Sector

Models that have to satisfy a regulator, not just a test set.

In lending, payments and financial crime, a model that performs well but cannot be explained is a model that does not ship. The explanation is part of the product.

01 / What is hard here

The problem underneath the problem.

Financial services has the hardest combination of constraints in applied AI: adversaries actively probing the system, regulators requiring documented reasoning for adverse decisions, and error costs that are large and wildly asymmetric between the two directions.

The technical consequence is that model choice is constrained from the start. The most accurate available architecture is frequently unusable, because it cannot produce reasons that survive an adverse-action notice or an internal credit committee.

02 / What we build

Work we take on in finance.

Credit decisioning support: Scoring with reason codes, tested for stability across segments and time, documented for model risk review. Built to assist a credit policy rather than to quietly replace it.
Transaction fraud detection: Velocity, network and behavioural features scored in real time, with rules deployable in hours for new attack patterns while retraining proceeds separately.
AML alert triage: Ranking of alerts by likelihood of genuine suspicion so that limited investigator capacity goes to the cases most likely to matter. Alert suppression is not something we build.
Reconciliation and exception handling: Automated matching across ledgers and statements with confident exceptions routed to people, which is frequently the fastest genuine return in a finance back office.
Document processing: Extraction from statements, KYC documents and agreements with calibrated confidence and mandatory human review below threshold.

03 / Constraints

What shapes the build.

These are design constraints agreed at the start, not compliance work bolted on before launch. Retrofitting them is expensive and usually incomplete.

Explainability: Adverse decisions need reasons that a customer, a reviewer and a regulator can each follow. This constrains model families before accuracy is even discussed.
Evidence retention: Every decision reconstructable years later: model version, feature values, threshold and outcome, retained per your obligations.
Fair-lending review: Decision and error rates examined across protected and proxy characteristics before launch, and monitored continuously afterwards.
Data residency: Where customer financial data may be processed, and by whom, decided at design time — including whether it may reach a third-party model provider at all.

04 / What we advise against

Where we would say no.

We would advise against fully automated adverse credit decisions without a human review path, and against any fraud model whose threshold has been set by a data team rather than jointly with the people who absorb the cost of both error types.

We would rather lose the work than build something that creates a liability for you or a harm for the people on the other side of it. If that rules out what you had in mind, it is better established now than at delivery.

05 / Other sectors

Adjacent work.

Different domains, the same delivery discipline — a measured baseline, a real evaluation set, and monitoring that catches decay early.

Next step

Working in finance?

Tell us which process is slow, expensive or error-prone. The diagnostic will tell you whether this is a problem worth solving with machine learning — including if it is not.

Request a diagnostic All industries