Why Your AI Doesn’t Know Anything About Your Deals

In the demo, the assistant is brilliant. Ask it about an account and it summarizes the history, surfaces the risk, drafts the follow-up — because demo data is clean, current, and singular. Then it meets your CRM. Ask the same question on a real Monday and it answers just as fluently — about the wrong duplicate, from a stage that changed two quarters ago, filling the gaps with plausible things nobody ever said. The disappointment that follows usually gets filed as “the AI isn’t ready.” It’s worth being precise about what actually happened, because the fix depends on it.

The model reads; it doesn’t know

An AI tool pointed at your sales data has exactly one source of truth: the records you gave it. It cannot tell a current field from one nobody updated since spring. It cannot tell which of three versions of the same company is real. Where a field is empty, it does what language models do — produces something plausible, delivered in the same confident register as everything else. The confidence of the answer and the quality of the record underneath it are entirely unrelated, which is precisely what makes the failure expensive: wrong answers that sound right get acted on.

Nobody in this story did anything wrong. The vendor shipped a capable model; your team ran the CRM the way every busy team does. The gap is structural: the tools assume a record layer that most revenue stacks were never wired to maintain. The AI didn’t fail. It faithfully summarized a fiction.

Three ways a good model gives a bad answer

Duplicates make it answer about the wrong entity — three versions of the account, each holding a different fragment, and the model reads whichever it found, presenting one fragment as the whole. Decay makes it answer from the past — deal values, stages, and owners that were true once, delivered in the present tense with nothing to flag their age. And gaps make it invent — not maliciously, but because completing patterns is what the machinery does, and an empty next-step field is a pattern begging to be completed.

Notice that all three are properties of the data layer, not the model. Swapping vendors re-runs the same experiment on the same records. Teams cycle through tools looking for the one that “gets” their pipeline, when no tool can read what was never written down — the search is happening one layer too high.

If you haven’t bought AI yet, you’re early — usefully early

The teams that get real value from AI on their sales data mostly aren’t the ones with the best tools. They’re the ones that arrived with a record layer worth reading: one record per account, a pipeline whose stages mean what they say, and fields that stay current because the system writes them rather than asking reps to. That is a buildable condition — and building it before the licences beats buying licences that renew while the data catches up.

So the readiness question isn’t “which AI should we buy?” It’s the plainer one: if you handed your CRM to an AI tool tomorrow, how much of what it told you would you trust? If the honest answer is “some of it — we’d verify anything important,” then the record layer is the project, and it’s a project worth doing even if AI never enters the picture. Every layer above the record inherits its trust — reports and humans included. The AI just inherits it faster and with more confidence.

Grounding, and its boundaries

Grounding is the unglamorous step the demos skip: connecting the tools you already license to your actual record layer, deliberately — including deciding, field by field, what the AI is allowed to read, on an explicit allowlist you review before anything is connected. Done right, the assistant answers about your deals instead of the world in general, and you can say exactly what it can and cannot see. That is a bounded piece of wiring, not a moonshot.

And it has an honest boundary: grounding cannot rescue a record layer nobody trusts. Connecting a capable model to records full of duplicates and decay just produces fluent summaries of the mess. If the data layer is the doubt, that is the first build — the sequencing isn’t a sales pitch, it’s how the layers actually stack. The readiest teams arrive at AI with records it can stand on. The rest arrive with licences.