Financial servicesFirst-hand

Fraud model: build or buy?

Data gravity and real-time latency pin it to self-host.

Open the live lab · pre-loaded to this scenario

Build vs Buy vs Fine-Tune

Context

A card issuer decides how to run its fraud model. A vendor is fast to stand up, but the training data can't leave the bank, scoring must happen in real time, and the model itself is a competitive edge.

The decision

Self-host / fine-tune wins here — data gravity, sub-100ms latency, and differentiation all point the same way, and the bank has the team to run it.

What most miss

The vendor's time-to-value is seductive, but on fraud the data can't move and the latency is non-negotiable — control isn't a preference, it's a constraint.

Stakes

A vendor round-trip that adds latency or moves fraud data off-prem isn't a cost line — it's a regulatory and loss-rate problem.

Takeaway · When data can't move and latency is a constraint, control stops being optional.

First-hand · Business of AI · verified 2026-07-03

Sources: Fraud-model build/buy — first-hand (cards & payments, American Express); Real-time scoring + data-gravity constraints

← All industries·See it in a full program storyline →