What is self-hosted AI governance?

Self-hosted AI governance is software for recording AI systems, applying policy to AI requests and keeping evidence, which the organisation installs and operates on its own infrastructure. Prompt content is inspected on the customer’s network, records are stored there, and the vendor is not in the data path. It is the opposite of a cloud service where requests pass through the vendor.

The term covers two different layers. The governance layer holds the inventory, reviews, approvals and evidence. The control layer, often an AI gateway, sits between users or applications and the model providers and decides whether each request may proceed. A product can be self-hosted for one layer and cloud-delivered for the other, so it is worth asking about each.

Why regulated organisations ask for it

The reasons are practical and usually come from the security, privacy or risk function rather than from a preference for running software.

  • One fewer party in the AI data path. An AI gateway sees the full prompt and the full response. If it is a cloud service, its vendor receives that content. Where that content contains personal data, the vendor is generally a processor under the GDPR, which brings a contract, a record and a due diligence task. Self-hosting keeps that content on your network.
  • Third-party risk in financial services. The Digital Operational Resilience Act (DORA) has applied to EU financial entities since 17 January 2025. It includes requirements for managing ICT third-party risk and for keeping a register of information on ICT service arrangements. A governance vendor operating in your AI data path is one more arrangement to assess and record.
  • Sovereignty and residency. Some organisations must show where data is processed and who can be compelled to disclose it. Software you operate in your own data centre, or in your own cloud account in a chosen region, is easier to describe than a shared service.
  • Isolated networks. Defence, critical infrastructure and some public bodies run segments with no route to the internet. A product that phones home for a licence or an update cannot run there.

None of this is a legal conclusion. Whether a vendor is a processor, whether DORA applies to an arrangement, and what a residency requirement demands are determinations for your organisation and its counsel. Self-hosting changes the facts those determinations are based on. It does not make them for you.

Cloud-delivered versus customer-operated

QuestionCloud-deliveredCustomer-operated
Where prompt content is inspectedIn the vendor’s environmentOn your network, before it leaves
Where records and evidence liveIn the vendor’s storage, in a region they offerOn storage you choose and back up
New processor in the AI data pathUsually yes, for any personal data in promptsNo, the vendor does not receive the content
Offline operationNot possiblePossible if licences and updates work offline
TenancyOften shared infrastructure, separated logicallySingle-tenant by construction
Who patches itThe vendor, on their scheduleYour team, through your change process

The trade-offs

Self-hosting moves work and responsibility to you. Decide whether that is acceptable before comparing features.

  • You operate and update it. Someone owns the host, the backups, the certificates and the upgrade calendar. A security fix reaches you only when your team applies it.
  • You need a host. Most self-hosted products ship as containers and need a Docker host or similar, sized for the traffic you route, with storage you protect.
  • Controls apply only to routed traffic. A gateway enforces policy on the requests that pass through it. Traffic that goes straight to a provider is not controlled, which is why discovery from logs and endpoints still matters.
  • No vendor view of your incidents. If the vendor cannot see your environment, support depends on what your team can collect and share. That is the point, but it changes how troubleshooting works.

A self-hosted product that still sends usage data, licence checks or error reports to the vendor by default is not giving you the property you are paying for. Test the claim on a network with no outbound route.

Questions to ask any vendor claiming on-premises support

  • Is any telemetry, usage data or error reporting sent to you by default, and can it be switched off entirely?
  • Does the licence need an online check, at installation or periodically, or can it be verified offline?
  • Can updates be downloaded once, carried in and applied with no internet access, and are they signed?
  • Where is evidence stored, in what form, and can we export it without your involvement?
  • Does inspection of prompt content happen on our network, or is any of it sent to your service for classification?
  • Is the control layer self-hosted as well as the governance layer?
  • Does any feature stop working if your company or your cloud is unavailable?
  • Is each deployment single-tenant, and who holds the administrator credentials?

A practical path to evaluation

  1. Write down the constraint. Name the requirement driving self-hosting: a processor concern, a DORA register entry, a residency rule or an isolated network. It decides which answers matter.
  2. Send the questions above. Ask for written answers, not a call.
  3. Install on a closed network. Run the trial on a host with outbound traffic blocked and watch what fails.
  4. Route one real workload. Point one application or team at the gateway and check what is recorded and what is refused.
  5. Apply an update offline. Carry an update in and apply it, so you know the operating cost before you sign.

How MERIDVAR does it

MERIDVAR is software you install and operate: a self-hosted, single-tenant appliance on a Docker host on your private network.

  • No vendor cloud. There is no MERIDVAR cloud and no telemetry. Nothing is sent to MERIDVAR.
  • Offline operation. Once installed it runs without internet access, including signed licences and signed updates.
  • Inspection on your network. The admission gateway applies prompt DLP, data residency, budgets and approved model lists before a request reaches the model provider, and a refused request never reaches it.
  • Evidence you hold. Decisions are written to an append-only, hash-chained evidence log on your host, with optional Ed25519 signatures.
  • Live dispatch. Admitted requests are dispatched to Anthropic, OpenAI, Azure OpenAI and xAI.

Controls apply only to traffic routed through the gateway. MERIDVAR is in beta. See the deployment model for how it is installed.

Sources: European Commission, what is a data controller or a data processor; Regulation (EU) 2022/2554 (DORA); EIOPA, Digital Operational Resilience Act.