Why AI pilots don’t reach production

Most lenders I speak to have now run an AI pilot. Rather fewer have anything in production. The gap between those two states is where most of the frustration in this market currently sits, and it is almost never explained by the technology.

The demo worked. That is usually not in dispute. What happened next is the interesting part.

Three failures that look technical and are not

Nobody agreed who decides. An AI-assisted underwriting tool changes the distribution of decision rights inside a credit function. Most credit policies are silent on this, because they were written when the only actors were people. Until someone answers whether the system can decline, or only recommend, and who owns the outcome when it is wrong, the tool cannot be deployed regardless of how well it performs.

There was no shared source of truth. Pilots are usually run on a clean extract. Production runs on the actual estate — where the same customer exists three times, the asset schedule is authoritative in one system and stale in another, and two teams disagree about what a completed application means. The pilot never encountered this because the pilot was given tidy data.

Nobody owned the exceptions. Every deployment generates cases the system cannot handle. If no named person owns that queue, it grows quietly until someone notices that the automation has created a new manual process rather than removing one.

AI does not fail because the tools are weak. It fails because organisations do not give it the context it needs to act well.

The success case has its own failure mode

There is a second pattern that gets less attention because it looks like progress.

The pilots that do succeed often succeed in isolation. A team builds something genuinely useful, deploys it in their function, and moves on. Six months later the business has several of these — each valuable, each with different data assumptions, different controls, and no shared governance. Nobody is responsible for the collection.

This is how a portfolio of useful tools becomes an unmanaged liability. The individual deployments were all sensible. The absence of a deliberate plan across them was not.

Why mid-size lenders are worse served

Read almost any enterprise AI framework and you will find the same assumed scaffolding: a board-approved AI vision, an AI risk committee, RACI charts, a governance forum, a transformation office. The advice is not wrong. It is written for organisations that already have those things.

A specialised lender with a few hundred people has a compliance lead wearing four hats and a change budget largely consumed by regulatory work. Handed a framework that assumes a programme office, the rational response is to conclude it does not apply and do nothing. Which is exactly what happens.

The gap is not appetite. It is the absence of an on-ramp.

What a proportionate on-ramp looks like

The things that actually determine whether a first deployment reaches production are small in number and mostly non-technical:

• Pick work the front line is already asking for. Enthusiasm from the people who will use it is worth more than theoretical value.
• Scope the path to production before the pilot, not after. Ask what would have to be true to run this for real, and start there.
• Decide the decision rights first. Recommend or decide. In writing. Before building.
• Name the exception owner. A person, not a team.
• Use real data early. The messiness is the requirement, not an obstacle to work around.
• Agree the off switch. What would cause you to stop, and who can make that call.

None of that requires new governance infrastructure. It requires a handful of decisions made deliberately, in advance, by people who have the authority to make them.

The lenders getting this right are not the ones with the best models. They are the ones treating AI as an operating-model change that happens to involve technology — which is a different job, and mostly a management one.