AI on top of a guess

Two numbers out of the same survey this week, published on 2 September. Five hundred US sales and revenue decision-makers were asked about AI in their process.

100% of them are using it somewhere. 20.6% describe their AI strategy as production-ready with measurable outcomes.

Adoption is finished. It's over, it's total, there is nobody left to convert. And roughly one in five has anything you could put a number against. The other four fifths are running tools they can't grade.

That gap gets blamed on the model most of the time. It almost never is.

Read further down the same survey

The interesting part isn't the headline, it's what the same 500 people said about the system underneath it.

55.6% said the information in their CRM is based mostly on subjective seller reporting. Not measured. Reported — by the person whose own numbers it reflects, typed in later, from memory, under pressure to look a certain way.

37.6% named CRM updates as their primary administrative bottleneck, and 31.4% called manual CRM administration the single greatest barrier in their pipeline. So the record is at once the most expensive thing they maintain by hand and the least trustworthy thing they own.

Then the consequence. Only 32% can instantly say why a deal stalled. 41% are slow to work it out or have no visibility at all. Another 27% can see their win and loss rates but cannot explain how anything moved between stages.

Read those four figures together and the 20.6% stops being mysterious. You cannot get a measurable outcome out of a system that cannot tell you why yesterday happened. No model fixes that. You are asking a very good reader to read a document nobody actually wrote.

This is not a sales problem

I've spent twenty years in operations, most of it in transportation and logistics, and I recognise this shape immediately. Swap the words and it is every dispatch floor I have ever walked onto.

The load record is the CRM. The timestamps get entered at the end of the shift, from memory, by the person who is also being measured on them. Everybody knows the arrival time in the system is approximate. Everybody works around it. Then somebody points intelligence at it and expects an answer.

The tool will give you one. That is the trap. It will hand back a fluent, confident, well-formatted answer built on a number that was a guess when a human typed it and is still a guess now that it has been averaged. You haven't added intelligence to the operation. You've added conviction to the guess.

What to do before you buy anything else

Map the record before you build on it. Pick the five fields your decisions actually rest on. For each one, write down where the value comes from: measured by a system, or typed by a person. That single sheet of paper will tell you more about your readiness than any vendor assessment, and it costs you an afternoon.

Get a baseline you can be embarrassed by. Sample the fields that turned out to be typed, against whatever ground truth exists, and count. A real error rate, on real work, written down. If you have no before, you can never prove an after — which is precisely how four fifths of a market ends up unable to describe its own results.

Fix the capture, not the analysis. Wherever a number can be captured at the moment it happens instead of remembered at the end of the day, move it. This is the unglamorous half, and it is where the return actually lives. Every hour spent making the record true multiplies through everything you point at it afterwards.

Keep the gate. Worth noting that the largest single group in that survey — 38.4% — picked a guided model with human oversight over anything running on its own. That is not timidity. When your inputs are self-reported, the human in the loop isn't a safety blanket, it is the only sensor you have.

When we have done this properly on an operation, the wins came from exactly that order of work. Inbound reaction went from three days and more to about sixty seconds, and daily escalations came down from seven to ten a day to three to five inside six weeks — not because the model was clever, but because by the time it ran, the thing it read was true.

Everyone bought the reader. Almost nobody wrote the document. Intelligence pointed at a guess just gives you a more confident guess.

Sources: Salesloft Revenue Benchmark Report — survey of 500 US sales and revenue decision-makers (2 Sep 2026)