The last mile of AI is deployment, and it is yours to own

The short answer

A demo proves a model can do the task. Production proves the business can rely on it. Between the two sit integration with real systems, the exceptions the demo never saw, the people who have to change how they work, and someone to run it on a Tuesday night.

That stretch is the last mile, and it is where most of the cost and almost all of the return is decided.

What the last mile contains

Integration. Reading from and writing to the systems where the work actually happens, with their permissions, their quirks and their downtime.

Exceptions. The cases outside the happy path, which in operations are often where the money is lost. Each needs a rule, a review or an escalation.

Adoption. The team using it in their daily work rather than around it. That means training, adjusted targets and a way to report when it is wrong.

Operation. Monitoring quality, handling incidents, updating rules when the business changes.

Why it gets underpriced

Proposals are written around what is easy to specify, and the build is easy to specify. Integration depends on systems the vendor has not seen, exceptions are discovered in use and adoption is hard to scope.

So the proposal prices the visible part, the project discovers the rest, and the budget conversation happens at the worst possible moment. Ask any vendor to show you how they priced integration and operation, separately from the build.

Who should own it at the end

You should. A system that only the vendor can operate is a dependency, not an asset. The healthy path is a handover by design: the delivery team leads at first, works alongside your team as the system stabilises, and then your team runs it with support available.

That handover has to be planned from the start, including who on your side will own it, what they need to learn and which documentation they will receive. Ownership of the code and the infrastructure should be yours from day one.

A test for any proposal

Ask three questions. What happens when the system meets a case it was not built for? Who on my team will run this in a year, and how will they learn? What does it cost to operate at full volume, per month?

Vague answers to any of the three mean the last mile has not been priced, and it will be paid for later.