Banco BMG: +30% credit conversion

The challenge

In lending, conversion is won or lost at the credit decision. Friction and hesitation in the approval and offer flow leak qualified applicants the bank already worked to reach. The opportunity was the decision itself, not more volume at the top.

What Studio Labs built

We embedded decision architecture at the exact point where the credit decision happens, shipped as a production agent. It runs under auth, RLS, audit logs, and continuous evaluation, aligned with LGPD from the start, inside the same stack the bank already operates.

Not a strategy deck. Decision logic running in production, owned by the team it was built with.

The result

Credit conversion up 30%, with the decision logic live in production at the point where applicants commit.

How it runs in production

The part that decides whether this holds is not the model, it is everything around it. The work runs inside the stack Banco BMG already operates, not beside it as a separate system, because a decision that lives in a parallel tool is a decision nobody acts on.

That means auth and row level security so the system only ever sees what the requesting user is allowed to see, rate limits and audit trails so every action is attributable, and continuous evaluation against real traffic rather than a fixed test set. Observability is wired in from the first day, because a system you cannot inspect on a specific request is a system you cannot debug at two in the morning.

What the team owns

The code belongs to the client from day one, in their repositories. We are not a dependency you rent indefinitely, and the engagement is not designed to make you need us next quarter.

We embed with the internal team while the work is built, so the people who will run it afterwards are the same people who watched it get made. We stay until it holds in production and the team can move it forward without us, which is a harder standard than shipping and walking away.

Where this pattern applies

The shape generalizes beyond lending and banking. Wherever a decision is made repeatedly, at volume, with a real consequence you can read in a number, the same approach applies: find the decision worth automating, build the system that runs it, wire it into the flow where applicants actually are, and hold it to a measurable standard.

What does not generalize is skipping the diagnosis. The credit decision was the right place to work here because that is where the decision lived. Somewhere else, it will be a different step entirely, and starting from the technology instead of the decision is the most common way this work fails. That is the reasoning behind how we build production AI, end to end, and why we wrote about how to prioritize AI use cases.