DORA and the case for fewer AI vendors
Every AI tool you add is another third party to oversee. DORA changed the maths on that, and self-hosting is one possible answer.
Every external AI tool a firm adopts is, in regulatory terms, another third party. Someone has to assess it, contract with it, monitor it, and plan for the day it fails or is breached. For a long time that overhead was easy to ignore. DORA made it explicit.
Since 17 January 2025, the Digital Operational Resilience Act has required financial entities across the EU to manage information and communication technology (ICT) third-party risk as a named risk: a register of information on contractual arrangements, due diligence before contracting, monitoring after, and exit strategies for services supporting critical or important functions. The Act does not forbid using external services. It requires their number, concentration and criticality to be assessed. Every additional AI vendor in the data path is one more entry in that register and one more dependency to defend.
This changes the maths on how a firm should assemble its AI stack. The instinct in the market is to bolt together best-of-breed services: one vendor for the model, one for retrieval, one for guardrails, one for monitoring. Each is a sensible choice on its own. Together they are four DORA ICT third parties, four contracts, four sets of due diligence, four points that can independently be compelled to disclose data or be breached.
Self-hosting changes that risk; it does not erase third parties. A customer-controlled application can remove a vendor-operated application cloud and reduce standing external access, but the software supplier, implementation partner, support arrangement and any external model may still be ICT dependencies that need assessment. The right DORA classification depends on the actual contract and function, not the word “self-hosted.”
This is not an argument for self-hosting everything. It is an argument for counting. Before adding the next AI vendor, a regulated firm should ask what it adds to the third-party register, and whether the same job could be done by something that runs where the data already lives. Often the objective answer is that fewer vendors, self-hosted, is the more defensible design.
GAIA Brain is being developed around that narrower premise: keep the application and evidence under customer control, make every external endpoint explicit, and reduce unnecessary point-to-point dependencies without pretending that customer-hosted software sits outside DORA.