Moving AI pilots into production: why they stall and how to get them live

Your AI pilot worked. It handled the task in the demo, a few people use it, and the business case looks real. Six months later it still sits outside the way your team does its job.

This guide is for heads of product, CTOs and innovation leads at UK scale-ups and enterprises in that position. It covers why AI pilots stall, what production means in practice, and a phased path from a proof of concept to a system your team runs every day. The question of who should own a pilot has its own essay, Your AI pilot worked. Now who owns it?. This piece is the practical how-to.

Why AI pilots stall before daily use

A pilot proves that a model can do a task. It does not prove that the business can run that task with AI every day. The gap shows up in four places.

• Nobody owns it. The pilot belonged to a project, and the project ended. Nobody answers for whether it is used, whether it is right, or what it should do next.
• It was built off to the side. The pilot runs in a separate tool, a notebook or a chat window, so people have to leave their real workflow to use it. When the week gets busy, they stop.
• Nobody maintains it. Models change, data changes and business rules change. A pilot with nobody to update its prompts, tests and integrations gets worse without anyone noticing, and people stop trusting it.
• The knowledge sits with one person. One engineer or one keen analyst understands how it works. When they move on, the pilot becomes something nobody wants to touch.

None of these is a model problem. Each one is about where the system lives and who works on it, so a newer model will not rescue a stalled pilot.

What "in production" actually means

An AI pilot is in production when four things are true.

• It is part of daily work. It reads from and writes to the systems people already use, such as the CRM, the ticketing tool or the approval workflow, with the same sign-in and permissions as everything else.
• It is owned. A named person decides what it should do, approves changes, and answers for it when it gets something wrong.
• It is maintained. Someone updates it when the model, the data or the process changes, and an evaluation suite shows whether each change made it better or worse.
• It is measured. The team agreed a measure before launch, such as time saved, cycle time or error rate, and someone reports it after launch. That measure is where any claim about AI ROI has to come from.

The Gemini Sales Sidekick that &above built for Google Cloud shows what the first point looks like. The team joined Cloud's Sales Hub, Salesforce and the intranet into one flow behind single sign-on, so sellers ask one agent instead of searching several internal sites. The agent has 5,000 monthly active users. The Google Cloud case study describes the build.

A phased path from pilot to production

Moving a pilot into production does not need a large programme on day one. Three steps keep each commitment small, and each step ends with a decision to continue, change course or stop.

A workshop to agree the problem

Put the people who own the workflow, the people who do the work and the people who run the systems in one room. Agree which task the AI does, where in the workflow it sits, what good looks like, and who will own it. If your pilot stalled, check which of these questions it never answered.

Sage started this way. A one-day AI Launchpad workshop with senior stakeholders from across Sage identified the opportunity, and the work became a working prototype of the Creative Intelligence Agent on Gemini and Google Cloud.

A proof of concept against real work

Rebuild or extend the pilot against real data and the systems it will run in, not a sample in a sandbox. Write the evaluation tests first, so the AI PoC answers one question: can this do the task well enough to become part of the workflow? You leave with a working prototype and a clear go or no-go on the business case.

For Tesco Retail Media, the team built an evaluation suite that tested Gemini against real approval rules and workflows before building the product. That validation set up the next phase, described in the Tesco case study.

An embedded squad to ship and run it

When the proof of concept says go, put engineers inside your team to build the production version. They work in your repositories, your standups and your production environment. They build alongside your people, so the knowledge stays with you when the engagement ends.

At &above, we hire the AI engineers, we manage them, and they work embedded in your team. An engagement starts with engineers on one scoped workflow and grows into a full squad with product management and design once the scope proves out.

At Tesco, our embedded engineers shipped the AI-native creative studio that cut creative compliance reviews from up to four weeks to days. The compliance engine sits inside the existing approval workflow, and a prompt and model configuration layer lets the team test rule updates and check model performance over time.

Who should do the work

There are three common options. Match the choice to the gap you have.

• Hire your own AI engineers. This works when you know what to build and can wait out the recruitment timeline. You also take on managing them. Should you hire an AI engineer? covers when a hire makes sense.
• Bring in forward deployed engineers. Senior engineers from a partner work inside your team on a scoped workflow. This fits when you know roughly what needs building and the gap is engineering capacity or AI expertise. What is a forward-deployed engineer? explains the model, and our forward deployed engineers page explains how we run it.
• Add product and design in a full squad. When the harder question is what to build, add product management and design to the same embedded team. For agents that act inside your own systems, our AI agent development work combines agent design, systems integration and guardrails in one embedded squad.

The options combine. A partner's engineers can get the first workflow live, and your own hires can take it on once the system and the use case are proven.

Budgeting the move without a big bet

The cost of moving a pilot into production depends on scope, integrations and the shape of the team, so this guide gives no figures. How much does AI product development cost? explains what moves the number. Before you compare quotes, weigh three cost drivers that a quote rarely shows.

• Recruitment time. Senior AI engineers take time to find and hire, and the pilot stands still while you search. Count the months of delay as well as the salaries.
• Management load. Someone has to set direction, review the work and keep the engineers productive. If that person already runs a team, something else slips.
• Maintenance risk. A system that nobody maintains gets worse. Budget for evaluation, prompt and integration work after launch, as well as the build.

To keep the commitment small, fund one step at a time. Pay for the workshop, then decide on the proof of concept. Pay for the proof of concept, then decide on the build with its result in front of you. When a partner hires and manages the engineers, the recruitment time and the management load sit with the partner, and the build starts with one scoped workflow.

How to tell it is working

A pilot has become a capability your team can run and improve when you can answer yes to these questions.

• Do people use it inside their normal tools without being reminded?
• Can you name the owner, and do they have time set aside for it?
• Has someone other than the original builder changed it and shipped the change?
• When a model, a rule or a process changed, did the team update the system and run it against its tests?
• Is the measure you agreed before launch moving, and does someone report it?

If an answer is no, the pilot is still a pilot. Go back to the step that deals with it. A missing owner is a workshop question. A missing integration is a proof-of-concept question. Missing maintenance is a question about who does the work.

Frequently asked questions

Who can help us get an AI pilot into production?

&above can. We hire the AI engineers, we manage them, and they work embedded in your team, starting with a workshop and a proof of concept before a squad builds and runs the production system.

Why do AI initiatives stall at the pilot stage?

They stall because nobody owns the pilot, it runs outside the tools people use every day, nobody maintains it, and the knowledge of how it works sits with one person. A newer model fixes none of these.

How much does AI product development cost in the UK?

It depends on scope, integrations and the shape of the team, so a figure without a scope is a guess. How much does AI product development cost? explains what drives the cost at each stage.

How do we budget for scaling AI without a high-risk long-term commitment?

Fund one step at a time: a workshop, then a proof of concept with a go or no-go decision, then a build that starts with one scoped workflow. Count recruitment time, management load and maintenance alongside the build cost.

What should I look for when hiring an AI development company?

Look for systems the company has put into production in real workflows, the people who will do the work, a small first step before a long programme, and a plan for who owns the system after launch. How to choose an AI consultancy lists the questions to ask before you sign.

If your pilot works and is not yet part of daily work, talk to us about the smallest next step.