and walks into the call wrong. Both outcomes are expensive, and neither is the model's fault.

The prerequisites are boring, and that's the point

What separates the stacks where AI lands from the stacks where it stalls isn't model choice or prompt craft. It's three unglamorous properties:

A shared data model. Sales and marketing agree on what an account is, what an opportunity is, what a touch is — one definition each, enforced in the systems rather than reconciled in spreadsheets. If your funnel reporting requires a human to explain which numbers to trust, an agent inherits that ambiguity and hides it behind fluent prose.

Ownership. Every data domain has a name attached — someone who can say yes, say no, and fix what's broken. "The CRM team" is not a name. Agents surface data-quality debt at conversational speed; unowned debt just compounds.

Consent boundaries. Which records an AI may read, and which it may act on, is a governance decision that has to be encoded before the first agent ships — not discovered afterward. Privacy regulation makes this a hard boundary, not a preference. That's all I'll say here, because it deserves its own piece; for now, treat consent as an AI-readiness input, not a compliance afterthought.

Behind those three sit three quieter ones that determine whether the foundation stays sound. Quality has to be monitored, not discovered — if duplicates and stale records only surface when a user complains, an agent will find them first and at scale. Access has to be governed deliberately: an agent is a new consumer of every permission decision you've ever made, including the sloppy ones, and "who can see what" questions that were theoretical in a reporting tool become live in a conversational one. And someone has to run all of this in month six — AI systems are operated, not launched. If that sounds familiar, it's the same rule that governs buy-vs-build decisions: operational complexity survives the build.

Govern, unify, then automate

The sequence I'm running is deliberate, and it's the opposite of how most AI initiatives get scoped:

In the stack I run, AI came last on purpose. First came governance and the unified commercial data foundation — one data model across CRM, marketing, and the web platform, with the reporting infrastructure sitting on a governed lakehouse. The AI-enabled workflows now rolling out across the platform (Slack + Agentforce) were funded by reallocating underutilized platform spend — capabilities we were paying for and not operating — into the layer we'd actually prepared the ground for.

The same lakehouse that made marketing attribution answerable turns out to be the substrate the agents consume — the governance work done for analytics was the AI prerequisite, we just didn't call it that at the time. I wrote about the buy-vs-build side of that foundation in Buy vs. Build for the GTM Stack: governance is the new moat, and as more of the stack becomes AI-assisted, clean, governed, joined data is the differentiated asset — which is a build, and always was.

Run the sequence backwards and you get the stalled pilot everyone has seen by now: agent first, then a scramble to explain its answers, then a "pause" that never un-pauses. The failure isn't dramatic. It's a slow leak of trust — and the first confidently wrong answer in front of a sales team costs more trust than ten right ones earn back.

Score your own stack

Before piloting anything user-facing, I'd score the foundation honestly across six factors:

1. Unified data model — one agreed definition of account, opportunity, and touch.
2. Ownership & stewardship — a named owner per data domain.
3. Consent & compliance boundaries — encoded, not assumed.
4. Quality monitoring — duplicates and staleness measured, not discovered by users.
5. Access governance — deliberate, auditable, least-privilege.
6. Operational capacity — someone runs this in month six. AI systems are operated, not launched.

I built an interactive version: the AI Readiness Assessment in this site's sandbox. Score each factor 1–5, adjust the weights to your context, and share the resulting URL with whoever you're trying to convince — the same pattern as the buy-vs-build tools. It's decision support, not certification: a low score doesn't mean "no AI," it means the first AI project should be the unglamorous one that fixes the foundation. The tool is deliberately sequence-aware — if a foundation factor scores below 3, it will tell you to fix that before anything else, no matter how the weights fall.

The sequencing decision

"AI on governed data" isn't a feature decision. It's a sequencing decision. The stacks where agents earn trust are the ones where the boring layers came first — and the leaders who look best in two years will be the ones who spent this year on data models and ownership while the demos got the applause.

Govern, unify, then automate. And be suspicious of any AI roadmap that starts at step three.