Why AI pilots fail hardest in sales

MIT’s Project NANDA looked at enterprise generative AI and found that 95 percent of pilots delivered no measurable impact on the profit and loss. Not low return. Zero. RAND put the broader AI project failure rate above 80 percent, roughly twice the rate of conventional IT projects. S&P Global found that 42 percent of companies abandoned most of their AI initiatives in 2025, up from 17 percent the year before.

Those numbers get quoted as evidence that AI is oversold. That is the wrong read, and for anyone running revenue there is a more uncomfortable finding buried underneath them.

Most enterprise AI budget is concentrated in sales and marketing. Return is lowest there.

You are spending the most in the place it works least. It is worth understanding why before you renew.

What actually fails

The recurring causes are consistent across both studies, and none of them are about the model. Unclear definition of success. Weak data underneath. Poor integration into the actual workflow. Chasing the technology rather than an outcome. Executive sponsorship that fades once the demo is over.

THE BUDGET IS CONCENTRATED WHERE THE PROCESS IS LEAST DEFINED.

MIT’s own framing is the useful one. They call it the GenAI Divide: organisations successfully deploy consumer-style tools that make individual people faster, and fail to build anything that makes the organisation smarter. You end up with quicker individuals and an unchanged company.

That distinction lands hard in a revenue org, because a sales team is mostly a set of undocumented individual methods already. Handing each of those people a tool that accelerates their own private approach does exactly what you would expect. Everyone gets faster at their own thing, and their own things still do not add up to a system.

Why sales is the worst place to start

Back-office automation produces the highest documented returns in the MIT data. That is not because finance is more sophisticated. It is because a closing process is written down.

An invoice has a defined shape. A reconciliation has a defined output. When the process is explicit, an agent can be pointed at it, its work can be checked, and the saving can be counted.

Now look at a typical sales process. Stage definitions three people would describe differently. A CRM where the fields are filled in the way each rep has decided is reasonable. An ICP that exists as a slide. Qualification that happens in someone’s head. Put an AI layer on that and you have automated an opinion.

The output looks impressive, because generated text always does. The scoring model ranks accounts against a definition nobody validated. The summariser produces notes on calls nobody was going to read anyway. The forecasting tool weights a pipeline whose stages do not mean anything consistent.

None of that is a model failure. The model did exactly what it was asked. The instruction was built on an undefined process.

The test before you buy anything

There is one question worth asking about any AI tool in a revenue stack, and it screens out most of them.

Could a competent new hire do this correctly if you handed them the written instructions?

If yes, automate it. The instructions exist, which means the output can be checked, which means the saving is real and countable.

If no, you do not have an automation opportunity. You have a definition problem wearing an automation costume, and buying a tool converts a cheap problem into an expensive one on an annual contract.

This is the same reason pilots die at the budget review. Somebody eventually asks what it produced, and the honest answer is that nobody can separate the tool’s contribution from the noise, because there was no documented baseline to measure it against.

What the 5 percent do differently

The minority that captures value is not buying better models. Boston Consulting Group surveyed 1,800 executives and found only 26 percent had generated meaningful financial value from AI. Morgan Stanley found only 21 percent of S&P 500 companies could cite a measurable benefit at all.

What separates the group that can is boring. They defined the outcome before the build. They picked a process that was already explicit. They measured against a documented baseline. They integrated into the workflow the person was already in, rather than adding a tool the person has to remember to visit.

In a revenue context that translates cleanly. Write the ICP down properly, then let a model score against it. Define what a stage means in observable terms, then let a model check whether a deal has met it. Get the five answers your reps need into usable form, then let a model retrieve them mid-call.

In every one of those, the AI is doing the second half. The first half is the work.

Where it genuinely earns its place

I am not arguing against any of this. Used properly it is the largest leverage available to a small revenue team right now, and the gains are real in one specific category: anything where the pattern is known and the volume is the constraint.

Reading two years of closed business for pattern. Drafting the first version of a sequence from a defined ICP. Turning a call recording into structured fields against a stage model that means something. Pulling the five accounts that match a trigger you have already defined.

Every one of those is a task where a human knows what good looks like and cannot do it at volume. That is the sweet spot, and it is also the shortest path to a number the CFO will accept.

The order

Define the process. Then automate it. Then measure against what it looked like before.

Do it the other way around and you join the 95 percent, having spent a year and a licence fee discovering that the tool was never the problem.