- Two Ways a Pilot Gets Its Task
- Why Demo-Picked Pilots Stall Before Production
- How to Move an AI Pilot to Production From the Team’s Own Work
- What to Do With a Pilot That Has Already Stalled
- One Housebuilder, Five Pilots
- Moving an AI Pilot to Production: What to Settle First
- Where GrowthNation Fits
- Find the Task Before You Fund the Next Pilot
- Frequently Asked Questions
Quick Answer: Getting an AI pilot to production depends less on the build than on where the task came from. A pilot picked from a vendor demonstration tests a task chosen to show the tool at its best. A pilot picked from the team’s own week tests work people will still be doing next month. Ask the people who do the work first, and let them set the test.
In Orgvue’s survey of 1,163 senior decision-makers, published in May 2026 by the workforce planning software company, 57 per cent said they had deployed AI because their competitors had. In the same survey, 78 per cent said AI projects had either failed (35 per cent) or stayed stuck in pilot (43 per cent). When the reason to start belongs to someone else, the task tends to arrive the same way: from a demonstration or a slide in a board pack. Moving an AI pilot to production then means persuading a team to adopt work it never chose.
Choosing a problem because it looks good in a meeting is one of the ways AI projects fail after the pilot itself has worked. The remedy comes before monitoring, integration or governance: where the task came from, how it was tested and what should change if it passed. A pilot that has already stalled can be read the same way.
Two Ways a Pilot Gets Its Task
In a 2025 survey of 448 enterprises by Omdia, an independent analyst firm, under a third said fewer than 5 per cent of their proofs of concept reach production, and only one in ten put the share over 40 per cent. Omdia blames enterprises and vendors underestimating what deployment involves, not faults in the technology. Part of that underestimate is set at the start of an AI pilot to production effort, when someone picks the task.
| Picked from a demonstration | Picked from the team’s work | |
| Where the task comes from | A vendor demonstration or a board request | The team’s own account of its week |
| Who described the work | The seller, in the seller’s terms | The people who do it each week |
| What it is tested on | Sample cases chosen to show the tool | Cases the team handled last month, awkward ones included |
| Who notices if it is switched off | The sponsor who chose it | The team whose week it changed |
| How it usually stalls | The team carries on the old way beside it | On a narrow gap the team can name |
| What a stall teaches | Usually blamed on the tool | Which part of the task to change next |
A demonstration still has a job: it shows what a tool can do. It cannot tell you which of your tasks it should do, and that question belongs before picking the platform, not after it.
Why Demo-Picked Pilots Stall Before Production
An independent evaluation of the NHS AI Lab by the University of Edinburgh, published in April 2025 and based on 85 interviews, found that the work its awards supported “was influenced by supplier interest” and that frontline staff “had limited involvement”. When an AI pilot to production plan starts from the tool like this, four things tend to follow.
- The seller described the task, not the team. A demonstration shows the textbook version of a job. The real version includes checks and judgement calls that never get mentioned in a sales meeting, so the pilot automates a task that exists mostly on paper.
- Nobody on the team asked for it. Running the pilot becomes one more job, so people keep the old way going beside it. NHS England’s lessons report on its AI Award found projects “were led by the technology suppliers rather than adopting sites”. Team buy-in comes far more easily when the task came from the team.
- Success was defined by the demonstration. The pilot is judged on whether the tool does what was shown, not on whether anyone’s week changed. In the NHS programme, the independent evaluations “were commissioned after projects had been selected”.
- Nobody planned to keep paying for it. A pilot funded as an experiment has no line in next year’s budget. Many sites in the Edinburgh evaluation “had no transition plans” to buy the technology once funding ended.
How to Move an AI Pilot to Production From the Team’s Own Work
Researchers from three US universities spent six months embedded in one company’s security operations centre. A vendor’s AI query builder could not see the organisation’s own logs, so it could only produce queries like those in its documentation. Tools built with the analysts, around problems they raised, stayed in use. It is a single-site preprint and the researchers built the tools themselves, but the route it shows for an AI pilot to production is easy to follow.
Ask the People Who Do the Task
Ask everyone involved, not only the manager, which part of their week they would hand over first, what goes wrong and which workarounds they rely on. The answers name the task in the team’s words, including the small repeated jobs nobody raises in a planning meeting. A manager’s list and a team’s list often differ, and the pilot should come from the second.
Let the Team Set the Pass Mark
The people who do the work should pick the test cases, including the awkward ones, and agree what good enough looks like before anything is built. Record how long the task takes today, so the result has something real to beat. A pass mark set by the team is harder to argue with than one taken from a vendor’s sample.
Decide Which Manual Step Stops If It Works
Production means the old step stops. If the team keeps doing the task by hand beside the new one, nothing has moved. Agree in advance which step is dropped if the pilot passes, and check how it connects to the surrounding process so nothing downstream breaks. A pass then ends in a changed week, not another meeting.
What to Do With a Pilot That Has Already Stalled
The part of a pilot that survives is often the part nobody planned. In one case study from Dublin City University, which followed a European equipment maker’s pre-sales team for 18 months, the use that stuck came from staff experimenting with the tool beyond its original purpose, while the later uses leadership envisioned were still plans. When an AI pilot to production effort has stalled, that points to three options.
