Written by: Agata Wojtas, Chief Commercial Officer, Digital Colliers
The DeepSeek round is a useful stress test for your AI governance. A single Chinese lab reportedly pricing at $74B on $400-500M of annualised revenue tells you two things at once. The model layer is still repricing violently, and the vendors your risk committee approved last quarter may not be the vendors your engineers want to use next quarter. If your bank has one model provider baked into production, that is now a concentration risk you have to name in writing.
DORA already made this your problem
DORA has been in force since 17 January 2025, and its third-party ICT register is not a light-touch ask. Every critical provider needs to be listed, assessed, and shown to be substitutable. Regulators are explicit that concentration on a single provider is itself a finding, not just an operational nuisance.
The practical read for financial services:
- Your foundation model vendor is a third-party ICT provider. Full stop.
- If that vendor is domiciled outside the EU, or subject to export controls, or funded by capital your compliance team cannot trace, that goes in the register too.
- Substitutability is not a slide. You have to be able to show, on demand, that you could move workload off a provider inside a defined window.
The DeepSeek story is not really about DeepSeek. It is about the fact that any lab you depend on today can be re-rated, restricted, acquired, or geopolitically fenced off inside a quarter. Your register has to assume that motion.
What model abstraction actually looks like
Most banks I see talking about "multi-model strategy" have one provider in production and a second one in a slide deck. That is not abstraction. Abstraction means the application code, the prompts, the evaluations and the logs do not know or care which provider served a given call.
The pattern operators are converging on looks roughly like this:
- A thin internal gateway between your applications and any external model API. All traffic goes through it. No direct SDK calls from application code.
- Prompts stored as versioned artefacts, not string literals in code. Each prompt has an ID, a version, an owner, and a target task, and it is portable across providers.
- An evaluation harness that runs the same test set against every candidate model on a schedule, so you have current evidence of which providers can serve which tasks at what quality.
- Structured logs at the gateway that capture prompt ID, model ID, model version, input hash, output hash, latency, cost, and the compliance flags relevant to the call.
Done properly, swapping a provider becomes a config change plus a re-run of the evaluation harness, not a project.
The regulatory clock is not just DORA
DORA is the near obligation, but the EU AI Act stack lands on top. Article 50 transparency obligations apply from 2 August 2026, and high-risk system obligations apply from 2 December 2027. Fines for high-risk violations reach up to €15M or 3% of global turnover, and that sits alongside GDPR exposure of up to €20M or 4% for the personal data that flows through the same pipelines.
If you cannot produce, per call, which model version generated which output on which customer's data, you are going to struggle with all three regimes at once. The logging discipline is not optional infrastructure. It is what makes the register truthful and the audit survivable.
Why most programmes will not get there in time
Around 95% of enterprise AI projects fail to reach production or ROI, and the ones that do reach production tend to be the ones that treated abstraction, evaluation and logging as day-one work rather than a later hardening phase. The banks that will be exposed in 2026 and 2027 are the ones running pilots wired directly to a single vendor SDK, with prompts pasted into notebooks and no independent evaluation.
A few practical questions to put to your own stack this quarter:
- If our primary model vendor became unavailable on Monday, what is the actual restore-to-service time, in hours?
- Can we name the model version that produced any given customer-facing output from the last 90 days?
- Do our prompts live in a versioned store, or in application code?
- Is our evaluation set run against at least two providers on a recurring cadence?
If the honest answer to any of those is no, the DeepSeek headline is not somebody else's story. It is a preview of the next audit question you will be asked.