Your pilot worked. That is what makes the next decision hard.
A pilot that fails is cheap to read: wrong tool, missing data, unwilling group. A pilot that succeeds and then goes nowhere costs more, because it hands you a positive result nobody can repeat.
Most writing about getting an AI pilot to production treats this as an engineering problem. For a 50-to-500-person company with AI licenses and no AI function, AI enablement is not an engineering problem. The pilot was a handful of people applying a model to their own work, and production means a colleague who was not in the room doing the same thing in week six of a busy quarter.
Orgvue’s 2026 survey of 1,163 senior decision makers found 43% of organizations with AI projects stuck in pilot and 35% reporting outright failures (Orgvue, May 2026). The stuck group is the bigger one, and it is the harder one to be in, because nothing visibly went wrong.
What “pilot to production” means when the pilot was a behavior
Search the phrase and you get one answer, repeated. Pilots stall, the ranking pages say, because the data was never production-ready, the model was never integrated with core systems, there was no evaluation harness, and no governance body had signed off. Fix those gaps and you cross the chasm.
Every item on that list is an asset a company buys or builds, and the list explains a pilot that did not work. It cannot explain yours.
Your pilot worked, people liked it, the write-up was honest, and four months later two people still do the thing and nobody else has started. No data pipeline caused that.
For a behavior, production has a four-part test, and all four parts have to hold at once:
- Someone who was not in the pilot does the task the new way.
- They do it without a nudge from whoever ran the program.
- They do it on a normal week, with the normal workload.
- A manager would notice if they stopped.
Most pilot write-ups prove none of these. They prove that eight supported people, on a chosen task, did better work than they would have otherwise. That is a real finding about the tool, not about your company.
The five conditions a pilot installs and a company cannot copy
The trap in one sentence: the five things that make a pilot succeed are the five that cannot be reproduced at scale, and a well-designed pilot installs all five on purpose.
Self-selection. Pilot groups are volunteers, or they are the people a manager named because they would take it seriously. You sampled the willing end of your company. A rollout draws from the middle, where the same instruction meets someone with a full calendar and no particular interest.
A facilitator in the room. When a pilot participant got a bad output, somebody who knew why was ten feet away. After rollout, the recovery path is a colleague who is also behind on work, so the second bad output ends the attempt rather than starting a question.
A curated task. Pilots run on a task chosen because a model is good at it, with inputs that someone cleaned up first. Ordinary weeks hand people the messy version of the same task, which is where the method either holds or quietly gets dropped.
Protected time. Half a day came off somebody’s calendar for the session, and often more for practice. Nobody clears time for the 40th user. The new method costs twenty extra minutes on the first hard document, and twenty minutes is the whole decision.
Novelty. Being early is motivating in a way that being 200th is not. The late adopter is not offered something new; they are asked to drop a method that already works for them.
None of this means the pilot was run badly. A pilot’s job is to find out whether the tool can do the work, and loading it with favorable conditions is the right way to answer that question. The error comes later, when a result about the tool gets read as a result about the organization.
What the firm-level evidence says about pilots that worked
The cleanest measurement of this gap is not a vendor survey. Anders Humlum and Emilie Vestergaard linked large-scale adoption surveys to Denmark’s administrative labor records and published the result as NBER Working Paper 33777, issued May 2025 and revised in March 2026.
Their finding has two halves that most commentary separates. Most employers in exposed occupations had adopted chatbot initiatives, workers reported productivity benefits, and new AI-related tasks were widespread. Using difference-in-differences, the authors then estimate precise null effects on earnings and recorded hours at both the worker and the workplace level, ruling out effects larger than 2% two years after ChatGPT launched.
Adoption, enthusiasm, and reported gains at the individual level; nothing at the level of the firm. That is the pilot trap measured across a national labor market rather than one company.
The same paper names what did move: employers absorbed AI through task reorganization, including new work in content generation, oversight, and integration. The structure of the work changed where anyone changed it. Where nobody did, use continued and output did not.