- Narrow it to the part people use. Find the piece anyone opens without a reminder, keep that, and drop the rest.
- Re-run it on a task the team named. Keep the tool if it still fits, but swap the demonstration task for one from the team’s own list.
- Stop it, and write down why. Some tasks need a plain rule rather than AI, and some need nothing at all. A short note, kept somewhere central, saves the next team repeating it.
A stalled pilot is evidence about the task, not only the tool, and it should shape the next decision about what moves first.
One Housebuilder, Five Pilots
Take a regional housebuilder with around 500 staff across a dozen live sites. Its board had approved two pilots after vendor demonstrations, and neither had made the move from AI pilot to production. Asking the planning, commercial and sales teams which part of their week they would hand over produced three more.
| Pilot | Where the idea came from | What the team said it needed | What happened | What changed |
| Homebuyer aftercare chatbot | A vendor demonstration | Defect reports logged with photos and a site visit booked | Answered questions few buyers asked | Narrowed to logging defects and booking visits |
| Site progress reports written from photos | A board request after a demonstration | Nothing new; site managers already report weekly | Nobody read the reports | Stopped |
| Planning condition tracker | The land and planning team | Conditions tracked across sites, not in one planner’s spreadsheet | Used weekly from the first month | Rolled out to every live site |
| Subcontractor document checks | The commercial team | Insurance and accreditation expiry dates chased before work starts | The old chasing list was dropped | In daily use |
| Plot handover packs | Sales progressors | Warranty papers and manuals gathered for each plot | Tested on last quarter’s handovers | In daily use |
The two demonstration pilots were narrowed or stopped, and the three team ideas reached daily use. The aftercare chatbot only reached daily use once it was rebuilt around what buyers actually report. The planning condition tracker, once a single planner’s spreadsheet, now runs across every live site.
Moving an AI Pilot to Production: What to Settle First
Before the next pilot is funded, settle six things with the team that will live with it.
- Trace every idea to a person who does the work. If it traces back to a demonstration, go back to the team.
- Ask more than the manager. The people closest to the task know its exceptions.
- Keep the vendor’s sample cases out of the test. Use last month’s real cases instead.
- Agree the pass mark before anyone builds. Write down the scale conditions too, so a pass leads somewhere.
- Put the running cost in next year’s budget. A pilot nobody plans to pay for has nowhere to go when it passes.
- Set the date you will decide, and tell the team. A fixed date makes the result easier to put to the leadership team.
Where GrowthNation Fits
GrowthNation starts with the choice of task. Its AI agent interviews everyone in a team, by voice or text and in confidence, so the task comes from the people who do the work rather than from a demonstration. Agents built from those interviews are served centrally and belong to the organisation, not to one person’s account. Hours saved are tracked per agent in daily use, against the starting count. If a pilot already came from the team and is stalling on integration or security, an engineering partner may be the better help.
Find the Task Before You Fund the Next Pilot
Ask one team which part of last week it would hand over first, then put that answer next to the pilot you are running. Getting an AI pilot to production starts with that list, because a task the team named is one it will miss if it is switched off.
Book a walkthrough and GrowthNation will interview one team, then hand you three things:
- The tasks the team would hand over first, in its own words, exceptions included.
- Where your current pilots match that list, and where they do not, so you can see which to narrow, re-run or stop.
- A working agent, built from a task the team chose and held by your organisation.
Hours are measured against where you started, so the saving can be checked.
Book a walkthrough: one team, one task list, one working agent. No deck.
Join the GrowthNation waitlist. New organisations are taken on in batches, so joining now puts your team in line for the next slot.
Frequently Asked Questions
How do you move an AI pilot to production?
Moving an AI pilot to production starts with a task the team named, not one taken from a vendor demonstration. The people who do the work choose the test cases, agree the pass mark before anything is built and decide which manual step stops if it passes. Budget the running cost and set a decision date before it starts.
Why do AI pilots stall after a vendor demo?
A vendor demonstration shows a task chosen to display the tool at its best, in the seller’s terms. The real version has exceptions the demonstration never met, and nobody on the team asked for the change, so people keep the old way going until the pilot quietly stops.
Who should choose the task for an AI pilot?
The people who do the work should name the task, and leadership decides which suggestions to fund. Asking everyone in the team, not only the manager, surfaces exceptions and workarounds a planning meeting misses.
Should you restart a stalled AI pilot or stop it?
Check which part of the stalled AI pilot anyone still uses without being reminded, and narrow it to that. If the tool fits but the task was wrong, re-run it on a task the team named. If neither applies, stop it and record why, so the next pilot does not repeat it.
How many AI proofs of concept reach production?
In Omdia’s 2025 AI Market Maturity Survey of 448 enterprises, under a third said fewer than 5 per cent of their proofs of concept reach production, 21 per cent put the share at 5 to 10 per cent, and only 10 per cent said more than 40 per cent get through.
How long should an AI pilot run?
An AI pilot should run long enough to cover at least one full cycle of the work, such as a month-end close, a rota or a quarter’s handovers, so the team meets its usual exceptions. Set the decision date before it starts.