The short answer
Most AI pilots stall because they are set up to prove a tool works, not to change how work gets done. To reach production, a pilot needs a named business owner, a mapped and simplified process to land in, clear points where people review the AI's work, and one number that has to move.
On this page
Key takeaways
- The model is rarely the problem. Ownership, process and measures usually are.
- Fix the process before you automate it, or the AI inherits the mess.
- Decide up front where a person must approve the AI's work.
- Agree the number that must move before the pilot starts.
Why do most AI pilots fail to reach production?
Because the hard part was never the model. MIT's Project NANDA found that 95% of generative AI pilots showed no measurable impact on profit and loss. RAND's interviews with practitioners found most blamed leadership and organisational causes, not the technology. The demo works; the business around it doesn't change.
What are the warning signs a pilot will stall?
- Nobody in the business owns the outcome. IT or a vendor does.
- Success means "it works", not hours saved, errors cut or faster turnaround.
- The process it sits in has never been mapped.
- There is no plan for what happens when the AI is wrong.
- Nobody has budgeted to run it after the pilot ends.
The five questions to answer before you start
- Which process, and who owns it? One process, one business owner with authority to change it.
- What outcome, and how will you measure it? Often hours or cost, but it might be turnaround, error rates or complaints. Record the baseline before you start.
- What is the simplest fix that isn't AI? Remove the steps that should not exist first.
- Where must a person decide? Name the checkpoints and how people will review the AI's work.
- What number must move, by when? Agree the go-live threshold before the first line of code.
Where does human-in-the-loop fit?
Human-in-the-loop means people review, approve or correct the AI's work at set points. It is what makes AI safe to switch on in a real business. Start with people approving everything, measure how often they change the output, and only widen what the agent handles on its own when the numbers say it is ready.
What we learned doing this in a bank
[ADD: one real, anonymised story from the home lending work. What was automated, what stayed with people, what went wrong, and the result in numbers. First-hand detail like this is what ranks and gets cited.]
Frequently asked questions
How long should an AI pilot run?
Long enough to measure the number you agreed, and no longer. A pilot that runs on without a go or no-go decision is drifting.
Should IT or the business own an AI pilot?
The business. IT and partners build and run it, but the owner must be the person accountable for the process it changes.
What is a good first AI use case?
A high-volume, rules-heavy task where a person can easily check the output, such as triaging requests or preparing documents for review.
Got a pilot that's stuck?
A loop starts with a Diagnostic that shows what it needs to land.