Two notes on the numbers quoted alongside this. The MIT NANDA figure that 95% of generative AI pilots showed no measurable profit impact is the most repeated statistic in the category, and its method has been disputed often enough to read it as a direction rather than a measurement.
The enterprise surveys also sample large: Orgvue’s North American and UK panels are organizations of 2,000 employees or more. The stall pattern carries down to a 180-person company, the remedies do not. Orgvue found 57% deployed AI because competitors had, a fair description of how pilots get authorized with no decision attached.
Why a second pilot is the most expensive answer
When a positive pilot leads nowhere, the common response is another pilot: different team, different tool, cleaner scope. It feels like progress and it is the costliest move available.
A second pilot re-tests a question the first one already answered, and it gives the same answer, because it reinstalls the same five conditions. AI enablement services are quoted in ranges rather than list prices, and hands-on workshop engagements sit in a market range of roughly $2,000 to $15,000 by cohort size and session count. The re-run is a five-figure confirmation of a result you already hold.
There is one case where a second pilot is the right call: the first produced no named deliverable. If nobody can point to a document they made with the model and say what changed about it, you have an impression rather than a result. The rules for reading a pilot result before you scale it cover how to set the second one up so it produces a decision.
Short of that, what you are missing is not evidence. It is a standard: one named task, one written method, one manager who checks. That is a cheaper purchase than a pilot, and the one most proposals leave out.
Five questions to ask before you scale
Each question maps to one condition from the pilot. A “no” tells you what to buy, and the answer is never another pilot.
| Question | If the answer is no |
|---|---|
| Will the next 40 people include the reluctant, not just the willing? | Budget for the reluctant: shorter sessions on their own documents, run by their own manager |
| When someone gets a bad output in week six, who do they ask? | Name two internal practitioners and give them hours, not a title |
| Does the method hold on the messy version of the task? | Re-run it once on your worst inputs before you write the method down |
| Has anyone cleared time for people who were not in the pilot? | Put the first month’s practice into working hours and say so in writing |
| Would a manager notice if the new method stopped? | You do not have a standard yet, and that is the whole gap |
Run these with the people who were not in the pilot. A program owner answering on their behalf gets five yeses and learns nothing, which is also how license dashboards end up overstating what a team can do.
Who owns the move from pilot to habit in an AI enablement program
Ownership has to change hands, and that is where programs break.
A pilot is owned by whoever ran it: the ops lead, the chief of staff, the person doing this as a second job. They can convene, schedule, and report. What they cannot do is tell a function to stop producing a deliverable the old way. Only a line manager with authority over that work can, and in most stalled programs no line manager was ever in the room.
Orgvue’s chief executive Oliver Shaw puts it at the level of the whole deployment: “Failed AI deployments are not a technology problem; they’re a workforce design problem.” The survey also found 84% of leaders saying they should have a deployment roadmap with clear return targets, a large majority describing something they have not built.
The handoff is a short conversation with a specific ask. One manager, one task their team does weekly, agreement that from a named date the expected method includes the model, and a commitment to ask for the working in review the way they already ask which report a number came from. The sequence underneath that conversation is laid out in the five stages from pilot to default behavior.
What to fund after the pilot: the AI enablement line items that matter
The Danish data points at one thing: what moves is the structure of work. Fund the work of changing it rather than another round of instruction.
Four line items, in the order they pay back:
- Thirty days of supervised practice inside real work. Short clinics on live documents in the month after the sessions, where the method meets the messy version of the task and someone is there when it goes wrong.
- A half day with line managers. What to ask for in review, what good working looks like, how to hand back work done the old way. The cheapest item in corporate AI enablement and the first one cut from a proposal.
- One written method for one named task. A page in the same place as the template people already open. “First-draft client updates are written with the model against the account file” is a method. “We use AI in client service” is a sentence.
- Two internal practitioners with hours. The people a colleague asks in week six. They are what keeps the method alive after the people who learned it move on.
