Why AI Pilots Stall After the Demo

An AI pilot stalls when the demo works but the daily work never does. Everyone liked what they saw in the meeting. Then Monday came, and the pilot hit real files, real handoffs, and a person who could say no.

Cited in this note

50%

In January 2026, Gartner wrote that at least 50% of generative AI projects were dropped after the first test run. The reasons were poor data, weak risk controls, costs that kept growing, and unclear value.¹ We read that list as a plain checklist. If any of those four is blank for your pilot, it is already sliding toward the dropped pile, even if nobody has said so yet.

Our decision guide fits this problem best. The steps that should come before a demo are in our implementation framework. The trouble we see most often is in our data note. A pilot that cannot point to the one place its facts live is not early. It is unfinished, and paying for another demo will not finish it.

We have sat through the awkward meeting that follows a nice demo. The owner feels embarrassed, and the staff are tired of sample data. Someone suggests a new vendor, as if the vendor were the missing piece. The vendor was never missing. What was missing was a decision.

Using AI is not the same as rolling it out

McKinsey’s 2025 survey found that 88% of people use AI regularly in at least one part of their business. A year earlier, that number was 78%.² About one third of companies are rolling AI out widely. Only 23% are rolling out an AI agent system somewhere, and 39% are still trying things out. No more than 10% are rolling out agents in any single part of the business. Among companies that earn less than $100 million a year, 29% have reached the wide rollout stage. Among companies that earn more than $5 billion, it is nearly half.

Those numbers describe a stalled pilot well. Use grew fast, but only a small group of companies can run a tool every day inside one team. A small company should not hear the big-company numbers as a dare. Even the very large firms work in a different way than a dispatcher and an owner do. Your pilot may be stuck because of your own files, and that has nothing to do with how big you are.

In June 2025, Gartner said that over 40% of AI agent projects will be canceled by the end of 2027.³ We read that as a warning about projects that never became a daily habit. It does not say the software is bad. Projects get canceled when the cost, the value, or the checking stays vague after the demo. We would rather you stop a pilot on purpose this month than let it drift into that pile and blame bad luck.

Say a pilot cannot name its data, the person who checks it, and the value it brings. Then it is already in the dropped pile. Turn it into a written routine, or stop it. A third choice, like waiting to see how the next demo feels, is how the same bill keeps coming back.

A worked example

Picture a small electrical company. This story is made up, and we did not measure it. The owner still prices the odd jobs herself. A dispatcher plans the day’s routes. Two technicians send photos from the job site. An office lead turns those photos and the customers’ texts into a quote whenever she can get the owner’s attention.

The company paid for a pilot that turns incoming texts into a tidy intake note. In the demo, the seller pasted in three clean conversations. The note came back with the address, the problem, and a next step. Everyone nodded, and the owner pictured the office lead going home on time.

Monday looked nothing like the demo. One customer changed the light fixture in a later text. Then the same customer sent a photo of the old fixture by mistake and asked if the price included the permit. The job history lived in a group chat that the pilot could not see. The price list was a spreadsheet on the owner’s laptop, named in a way only she could remember. The office lead could not tell if a summary was safe to send, because the note never said which messages it had skipped.

The data was poor in a clear way. The facts lived in many places, and the pilot acted as if they lived in one. Nobody was named to stop a quote that sounded sure but was wrong. The cost started as a small monthly fee. Then it grew when the seller offered a second task while they were already set up. The value had been described as saving the office lead’s afternoon. But nobody had written down what that would look like on a Thursday when the phones were awful.

So the pilot stalled at the handoff, which is where most pilots stall. The demo had no handoff, because it ended at the summary. The real work ended when a person had to put the company’s name on that summary. The office lead stopped trusting it, yet the tab stayed open. The owner kept paying, because canceling felt like admitting the demo had fooled her. It had not fooled her. It had shown her a different company, one with clean message threads and a price list the software could read.

How we’d run it

We do not try to bring a stalled pilot back with a pep talk. We make a short decision in a set order, and we are fine if the answer is to stop.

First, we write down the original promise in one sentence. We use the words the owner remembers from the sales meeting. For the electrical company, it might be that the office lead gets an intake note she can trust from the customers’ texts. If nobody can write that sentence, the pilot never had a job to do. We stop there and say so out loud, because vague hope is not worth another month of payments.

