The Sovereign AI Balance Report: Banking Edition

The products are commoditised, and after the Reset the intelligence is too: every bank can put a frontier model on every desk. What is not commoditised is the customer relationship, who a bank acquires, how precisely it personalises, and why an account stays instead of switching. That relationship runs on transaction histories, behavioural signals, and life events, the richest first-party dataset in the economy, and the most closely regulated.

This is not an argument against frontier models. Claude and Codex are extraordinary at open-ended reasoning, synthesis, and code. It is an argument about placement: the models that read your customers' lives in order to grow your book belong on infrastructure you control, and the frontier's capability premium applies to the edges of the bank, not to the relationship at its centre.

A model that personalises credit offers engages GDPR Article 22's limits on solely-automated decisions, the EU AI Act's high-risk classification for creditworthiness assessment, the consumer-credit rules, and supervisory model-risk expectations, all at once. And the same customer data that powers next-best-action also powers the AML and fraud controls, so the perimeter around it is watched by seven overlapping regimes. Every one of those obligations is easier to meet on infrastructure you control than through a third-country API.

DORA compounds the point: a frontier lab serving a critical or important function becomes a regulated ICT third party, with register, exit-plan, and concentration-risk obligations. The throughline across GDPR, the AI Act, DORA, EBA outsourcing guidance, PSD2/3, MiFID II and the AML package is the same, the European stack systematically rewards control over where customer data sits, who can touch it, and whether you can prove what the model did with it.

At the sovereign end sit the workloads where the customer and the regulation meet: retention and churn models over account history, personalisation and next-best-action, cross-sell across products, credit decisioning, and the AML and KYC controls that guard the same data. These run on self-hosted, open-weight models inside the perimeter, governed and auditable end to end, where an Article 22 explanation and the AI Act's documentation are straightforward to produce.

At the frontier end sit low-sensitivity, high-capability tasks: campaign content and copy, research synthesis, public-data analysis, and coding on non-core repositories. The instructive case is software engineering, frontier coding agents earn their keep on most repos, but core-banking logic and secrets belong on self-hosted code models. Even the developer tooling needs the split.

The figures here are modelled benchmarks for a representative universal bank, not survey statistics. A digital-only retail bank tilts further sovereign, its whole business is the customer data; a markets-heavy investment bank shifts slightly toward frontier. The durable claim is the two-thirds sovereign centre of gravity, and that each customer workload, in a specific bank, has a calculable right answer.

Inventory every AI workload that touches customer data, live and planned, most banks undercount by half, then classify each against GDPR Article 22, the AI Act risk tiers, and your own sensitivity scale rather than by department.

Move the growth stack first. Retention, personalisation, and next-best-action are high-volume and high-sensitivity at once, which is exactly where self-hosting wins on cost and control together. A sovereign retention pilot is the cleanest proof, because a few points of churn show up on the P&L within a quarter.

Then make placement a standing governance decision, not an accident: every new customer model gets scored and placed by default. The portfolio split above is the destination; the routing discipline is how you get there without slowing the bank down.