“Our AI pilot produced no results” — what to do in the next 30 days

If your AI pilot produced no results, you are not unusual: CEOs surveyed by IBM in 2025 reported that only 25% of AI initiatives had delivered expected ROI. Spend 24 hours freezing and exporting, 7 days finding out what “nothing” means, and on day 30 choose: fix, re-scope, switch or stop.

• Is it urgent? Rarely. The money is spent. Only the renewal or notice date, the next budget date and your access to the pilot’s data run on a clock.
• Next 24 hours: no expansion, no angry cancellation. Find the notice date, export every log, and write what “nothing” means in numbers.
• Next 7 days: the Readability Check (was the result too small to see?) and the break-point tally (where did 20 real cases stop?).
• Day 30: one of four verdicts from the threshold table below. Stopping is one of them.
• The honest possibility: the use case was wrong. RAND’s test: if a problem is not worth a team for a year, it is not worth committing to at all.

“Our AI pilot produced no results”: what is actually happening

“Produced nothing” covers three different situations. Nothing was measured: there was no baseline, or the instrument changed mid-pilot, so nobody can say whether anything moved. Too small to read: the pilot ran on too little volume for any plausible lift to rise above ordinary noise. A real zero: the counts were large enough to read, and the outcome metric did not move, or fell.

Only the third is a failed pilot. The first two are unfinished experiments, and rescues go wrong when a sponsor treats all three the same way. If you already know it is a real zero and want the cause, the symptom-and-test table in why AI adoption fails: the five causes and the test for each covers the diagnosis. This page covers the 30 days after it. An AI pilot that produced no results has measured nothing, measured too little, or measured a genuine zero, and only the last one is a verdict.

Is a failed AI pilot as urgent as it feels?

Usually not. A pilot that produced nothing has mostly already cost what it will cost, and a single one should not end a career. In the IBM Institute for Business Value CEO study (2,000 CEOs, 33 countries, February to April 2025), only 16% of AI initiatives had scaled enterprise-wide, and 64% of CEOs said the risk of falling behind drives them to invest in some technologies before they clearly understand the value. Many pilots begin without a stated value case, so many end without one.

What does run on a clock:

• The renewal or notice date. An auto-renewal inside 30 days is the one real deadline.
• Data access. When a pilot account closes, transcripts, logs and prompts often go with it.
• Your team’s goodwill. Staff told to keep using a tool nobody defends stop quietly.

Two quiet weeks inside a pilot is noise; a readable zero across the full window is a finding. The only deadline in a failed AI pilot is usually the renewal date, and everything else can wait seven days.

What to do in the next 24 hours

None of this costs money or involves a vendor:

1. Freeze scope. Add no users, volume or budget. Do not cancel yet either.
2. Find the contract dates. Renewal date, notice period, data-return clause. If the notice window closes inside 30 days, ask the vendor in writing to extend the pilot terms to your decision date. What you own when you leave a vendor lists the data and exports to claim.
3. Export everything. Raw logs, transcripts, outputs and timestamps, not a dashboard screenshot.
4. Write the “nothing” sentence. Three numbers: the metric the pilot was meant to move, its value before, its value during. If you cannot fill in the first, the pilot measured nothing, and that is your answer for now.
5. Give the sponsor a date, not a verdict. “A recommendation on the 30th” buys the time the next two steps need.

The Readability Check: was “nothing” too small to see?

The Readability Check is our own rule of thumb, a rough Poisson approximation rather than a statistical test. It asks whether the pilot ran on enough volume for a real effect to show. Multiply the pilot’s volume by your baseline success rate and call the result E: the successes you would expect with no change at all. Treating the count as Poisson, random variation in it is roughly its square root, so any result inside E ± 2√E cannot be told apart from no change.

A worked example with illustrative inputs: 1,500 leads × a 2% baseline booking rate = 30 expected bookings. 2 × √30 = 11, so anything from 19 to 41 bookings is noise. The pilot needed 42 or more (a 40% lift) to show a win at all. “We got 34, so it did nothing” is the wrong reading. The right one is “at this volume, the pilot could not have shown a lift under 37%.”

This is a rough approximation, not a significance test, and it assumes a stable baseline. If your baseline is itself a count from a similar-sized period, widen the band by about 1.4×. One case runs the other way: an actual zero against an E of 5 or more has under a 1% chance of being random. That is a broken chain, not a weak one. To size the rerun, use the volume and stop-condition method in how to run an AI outbound pilot that can fail. A pilot with fewer than about 30 expected successes cannot show any lift smaller than about a third, so its “nothing” is silence, not evidence.

