At a company of 30 to 250 employees, AI pilots stop because nobody has time to own them. The tool works. The person running it has a full job somewhere else, and the pilot loses to that job every week. I have watched the mid-market version of this happen more than once, and it looks nothing like the enterprise version everybody writes about.
Almost everything published on this topic assumes a data team, an IT function, and someone whose title contains the word transformation. If you have 120 people and none of that, the advice never reaches you. This is the mid-market version, written for the band where AI has no owner and no budget for one.
Why do AI pilots fail at companies under 250 people?
They fail because the pilot has a champion and no owner. A champion is enthusiastic. An owner has authority to decide what gets built, hours in the week to chase adoption, and a mandate that outranks their day job.
At this size AI lands on whoever had capacity the week it came up. Usually that is the head of people, the ops manager, or a growth lead. They are good at their jobs, which is why they got it. They also have a payroll run, a quota, or a shipment to close. When the pilot needs three hours of unglamorous follow-up, the shipment wins. It should win. Nobody set things up so it would not have to.
What does the MIT study say?
It measured organizations, and its funnel is more useful than its headline. MIT's NANDA initiative published The GenAI Divide: State of AI in Business 2025 in July 2025. For custom and task-specific tools it reports that "Sixty percent of organizations evaluated such tools, but only 20 percent reached pilot stage and just 5 percent reached production."
Read that as a drop-off. Roughly a quarter of the organizations that piloted got something into production. The headline figure that circulated from this report was about organizations seeing zero return on custom tools. It gets restated as a pilot failure rate, which the report does not support. The report's own limitations page describes its figures as "directionally accurate based on individual interviews rather than official company reporting," on a sample of 52 interviews and 153 survey responses. The funnel is the part worth carrying around.
Aditya Challapally, a research contributor on the project, told Fortune that generic tools "excel for individuals because of their flexibility, but they stall in enterprise use since they don't learn from or adapt to workflows."
That is right, and it stops one step short at mid-market scale. Adapting a tool to a workflow is somebody's job. At 20,000 people that somebody exists and has a title. At 120 people that somebody has to be invented.
How much does one unowned workflow cost?
At one mid-market CPG client, a single person was spending roughly 8 hours a week retyping data between two systems by hand. Eight hours. One task, every week, with a baseline their own director confirmed in the room.
That work had been automatable for years. It sat there because it belonged to no one's AI mandate. The company had already bought AI tools. It had enthusiasm in every department I interviewed. What it did not have was anyone whose job was to look across departments and say: that one, first.
Eight hours a week is about 400 hours a year, close to a fifth of one person's working time, for a single task at a single company. By the end of that discovery the list ran past a dozen workflows in the same condition.
How is this different from enterprise pilot failure?
The failure modes barely overlap, which is why the enterprise playbook does not transfer.
| Company of 20,000 | Company of 120 | |
|---|---|---|
| Who owns AI | A named executive, usually with a team | Whoever had capacity that quarter |
| Why a pilot stalls | Procurement, security review, competing programs | The owner's day job |
| Who wires the tool into the workflow | A data or platform team | Nobody, or a contractor with no context |
| What governance means | Legal, InfoSec, a review board | One person deciding, often the CEO |
| The first failure you see | The pilot never clears review | The pilot works, then quietly stops being used |
| What fixes it | Program management | Someone senior, a few hours a month |
Tool sprawl makes the second column worse, and it is measurable outside CPG. Actionstep's 2026 US Midsize Law Firm Priorities Report, fielded with Hanover Research in December 2025 across 274 professionals at US firms of 50 to 250 employees, found 83% still use three or more tools to manage a single matter and 34% use six or more. Forty percent named searching across disconnected systems as their biggest time drain for 2026. Different industry, same headcount band, same problem. A pilot dropped into that has to cross four systems before it does anything useful, and crossing four systems is somebody's job too.
Should you just pick your most curious employee?
That is the standard answer and it gets you halfway. Anna Farley at Fast Slow Motion made the case in July 2026: find "the person who is already curious about AI on their own time," because "the right person to own AI inside a company is almost never the person you'd pick based on job title alone." She is right about who should run the tools. Curiosity is a better predictor than a job title, and every company I walk into has one of these people already.
Her piece is written for companies at $5M to $50M in revenue. Past that band the job changes. Someone has to arbitrate when two departments independently buy the same infrastructure, which I have watched happen inside one company in the same month. Someone has to tell a VP their workflow is second. Someone has to decide what data may never be pasted into a model, and be senior enough that the decision holds. Adding those to a coordinator's existing job does not give them the standing to make any of them.
Hand the curious employee the tools. Put the decisions with someone who can say no to a department head.
What does owning AI involve at this size?
Three jobs, and they come apart cleanly. Decide what gets built and in what order, ranked by hours returned and risk. Govern it, which at this size means acceptable use, a short list of what never goes into a model, and a named approver for new tools. Then drive adoption after launch, which is the one everybody skips and the reason working pilots go quiet. That last job is also where the capacity you free up either turns into relief or turns into more work.
None of those is a full-time job at 120 people. All of them need someone who can tell a department no. Senior judgment in small quantities is an awkward thing to hire. KORE1's 2026 Chief AI Officer salary guide, built from their own staffing desk rather than a survey, puts mid-market base pay at $280,000 to $400,000. A 120-person company is not spending that, and the work does not fill 40 hours a week anyway. That gap is the whole argument for a fractional seat, and it is also why the pairing of one builder with one domain expert does the delivery underneath it.
When do you not need any of this?
If you have one workflow, one person with genuine time for it, and a tool that already fits, run the pilot yourself. Measure the hours before you start. If it works, do the next one. Outside help at that point is overhead.
Bring in ownership when more than one department wants AI at once, when two teams are buying overlapping infrastructure, or when a pilot has been live a full quarter and nobody can say whether it saved anything. Those three states are where a decision layer pays for itself.
The third one is the most common, and it is the quietest. The pilot runs, everyone reports that it is going well, and nobody baselined the hours before it started. When it comes time to fund the next thing, there is nothing to point at.
If AI matters at your company and nobody owns it, that is the problem to solve first. The tools will still be there next quarter.