Why AI pilots stall before production, and the test that predicts it

Most AI pilots stall at the production gate because nobody can state the result in dollars the budget approver owns. Gartner found only 48% of AI projects reach production, after 8 months on average, and 49% name demonstrating value as the top barrier. The predictive test: count the assumptions between your pilot metric and a P&L line.

• How many make it: anywhere from 12% to 54% of pilots reach production, depending on the survey and what it counted (Gartner, IDC, S&P Global, Lenovo/IDC).
• How long it takes: 8 months on average from prototype to production (Gartner). MIT NANDA found enterprises took nine months or longer, against 90 days for the fastest mid-market firms.
• Most common cause: the value cannot be demonstrated (49%, Gartner), followed by data and technology readiness (43% each, Informatica CDO Insights 2025).
• The predictive check: the Link-Count Test. If three or more unmeasured assumptions sit between your pilot metric and a P&L line, redesign the metric before the pilot starts.
• Cost of running the diagnosis: about one analyst-day per pilot, using data you already hold.

How many AI pilots actually make it to production?

Somewhere between one in eight and one in two. The published figures disagree because each one counts something different. One survey counts projects, another counts proofs of concept, and MIT counts only deployments with a sustained P&L effect. Put them side by side and never stack them:

The quotable version: the AI pilot-to-production rate is 12% to 54% depending on who counts, and MIT’s widely quoted “5%” is not a pilot failure rate. MIT reports that 60% of organisations evaluated task-specific tools, 20% reached the pilot stage and 5% reached production. All three are shares of organisations, not of pilots, so dividing one by another does not give a pilot success rate. MIT itself calls the figures “directionally accurate”. They come from interviews, not company reporting.

Why is my AI pilot stuck? Symptom, cause and the test that tells them apart

A stalled pilot usually shows one visible symptom, but several different causes can produce it. The table below puts the causes in order of how often they are reported. Each row uses the highest share found in a survey that asked about that cause. Gartner’s 49% and Informatica’s figures come from different respondents (AI leaders versus 600 chief data officers), so read the order as approximate, not as one ranked list.

Row 1 is the only row where the pilot can succeed technically and still die. The other five are real engineering or organisational debts. Row 1 is a measurement choice, made on day one, that nobody revisits. For row 4, our page on getting a sales team to actually use an AI agent covers the adoption levers in detail.

The Link-Count Test: the check that predicts a stalled pilot

The Link-Count Test counts the unmeasured assumptions that sit between a pilot’s headline metric and a line on the profit and loss statement. Each link is a step the approver has to take on trust: that self-reported time savings are real, that saved time gets redeployed, that redeployed time produces revenue or avoids a hire. Every link is a point on which a finance team can say no.

This is our decision rule, not a validated model. We know of no dataset that tests it. It rests on two published observations. First, Gartner’s survey ranks demonstrating value above talent, technical, data and trust barriers. Second, MIT NANDA reports that sales and marketing attract investment “because outcomes can be measured easily”. MIT also quotes a Fortune 1000 procurement VP asking how to justify a tool that “won’t directly move revenue or decrease measurable costs”. That is a three-link pilot, described by the person who has to defend it. The four-number AI business case for a CFO is the document that has to carry those links once the pilot ends.

Worked example: two pilots with the same budget, and only one survives the gate

These inputs are illustrative. Substitute your own.

Pilot A, a writing copilot for 150 support staff. Users self-report 25 minutes saved a day. If every link holds: 150 × 25 ÷ 60 = 62.5 hours a day. Over 220 working days that is 13,750 hours, and at a $55 loaded hourly cost it is $756,250 a year. The metric has three links: self-report to real time, real time to redeployed time, and redeployed time to a cost avoided. Let each link hold only half the time and the value is $756,250 × 0.5³ = $94,531. The pilot’s value ranges eightfold on assumptions it never measured, and that range is what the gate meeting argues about for months.

Pilot B, AI appointment setting on a dormant CRM for eight weeks. It produces 60 held sales appointments. Your CRM shows a 20% close rate on held appointments at a $9,000 average deal: 60 × 0.20 × $9,000 = $108,000 in expected closed revenue. That is one link, the close rate, and you can test it by following those 60 appointments through to closed deals within one sales cycle. The link can fail: appointments booked by AI may close below your referral average, which is exactly why you track the cohort rather than assume it. To show the pilot’s results are incremental rather than pipeline you would have won anyway, run a holdout, as set out in how to run an incrementality test on lead gen spend.

Pilot B survives because a pilot that books real appointments has a revenue number on day one, not because the technology is better. This is the class of deployment AI appointment setting belongs to. It is also why LeadsNow can report 50,769+ AI-booked sales appointments since 2017 as a count rather than an estimate. There is an honest concession: MIT NANDA argues this measurability bias sends budgets to sales and marketing while “the highest-ROI opportunities in back-office functions remain underfunded”. A metric that is easy to count is not automatically the most valuable one.

What does running this diagnosis yourself cost?

Running it in-house takes about one analyst-day per pilot. Allow two hours to write out the link count and agree it with finance, half a day to rerun the pilot on 200 uncleaned live records, and an hour to list the write-back systems and name an owner. You need read access to the CRM or ticketing data, and nothing else. The skill that is hard to find is not technical: it is someone senior enough to tell a sponsor that their three-link metric will not pass. Do this before the pilot starts, not at the gate. Our guide to running an AI outbound pilot that can fail covers the stop condition, control group and sample size that go alongside it.

Frequently asked questions about AI pilots reaching production

What percentage of AI pilots make it to production?

Between about 12% and 54%, depending on the study. Gartner found 48% of AI projects reach production. IDC reported 4 of every 33 proofs of concept graduating, according to CIO.com’s March 2025 report. S&P Global found 46% of projects scrapped between proof of concept and broad adoption.

How long does it take to move an AI pilot into production?

Gartner’s survey of 644 organisations found an average of 8 months from AI prototype to production. MIT NANDA reported that top-performing mid-market firms took about 90 days from pilot to full implementation, while enterprises took nine months or longer.

What is the most common reason AI pilots do not scale?

Difficulty estimating and demonstrating value, reported by 49% of respondents in Gartner’s survey. Informatica’s survey of 600 data leaders ranked data (43%) and technology (43%) as the top obstacles to moving GenAI from pilot to production, ahead of people and process (35% each).

Is it better to build or buy if we want the pilot to reach production?

In the MIT NANDA sample, external partnerships with customised tools reached deployment about 67% of the time, against about 33% for internal builds. MIT describes its figures as directional, based on interviews rather than company reporting.

Does a successful pilot mean the AI will work in production?

No. Production changes the data, the users and the population a metric is calculated on. Rerun the pilot on uncleaned live records, and compare production results only against the same segment the pilot ran on, never against a blended total.