What to do in the next 7 days: the break-point tally

The break-point tally finds where the pilot’s chain stopped. Pull 20 real cases the pilot touched, at random, not the ones the vendor showcased, and mark the first step where each stopped:

1. Input: the AI never received the case, or received it incomplete.
2. Output: the AI acted, but wrongly.
3. Handoff: the output was right, but no person or system acted on it.
4. Outcome: the handoff happened, but the business result did not follow.

If one step holds 12 or more of the 20 (our working threshold, not a published one), you have a single break, and a single break can be fixed. Spend days 3 to 7 changing that one thing, and only that, on a small fresh batch. If the stops spread across three or four steps, no single change will rescue the pilot. The break-point tally turns “the AI pilot didn’t work” into a count of where 20 real cases stopped, and the step holding the most stops is the only thing worth fixing this week.

Day 30: fix, re-scope, switch or stop?

By day 30 you have a Readability Check result, a break-point tally and a one-week test of one fix. Match them to a verdict. The thresholds are our working rules for this decision, not a published benchmark.

If your results match two rows, take the higher one: a cheap rerun comes before a switch. A 30-day AI pilot rescue should end in exactly one of four verdicts (fix, re-scope, switch or stop), each triggered by a count, not by how the pilot felt.

When the answer is that this use case was wrong

Sometimes the pilot was fine and the question was wrong. RAND’s 2024 report, drawn from interviews with 65 experienced data scientists and engineers, recommends committing each team to a specific problem for at least a year: “If an AI project is not worth such a long-term commitment, it most likely is not worth committing to at all.” It also names chasing the latest AI advances “for their own sake” as one of the most frequent pathways to failure.

So apply the one-year test to the use case, not the pilot: would you fund a team to own this problem for a year if AI were not involved? If not, the honest verdict is switch or stop. Use cases that survive a rescue usually have an outcome someone already counts, such as invoices paid, tickets closed or sales appointments booked and held. A booked appointment is countable the day it happens, which is why LeadsNow can state 50,769+ AI-booked appointments since 2017 as a tally rather than an estimate. Whatever you choose next, carry it into the budget round as the four-number AI business case a CFO will sign. If nobody would fund the problem for a year without AI, the pilot did not fail; it answered the question.

What running this rescue yourself costs

Doing it in-house is realistic. Our estimate for one pilot:

• 24 hours: about 2 hours of the pilot owner’s time, plus 30 minutes from whoever holds the contract.
• Readability Check: an hour with a spreadsheet, once you have the baseline rate.
• Break-point tally: 1 to 2 analyst days to trace 20 cases end to end, longer if the logs are poor (which is itself a finding).
• One-week fix test: whatever the fix needs, plus a week of the pilot’s run rate.

The hard skill is not technical. It is someone with the standing to tell a sponsor that a pilot was unreadable or that the use case should go. Without that person, pilots drift into a second quarter with nobody defending them and nobody ending them. Rescuing a failed AI pilot takes about two analyst days plus the fix itself, and the scarce input is the authority to say stop.

Frequently asked questions about AI pilots that produced no results

What should I do first when an AI pilot produces no results?

Within 24 hours, freeze scope, find the renewal and notice dates, and export every log and transcript. Then write one sentence with three numbers: the target metric, its value before the pilot and its value during. If you cannot fill in the first number, the pilot measured nothing.

How many AI initiatives actually deliver ROI?

A minority. In the IBM Institute for Business Value’s 2025 CEO study of 2,000 CEOs, surveyed from February to April 2025, CEOs reported that only 25% of AI initiatives had delivered expected ROI over the last few years, and only 16% had scaled enterprise-wide.

Should we kill an AI pilot that showed nothing after 30 days?

Not on that fact alone. Check the volume first: if the pilot expected fewer than about 30 successes, it could only have shown a lift of roughly a third or more. Stop when a readable pilot (100 or more expected successes) sat inside its noise band and the break-points were spread across three or more steps.

How do I tell leadership our AI pilot failed?

Give a date first, then one of four verdicts: fix, re-scope, switch or stop, backed by the Readability Check and the break-point tally. A stop memo that shows the counts is a result. A pilot left running unexamined is not.

Is it the vendor or our data?

The break-point tally tells you. Stops at the input step point to your data or integration. Stops at the output step point to the tool or its configuration. Stops at the handoff or outcome steps point to your own process, and a new vendor will not fix those.