AI Registry
Last Updated: September 10, 2026 | By Mihail Sebastian | AI Dictionary
An organization's inventory of approved AI assets: models, applications, agents, and MCP servers, with ownership, approval status, and review dates.
What is an AI Registry?
An AI registry is an organization’s inventory of approved AI assets: the models, applications, and agents in use, each recorded with an owner, an approval status, and a risk classification.
It is the counterpart to shadow AI: the registry defines what “approved” means, and anything outside it is by definition unaccounted for. Increasingly the inventory covers not just models but the tools agents reach, including MCP servers.
How an AI Registry Works
A registry runs on a lifecycle. At intake, whoever wants a new model, application, or tool submits it with the intended use case. At review, a governance function assesses the risk, checks the vendor, and classifies the asset.
Approval attaches conditions: approved for internal data only, approved with human review of outputs, or rejected. Each entry names an owner who answers for the asset, and periodic re-review catches entries whose risk changed since approval.
The registry earns its keep only if the path through it is fast. When approval takes months, teams route around it, and the inventory stops describing reality.
Regulation pushes in the same direction. The EU AI Act obliges organizations to know which AI systems they operate and to classify them by risk, and a maintained registry is where that evidence naturally lives. It also hands an AI audit its starting list.
AI Registry vs AI Gateway
The registry is the record of what is approved; the AI gateway is the runtime checkpoint that enforces it. One is a policy artifact consulted by people, the other a piece of infrastructure sitting in the traffic path.
They work as a pair: the registry decides, the gateway executes the decision on every request.
| Criterion | AI Registry | AI Gateway |
|---|---|---|
| Role | Record of what is approved and who owns it | Enforcement of policy on live traffic |
| When it acts | At intake and review, before use | At request time, during use |
| What it holds | Owners, approval status, risk class | Logs, quotas, credentials, guardrail rules |
| Without it | Nobody knows what AI is in use | Rules exist on paper, nothing applies them |
Example of an AI Registry
A data science team wants an external LLM API for contract summarization. They file an intake request naming the provider, the data involved, and the use case.
The review board classifies the use as limited-risk but flags that contracts contain customer data, so approval comes with a condition: access only through the company gateway, with logging on. The entry lists the team lead as owner and sets re-review in twelve months.
Six months later a colleague searches the registry, finds the approved API, and reuses the entry instead of putting a new vendor through review from scratch.
Related AI terms: AI Governance · Shadow AI · MCP · AI Audit · Compliance
Did you like the AI Registry gist?
Learn about 250+ need-to-know artificial intelligence terms in the AI Dictionary.