External cost peaks during the pilot and falls away, while the items that decide whether the money worked are the cheap ones at the end. A proposal whose price curve runs the other way is selling delivery days, which is the argument behind treating a rollout as a sequencing problem before a budget problem.
For most mid-market teams that is the whole program, and it is bounded. If a workflow needs a system rather than a method, that is another purchase in another bracket, with custom automation work running roughly $25,000 to $150,000 by scope. How to choose between them runs through the playbook from license rollout to measurable team capability, and the case for funding capability rather than more licenses is made in the argument against mistaking licenses for capability.
Frequently asked questions
Why do successful AI pilots fail to scale?
Because the conditions that made the pilot succeed cannot be copied. Volunteers, a facilitator in the room, a chosen task, cleared time, and novelty all raise a pilot’s chance of working, and none of them exist for the 40th user in an ordinary week.
What is the difference between an AI pilot and production use?
A pilot proves a model can do a task under favorable conditions. Production means someone who was not in the pilot does the task the new way, unprompted, on a normal workload, and a manager would notice if they stopped. All four conditions, not any one of them.
What percentage of AI pilots make it to production?
No single credible number exists, and the widely quoted ones measure different things. Orgvue’s 2026 survey found 43% of organizations with projects stuck in pilot and 35% reporting failures. Those figures come from a panel of large organizations, 2,000 employees or more in its biggest markets, so read them as a pattern rather than a rate for your size.
How long should an AI pilot run?
Weeks for the sessions, then a 30-day window you supervise. The pilot answers whether the tool works; the month after it answers whether the method survives without you in the room. Programs that end at the session boundary measure only the first question.
Should we run a second pilot if the first one stalled?
Only if the first produced no named deliverable. A positive pilot with a real output has already answered the question a second pilot would ask, at a market cost of roughly $2,000 to $15,000 per engagement. What a stalled program is missing is a written method and a manager who checks it.
Does a pilot with volunteers tell us anything useful?
It tells you what your most willing people can do with the tool, which is worth knowing and is the upper bound rather than the average. Treat the result as a ceiling, and discount it before you plan a rollout against it.
Who is supposed to own the move from pilot to habit?
A line manager with authority over the work. A program owner can run sessions and report on them, but changing what a deliverable is supposed to look like takes someone who can hand work back. Programs with a single owner throughout stall where that authority has to transfer.
How do we know whether a pilot result will survive a rollout?
Ask the five questions above of people who were not in the pilot. Will the next group include the reluctant? Who do they ask when an output is wrong? Does the method hold on messy inputs? Has anyone cleared their time? Would a manager notice if it stopped?
What should we budget for after the pilot?
Supervised practice for 30 days, a half day with line managers, one written method for one named task, and hours for two internal practitioners. Those four items cost a fraction of the pilot and decide whether it mattered.
Can we skip the pilot entirely?
Often yes, if people already use these tools informally and can show you work they produced with them. Personal use is evidence that the tool fits your material. It is no evidence that the method survives a busy quarter, which is the question the month after a pilot answers.
Key takeaways
- A pilot that succeeds and goes nowhere is the common case, not the unlucky one. Orgvue found more organizations stuck in pilot (43%) than reporting outright failure (35%).
- Five conditions make pilots work and none of them scale: self-selection, a facilitator present, a curated task, protected time, and novelty. A good pilot installs all five deliberately.
- Firm-level evidence separates adoption from results. Humlum and Vestergaard find widespread adoption and reported productivity benefits alongside precise null effects on earnings and hours, ruling out effects above 2%.
- The same study names what does move: task reorganization. Fund the work of changing what the work looks like, not another round of instruction.
- A second pilot is the right answer only when the first produced no named deliverable. Short of that, the missing piece is a written method and a manager who checks it.
If you are holding a positive pilot and no habit, bring it to a free AI-readiness call. We will run the five questions against your own result and tell you which of the four line items your next dollar belongs in.
Arthur Wandzel
