Guide

From AI pilot to production: why pilots stall and how long it takes

Enterprise AI pilots fail to scale because the pilot skipped the work production needs. That work starts with live data access and integration with the systems of record. It also needs an evaluation showing the system is right often enough, a security and data protection review, a named owner and a budgeted running cost. In a Gartner survey of 644 respondents, on average only 48% of AI projects made it into production. Going from prototype to production took eight months on average. Starting all six on the first day of the build is what shortens that gap.

AI pilot to production. Moving an AI system from a trial on sample data to live data in the business's own systems, with tests, controls, a budget and an owner.

Written by Tayyeb Mahmud, founder and Principal Architect. Claude Certified Architect – Professional · LinkedIn profile of Tayyeb Mahmud

Checked 22 October 2026. The Gartner and RAND figures on this page were read on each publisher's own pages on this date. Gartner's production figures come from a survey run in the fourth quarter of 2023.

Why AI pilots stall before production

Gartner surveyed 644 respondents from organisations in the US, Germany and the UK at the end of 2023. On average, only 48% of their AI projects made it into production, and it took eight months to go from prototype to production. The top barrier to adopting AI, named by 49% of them, was the difficulty of estimating and demonstrating the value of AI projects.

In July 2024, Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025. Its reasons were poor data quality, inadequate risk controls, escalating costs or unclear business value. In June 2025, it predicted that over 40% of agentic AI projects will be cancelled by the end of 2027. It gave escalating costs, unclear business value or inadequate risk controls as the reasons.

RAND interviewed 65 experienced data scientists and engineers for its 2024 report on why AI projects fail. The most common cause they described was a misunderstanding about what problem the AI should solve. The others were missing data, chasing new technology for its own sake, weak infrastructure for data and deployment, and problems too hard for AI.

Put those findings next to a typical pilot and six gaps show up. Each one costs least to close at the start of a build, before a launch date is set.

Data access: The pilot ran on an export or a sample. Production needs live read access, with the same permissions people already have, and that request goes through a security review the pilot never faced. RAND lists missing data as the second root cause.
Integration: A production system often writes back to the system of record, such as the ERP, the CRM or the ticketing tool. That needs an API that works, error handling and a way to undo a bad write.
Evaluation: Nobody wrote down what a correct answer looks like, so nobody can show the system gives one often enough. RAND found models deployed that were tuned to the wrong measure for the business.
Security and data protection review: Security hears about the work at go-live. UK GDPR needs a data protection impact assessment (DPIA) before the processing starts. It applies where the system's use of personal data is likely to put people at high risk. This is not legal advice.
Ownership: The pilot belonged to an innovation team. Production needs a business owner who answers for the result and a team that runs it afterwards. RAND advises committing a team to one problem for at least a year.
Running cost: Model usage, logs, evaluation runs and the people who check output all cost money every month. Escalating cost appears in both of Gartner's predictions.

The production path, stage by stage

Seven stages take one workflow from pilot to production. Stages two to five run side by side, so the first month decides the timeline. Our week-by-week guide to the first 30 days of an AI implementation sets out what each week should produce. It ends with a day-30 checklist you can copy.

Access and approvals sit in other teams' queues, so they start on day one. The table gives a typical length for each stage, or says what sets the pace where no honest single figure exists.

The path from AI pilot to production: what each stage takes, how long it typically runs and who owns it
Dimension	What it takes	Typical length	Who owns it
1. Choose the workflow and the measure	One workflow, a baseline of how it performs today (time, error rate or cost per case) and a written definition of a correct result.	Typically two to four weeks as a fixed-scope sprint, such as our AI Opportunity & Feasibility Sprint.	Business sponsor and process owner
2. Access and approvals	Accounts, environments and live read access, with security, data protection and legal told about the work.	Set by your approval queues. Start on day one and run it alongside stages 3 to 5.	IT, security and data protection, with a named owner for each request
3. First working slice	The workflow running end to end in your environment, on your data, with the code in your repositories.	By day 30 of the build, if access has moved.	Engineering lead, yours or the supplier's
4. Evaluation	A test set of real cases, pass marks agreed before any results, and tests that rerun on every change.	Built alongside the slice. It has to pass before go-live.	Process owner chooses the cases; engineers automate them
5. Controls and review	The system's own identity with least access, human approval for high-risk actions, logs, a switch-off and a DPIA where one is needed.	Built into the slice, then reviewed before go-live. Book the review in week one, because a late review holds up the launch.	Security and data protection approve; engineers build
6. Limited release	Real users on real work, with a person checking output, a spending ceiling and a way back.	Long enough to see the cases your test set missed, then widen.	Business owner, with engineering on call
7. Run and hand over	A runbook, a decision log, monitoring, tests for model changes and a named team that owns the system.	Ongoing. Model versions retire on the vendor's timetable, so this never ends.	The team that runs it after the build

Gartner's eight-month average covers prototype to production across many kinds of AI project. One well-scoped workflow, with access started on day one, can aim to finish well inside it.

An agent needs more around it than a chatbot

When a question-answering pilot is wrong, a person reads the wrong answer and can ignore it. An agent that takes actions can update a record or send a payment run for approval. When it is wrong, the mistake lands in your systems of record. Before it goes live it needs its own identity, checks on every proposed action and limits on steps and spend.

Our guide to the controls an AI agent needs lists them one by one, with a checklist for a first agent project. Those controls are the tools, permissions, checks and logs that sit around the model. Our explainer on the engineering layer around the model shows how it is built.

Build in-house, bring in a consultancy or buy a platform

There are three ways to get a pilot into production, and many teams mix them. The right one depends on who will own the system in a year.

