The question regulated buyers actually ask
When a financial-services team evaluates an AI agent, the real question is never “does it work?“. It is “can we deploy this without failing our resilience obligations?“. The honest answer is yes, provided the controls were designed in from the first commit rather than bolted on before an audit. DORA is useful here precisely because it is specific: it turns “governance” from a slide into an engineering checklist.
What DORA asks, translated to engineering
The EU’s Digital Operational Resilience Act has been in application since January 2025. It covers financial entities across banking, insurance and investment services, and it reaches the ICT third parties that serve them. If you operate in the EU, or serve EU customers from the UK, its requirements are on your desk, and the UK’s own operational-resilience regime from the PRA and FCA asks a recognisably similar question. Four of its asks land directly on any agent that touches a regulated process:
- Registered, assessed ICT risk. The agent is an ICT service. It belongs in your ICT risk management framework, assessed and documented, not running in a gap between departments.
- Third-party terms that survive scrutiny. Audit rights, subprocessing disclosure, where data lives, termination and exit assistance: the contract has to say them, and your third-party register has to show them.
- Resilience you can demonstrate, not assert. Tested changes, restore rehearsals, and a rollback that is a defined, rehearsed operation. An assertion of resilience is worth nothing; a rehearsal log is worth everything.
- Incident readiness on a clock counted in hours. Detection, classification, and a reporting path your team has actually walked, before the day it matters.
Underneath all four sits one property: an audit trail. Every agent action attributable, logged, and reconstructable months later.
How these controls ship in practice
This is how Nugawi, an applied-AI consultancy in London, builds production deployments for regulated sectors:
- Attribution by construction. An access matrix, and a record of who requested each action, with release history and approvals. Any action can be reconstructed: who, what, when, approved by whom.
- Tested releases, rehearsed rollback. The same tested package is promoted to live, never a rebuild, and rollback is rehearsed like a fire drill, because the one you have never run is the one that fails.
- Data where you chose it. UK or EU residency, clear controller and processor terms, secrets out of delivery packages.
- Incident support in scope. Monitoring and a defined incident path inside the engagement, not a best-efforts shrug.
- Exit designed in. Code, data and attachments return in usable formats with handover help. A vendor whose retention strategy is switching costs will fail your third-party review, and deservedly so.
These controls are standard in our 90-day production build, and an agent audit maps your existing agents against them in weeks, not quarters. If financial-services deployment specifically is the question, our Claude consulting practice covers the data-residency and governance detail.
Five questions before an agent touches a regulated process
Lab-owned platform or independent partner, ask these before anything goes live:
- Where does every prompt, document and output physically live, and under whose terms?
- Can you reconstruct any action six months later: who, what, when, approved by whom?
- What exactly is the rollback, and when was it last rehearsed?
- If you terminate the vendor tomorrow, what happens to the system?
- Which register does this land in, and who signs it off?
The pattern generalises
The same controls answer customs audits in logistics and tenancy-data duties in housing, under different articles: access matrices, audit trails, residency, rehearsed recovery. We ship them into logistics and real-estate deployments the same way.
Resilience is not a document you file. It is a property of a well-built agent. If your agent cannot pass the five questions above, no policy paper will save it, and if it can, the policy paper writes itself. A Discovery Sprint maps your regulated workflows against those questions in a few weeks, with a written recommendation and a firm price.