Second, we turn the four reasons projects get dropped into questions, not blame. Is the data good enough to back up the sentence? In this office, can the note see the whole message thread, the right photo, and the price list? Are the risk controls real, meaning we know who may send a quote and who must stop one? Is the cost still the cost you agreed to, or did it pick up a second task? Is the value clear enough that the office lead would notice if it went away? Any blank answer is a good reason to pause the spending.

Third, we pick the person who checks the work before we talk about a better prompt. In this office, the checker is the office lead. The owner backs her up on jobs above the dollar line they already use for quotes done by hand. If nobody can name that person, the pilot is only a toy. We also sit with her and watch her handle one real message thread. Watching can bruise the seller’s story, but it is very kind to the business.

Fourth, we choose a written routine or a stop, and we do it in the same meeting. A written routine says which places the note may pull from. It names the message thread, the photo rule, and the price list, and the office lead still sends every quote. A stop means the subscription ends. The one-sentence promise goes into a folder until the price list lives somewhere more than one person can find it. We do not book a second demo with cleaner samples, because cleaner samples are how the first meeting gave a false picture.

Fifth, we tell the staff about the decision in plain words. If we keep the pilot, they hear what will change on Monday. If we stop it, they hear that stopping was the plan and that it says nothing bad about the office lead. People will keep feeding a dead pilot if they think the owner is ashamed. Shame costs a lot of money, and a clear sentence costs almost nothing.

Common mistakes

Calling the stall a delay is a polite way of not telling the truth. The dispatcher says the team is still rolling it out. A real rollout has a date and a named person. A stall has an open tab and a bad feeling. If the office lead has quietly gone back to the old message thread, you are not late. You have already picked the old way, and you are paying twice, once in money and once in pretending.

Blaming the AI model for facts that live in many places wastes the next month. The summary was shaky because the price list sat in a private file and the photo did not match the text. A newer model will sound even more sure about the same gap. Sounding sure is not the same as knowing. We have watched owners swap tools and repeat the same stall with better grammar.

Letting the seller decide what counts as value is how costs creep up. Time back is not real value until the office lead agrees on which hour comes back and what she will do with it. If she says she will finally eat lunch, believe her. Then check whether the note really gives her that lunch. If the note adds a new review chore that eats the same lunch, the value was only a slogan.

Skipping risk controls because it is just office work is how a bad quote leaves the building. Office work is your voice to your customers. A wrong intake note turns into a wrong visit, a wrong part, and a customer who thinks you never listen. The control has to be a named person. A box in the software that says a human will check does not help if nobody’s day includes that check.

Paying for a fresh demo to avoid the stop talk is the costly mistake. Gartner’s June 2025 forecast says that over 40% of AI agent projects will be canceled by the end of 2027. This habit is what that looks like across many companies. You can be a project that ends with a clear reason, or one that ends later after a longer bill. We would rather you take the shorter bill.

What we’d do

We would score the stalled pilot before we let anyone pay for another demo. That is a strong opinion, and we stand by it. A second demo feels hopeful, but it is usually a way to avoid telling the seller and the staff that the first promise had no data behind it. The decision guide is the page we open in that meeting. If the score says the business is not ready to decide, we still decide. Not being ready is a good reason to stop spending.

We would not comfort a small company with the McKinsey usage number. Saying that 88% use AI regularly, up from 78% a year earlier, only tells you people opened a tool. It does not tell you the office lead trusts a quote. About one third of companies are rolling AI out widely, 23% are rolling out an agent system somewhere, and 39% are still trying things out. No more than 10% are rolling out agents in any one part of the business. That is the honest picture. And 29% of firms under $100 million in revenue have reached wide rollout, against nearly half of firms over $5 billion. So the middle is hard even before you count being small. Being small is not what stalls a pilot. A missing checker is.

We would also do the data work first, every time. That is why we send owners to the data note before we ever talk about prompts. If they decide to go on, the order of the steps is in the implementation framework. It does not call for a new pilot with a new logo. Our bet is simple. A stopped pilot with a written reason makes a healthier company than a live pilot that nobody will sign their name to.