Why most AI pilots die at week three

I went in expecting to write about failed pilots. The sourced data told a different story. The pilot does not blow up. It just never becomes anything. Here is what the numbers say, and what to do about it.

Most trials never start

Use is rising, but the smallest shops are not moving.

The Census Bureau's survey put national AI use at 19.8%, with 20% to 23% of businesses expecting to use it within six months. Firms with 4 or fewer employees sit under 20%, while firms with 250 or more employees sit at 37%, by the same Census survey broken out by size. Firms with fewer than 20 employees showed no significant change over the survey's collection window, the Census Bureau reported. Small shops are not trying and quitting. Most are not trying.

The gap between people and companies is wide. About 18% of US firms had adopted AI while about 41% of workers reported using generative AI for work, by the Federal Reserve Board's comparison of the Census survey with the Real Time Population Survey. People try tools. Businesses mostly do not run them.

It is a tool, not a system

Most businesses that use AI use it in a few spots and change nothing else.

A Census Bureau working paper on the survey's AI supplement found 57% of adopting firms use AI in 3 or fewer business functions. The same paper found 66% of adopting firms use it only to augment tasks, and just 2% reported any headcount cut from it. Among workers, 12% use generative AI daily and 35.2% use it at least weekly, by the Real Time Population Survey of 5,000 to 6,000 people as summarized by the Federal Reserve Board. Add it up and using AI mostly means someone on staff opens a chat window on a Tuesday.

Nobody paid for the setup

Half of small users spent nothing on training, vendors or data.

About 50% of small firms that use AI made zero investment in it, meaning no staff training, no vendor and no data work, compared with about 40% of large firms, by the SBA Office of Advocacy's analysis of the Census AI supplement. Small firms trail large ones most on staff training, an 8.6 point gap, then on hiring vendors or consultants at 5.2 points and on data practices at 4.2 points, in the same SBA analysis. This is the closest thing to a cause I could find. A tool nobody was trained on, that nobody owns, that touches no real data, has no reason to survive week three.

The same Census working paper found firms that spread AI across more functions and invest in operations around it show better commercial performance than shallow adopters. The authors call it a correlation, not a cause. But it points the same direction as the SBA gap.

The big ones quit too

Two famous reversals took a year or more, not three weeks.

Klarna said its AI assistant handled two thirds of customer service chats in its first month and did the work of 700 agents, according to Fortune's report on the company's own disclosures. Klarna then hired no human workers for about 12 months under that plan, Fortune reported. After that its CEO told Bloomberg that cost had been too predominant a factor, the result was lower quality, and the company began recruiting humans again, per Fortune.

McDonald's ran an automated voice ordering test at drive thrus with IBM for roughly 3 years before switching it off, Restaurant Dive reported from the company's statement. An analyst report had called the technology underperforming well before the shutoff, per Restaurant Dive. McDonald's said voice ordering would still be part of its future, in the same report. So even the giants, with training budgets and data teams, took years to decide.

What stays human

Klarna's lesson is the clearest one on the record.

Klarna's reversal was about the quality of human contact, not the cost of the software, by the CEO's own account in the Fortune report. The upset customer, the escalation, the call that does not fit the script, those went back to people. Klarna even piloted freelance support roles at 400 Swedish krona, about $41, per Fortune. In a service business that is the estimate that turns into a fight, the customer who wants the owner and the sub who did not show. Keep a person on those.

What we got wrong

The headline is a frame, not a measurement.

I could not find a primary survey with a published method that reports how many pilots reach production. The numbers that exist come from vendors or analysts, so I am not printing a failure percentage. Nothing I found measures a week three timeline either. The two named reversals ran about 12 months and about 3 years, by Fortune and Restaurant Dive.

For the smallest firms, the top reason for not using AI is not a failed trial. Nearly 82% of firms with fewer than 5 employees that do not plan to use it say it is not relevant to their business, while 6.7% cite lack of knowledge and 6.3% cite privacy, by the SBA Office of Advocacy's analysis of the Census AI supplement. So the honest headline is that most small business AI never becomes a system in the first place. It is a light tool that nobody set up to hold a job.

How to make it live

Give it a job, an owner and real data before day one.

Pick one job with a miss you can count, like a missed call or a late invoice. Put a name on it, one person who checks the output every day for the first month. Feed it the real customer list and the real price sheet, not a sample. Tell the crew what it does and what it does not do.

That is the whole plan. The SBA gap numbers above say training, a vendor and data work are exactly what small firms skip, by the SBA Office of Advocacy's analysis of the Census AI supplement. The Census working paper says breadth and operations investment go with better results, by its analysis of the same supplement. Do the boring three things and the pilot has a reason to be there at week four.

If you would rather not do the setup yourself, book the $199 consult and we will audit your front office first. We look at the calls, the invoices and the jobs you cannot fill, then tell you what to automate and what to keep human. The audit comes before any install.