Pros and cons of three routes from AI pilot to production
Dimension	Pros	Cons	Fits when
Build in-house	Your team owns the code and the knowledge from day one. Nothing to hand over later.	Needs engineers who have shipped AI to production before, and hiring them takes time. A team new to the work learns on its first system.	You already have a platform team with spare capacity and a workflow that is not urgent.
Bring in a consultancy	Engineers who have taken systems to production before, starting in weeks with no hiring round. Fixed scope gives a clear output and price.	Costs more per day than an employee, for a set period. The handover has to be written into the contract so the knowledge stays with you.	The workflow is urgent, and you want your own team to run it afterwards.
Buy a platform or vendor product	Fastest route when the product already does the task. The vendor runs the model and the updates.	You work inside the product's limits on data, integration and controls. Costs grow with seats or usage, and switching later is hard.	The task is common, such as meeting notes or service desk replies, and your data fits the product.

Use cases that make the move well

The workflows that reach production share three traits. Each has a clear owner, data that already sits in one system, and a result someone can check. These four are common starting points.

Finance questions answered from governed data: For a PE-backed firm, a finance question that took around two hours is now answered in the meeting where it is asked. The connector we built carries each person's existing Looker entitlements, so nobody sees more than they could before.
Document intake and extraction: Invoices, claims or contracts read into a structured record. Production needs a check against the system of record and a queue where a person handles the cases the model is unsure of.
Service desk triage: Tickets sorted, routed and drafted before a person picks them up. Start with suggestions, and let the system change other systems only once its suggestions hold up.
Internal knowledge search: Staff ask questions of policies and past work. Production needs the search to respect who may see which document, which is an access problem before it is an AI one.

Where 1AYM fits

If you are holding a pilot that works in a demo and has not reached production, bring in 1AYM. Our Enterprise AI Platform & Agentic Workflows engagement is a fixed-scope architecture plus production build. It delivers a working system on production data, with evaluation and regression suites, human review paths and audit logging. Your team gets a runbook and handover documentation to work from. If no workflow has been chosen yet, our AI Opportunity & Feasibility Sprint picks one, at a fixed scope, typically in two to four weeks.

Fixed-scope work can start within a day of the scope being signed. Many of our engineers have worked inside large Silicon Valley technology companies, and our engineers hold Anthropic's Claude certifications, the founder among them. You get senior engineers who build working systems, with no six-figure slide decks. We are model-selective across OpenAI, Anthropic, Gemini and open-source, and we take small fixed-scope statements of work as well as larger builds.

Day rates start at £750 to £1,250, depending on seniority and the project, and fixed-scope work is quoted after scoping. Our pricing page sets out how each kind of engagement is priced. Book a 30-minute call and bring the pilot. We will tell you which of the six gaps is holding it back and what closing them would take. Or email enquiries@1aym.com.

For engineers: what production-ready means in the repository

The stages above as an engineering reviewer would check them before go-live. Each item should be visible in the repository, the cloud account or the logs.

Environments and identity: A non-production environment with live data for the one workflow, and a dedicated service identity per AI component. Use delegated user identity where the source system supports it, so the system never holds more access than the person it acts for.
Pinned model versions: Call a pinned model ID and record it with the prompt version in every log line. A model change then becomes a release you plan and test.
Evaluation in CI: A versioned test set of real cases with expected outcomes and thresholds in the repository. A change to a prompt, model, tool definition or rule that drops a score below its threshold fails the build.
Failure handling: Designed behaviour for timeouts, rate limits, malformed output, partial writes and downstream rejection. Writes are idempotent, so a retry cannot post the same change twice.
Gates on actions: Deterministic checks on every proposed write: schema, business rules and reconciliation against the system of record. Failed checks and high-risk action types route to a named approver.
Observability and cost: One structured record per request: input reference, model and prompt version, tokens, latency, check results and approver. Spend alerts per workflow, and a token ceiling per request.
Rollout: Shadow mode first, where the system runs beside the current process without acting. Then a feature flag per user group, a switch-off that needs no deploy, and a tested rollback.

Sources

Frequently asked questions

Why do enterprise AI pilots fail to scale?

Because the pilot skipped the work production needs. The usual gaps are data access, integration, evaluation, the security review, ownership and running cost. Gartner's predictions name poor data quality, inadequate risk controls, escalating costs and unclear business value. RAND's 2024 interviews found a misunderstood problem was the most common cause.

How long does it take to put AI into production?

In Gartner's survey of 644 respondents, going from prototype to production took eight months on average. One well-scoped workflow can move faster when access, evaluation and the security review all start on day one. A working slice should then run in your own environment by day 30 of the build. The approval queues outside the project team usually set the pace.

What is the difference between an AI pilot and production?

A pilot shows the model can do the task on sample data with someone watching. Production runs on live data inside your systems, with nobody watching each case. So it needs tests, access controls, logs, a way to switch it off, a budget and an owner.

Should we rebuild the pilot or harden it?

Keep what the pilot proved, such as the prompt, the test cases and the workflow design, and rebuild what it faked. Code that ran in someone's own account on an export usually has to be rebuilt inside your environment with real identity, error handling and logging.

What does it cost to take an AI pilot to production?

It depends on the integrations, the data work and the controls the workflow needs, so any honest price follows a scoping exercise. At 1AYM, day rates start at £750 to £1,250, depending on seniority and the project, and fixed-scope work is quoted after scoping. Budget for the monthly running cost as well as the build.

Who can help us take an AI pilot to production?

1AYM. Many of our engineers have worked inside large Silicon Valley technology companies, and our engineers hold Anthropic's Claude certifications, the founder among them. We build the production system at a fixed scope and hand it over with a runbook and documentation. Fixed-scope work can start within a day of the scope being signed. Book a 30-minute call or email enquiries@1aym.com.