After go-live: reuse, cost and who owns the value

The first model in production felt like the finish line. It was the start of a different problem: who reuses it, who pays for it, and who owns the value.

Summary. Getting the first model into production felt like the finish line. It turned out to be the start of a different problem: other parts of the business wanted to use it, and cost and ownership needed to follow that use.

Context

A large regulated bank with a shared machine learning platform and its first enterprise model in production: a transaction categorisation service handling millions of customer transactions a day. The route to production had been built and other data science teams were starting to onboard.

The problem

Then three questions came up that nobody had needed to answer before.

Reuse. Other parts of the business could see value in the model's output. Analytics teams wanted it as a foundation for their own work, and customer processes could use it to make better decisions. Each could have asked for its own version. That would have been slower, riskier and harder to govern than reusing what was already running.

Cost. A shared platform carries its cost in one budget, and the value it creates lands in other people's numbers. As more teams arrived, each needed to see what the platform cost for its own work.

Ownership. When a model serves more than one consumer, someone has to own changes, monitoring and retirement. If the platform team absorbs all of that, it becomes an operations team for everyone else's models and stops building.

My role

Product owner for the platform. I built the cost forecasts and the tenant cost showback. On the newer Azure platform, teams that bring a model own and fund it. The model and its business uses belong to the teams that built and use them.

Constraints

Any reuse had to stay inside the governed route. A second consumer could not become a shortcut past the controls the first one went through.
The platform could not become a line item the business saw only as a bill.
The platform team had to keep building the platform, rather than drift into running every model on it.

Stakeholders

The data science teams who built and owned models. The business areas that used their output. The platform's leadership, who held the budget. And the control functions, who needed every new use to be as well governed as the first.

Decisions

Make cost visible to the team that owns the model. Tenant cost showback lets each team see what its own workloads cost.
On the newer platform, teams that bring a model own it and fund it, rather than the platform carrying it.
Reuse stays governed. A new use of an existing model goes through the same route to production, with approval staying with each control function.
Forecast cost with the roadmap, not after it. The budget conversation should follow the demand already in quarterly planning.

Approach

I built the platform's cost forecasts from the demand in quarterly planning, with a multi-year view, so the budget proposal followed the roadmap instead of arriving after it. I broke the platform's cost down by tenant so each team could see what its own workloads cost. Onboarding already carried responsible AI checks, and teams joining the newer cloud platform came on owning and funding their models. The business areas brought the new uses of the first model. My part was keeping each one inside the governed route and making its cost visible.

Outcome

The first model's output is now the analytics layer for customer and business banking teams, and part of a customer hardship assessment. Teams on the newer cloud platform own and fund their models from the start, and each team can see what its models cost.

The benefit side is measured by the teams that receive it, not by the platform. That is the unfinished part.

Lessons

Reuse is where the return is, and it rarely happens by accident. Someone has to make it the easier option.
Showback made cost something teams could talk about before any money moved.
Write ownership down before the second consumer arrives, not after.
A platform should not try to book the benefit. It should make it easy for the teams who own the benefit to see what it costs them.

What I would do differently

Agree with the first areas that reused the model, at the start, how they would describe its benefit in their own numbers. Cost was easy to make visible. Value needed the people who received it to name it, and that conversation is easier early than late.

Confidentiality

A generalised account. It contains no internal figures, documents, system names or people, and nothing beyond what is already public about the organisation.