Solutions
AI app governance - an IT approval workflow for AI-built apps
AI app governance is the set of rules and checks a company applies to apps built with AI: who may build, what data each app may reach, who approves a release, where it runs and what it may cost. Diakrin enforces these rules on the server for every app: data access needs IT-approved grants, every production release needs an approved plan, from a reviewer who is not the requester where your policy requires it, budgets are checked before deployment, and each step leaves evidence for audit.
Last reviewed
AI app governance is the set of rules and checks a company applies to apps built with AI: who may build, what data each app may reach, who approves a release, where it runs and what it may cost. The hard part is not writing the policy. It is enforcing it on every app, every release and every coding agent, without slowing builders to a stop. Diakrin enforces these rules on the server and records the evidence.
What is AI app governance?
General AI governance frameworks, such as the NIST AI Risk Management Framework ↗, cover how a company governs, maps, measures and manages AI risk. AI app governance applies that to a narrower and fast-growing case: business apps that employees build with AI coding agents. Those apps are ordinary software once built, but they are created by people outside engineering, often quickly, and often by an agent acting on their behalf. That changes who IT has to trust and where the checks need to sit.
What should IT require before an AI-built app reaches production?
Diakrin's deployment review checklist asks IT to confirm each of these for the exact app and environment:
| Check | What IT confirms |
|---|---|
| Environment | The exact test, staging or production target |
| Audience | Who should open the deployed app |
| Data | Approved sources, operations and record-level access |
| Runtime | A connected, compatible company-cloud runtime |
| Cost | Applicable limits, estimates and retained costs |
| Approval | The person authorized to accept the plan |
| Recovery | How a failed release is diagnosed or rolled back |
A deployment is ready for team use only after its real address works for an authorized user, unauthorized access is denied, the intended workflow succeeds and its owner confirms the result. See Review capabilities and deployment.
How does an IT approval workflow for AI-built apps work?
Approval happens at three points, and each one is separate:
- Who may build. IT decides who can create apps and whether creating one needs approval. Builders work only within assigned apps.
- What each app may reach. The app requests a capability, IT approves a named operation with limits, an authorized owner configures the connection, and your cloud verifies it. Only the last state means the app can use it.
- What goes live. Diakrin prepares a release plan with the exact image, target environment, cost estimate, access implications and expiry. Under your approval policy, an eligible reviewer who is not the requester approves or denies it. Promotion applies that approval to that release only.
How do you govern the coding agents themselves?
An agent connected to Diakrin acts as the person who connected it. Every call is checked on the server against that person's current role and app access, so it can never do more than they can. It cannot approve its own requests or reach another organization. This helps limit what OWASP calls excessive agency in its top risks for LLM applications ↗: the connection never grants more than the person's own access. Removing a person or changing their role invalidates their coding-tool connection. See Connect your coding tool.
What evidence does a CISO get?
- Separate states. Saved, requested, approved, built, previewed, deployed and verified live are recorded separately, and none stands in for the next.
- Operation history. Builds and deployments keep their stages, receipts and approvals, so you can see what happened and who approved it.
- Data inventory. On AWS with BYOC (not yet on Google Cloud or self-hosted), the runner in your cloud reports what each app stores, by size and name, never by value, with a risk score that explains itself.
- Cost by app. Estimates name their price source, and anything Diakrin could not price is listed as unknown, never as zero.
- Data controls. IT can set retention, export, quarantine and delete an app's data; your cloud applies each control and returns a signed receipt.
See Manage access and security.
How it works with Diakrin
- Set roles and the creation policy. Choose who are Builders, who are IT administrators and who are Technical administrators. See Invite your team and set roles.
- Connect identity. Connect company SSO and, optionally, SCIM so offboarding removes access. See Connect your company SSO.
- Set budgets and approval policy. Every app has a monthly ceiling; production releases follow your approval policy.
- Review access requests. Approve or deny each named data operation, with separate app and employee grants.
- Review release plans. Approve or deny the exact plan before promotion. See Preview, approve and deploy.
- Operate. Track health, cost, data and activity; change audiences, quarantine data or stop a runtime when needed.
If unsanctioned apps already exist, start with shadow AI apps.
Frequently asked questions
What should IT approve before an AI-built app reaches production?
The exact environment, who can open the app, which data sources and operations it may use, that it runs on a connected company-cloud runtime, its cost limits and estimate, who is authorized to accept the plan, and how a failed release is diagnosed or rolled back.
Can a builder approve their own release?
No. Under your approval policy, an eligible reviewer who is not the requester approves or denies the exact plan. On Google Cloud, production always needs a current independent approval.
What does IT approve when an app asks for data?
A named operation with resource and field limits. The app and each employee need their own grants. A requested capability is not an approved one, and an approved one still needs a configured, verified connection.
Can IT limit what an app costs?
Yes. Every app has a monthly budget ceiling, and deployments that would exceed it are refused at admission. The cost shown is an estimate priced from public on-demand prices, not your cloud bill.
How does a CISO see what AI-built apps exist and what data they hold?
IT sees every app's cost, data and activity in one place. On AWS with BYOC (not yet on Google Cloud or self-hosted), the app data inventory reports database and table sizes, estimated row counts, column names, file counts, log retention and secret rotation age, never row values or file contents.
What happens when someone leaves the company?
Removing them from the Diakrin organization revokes their app and project grants, ends their recorded app sign-in sessions and stops their coding-tool tokens. With SCIM, deactivating them in the identity provider removes their membership on the next sync.
Does approval authorize every later change?
No. Each production release needs its own approved plan, and preview approval never authorizes production. Rollback reuses an earlier plan only while its approval is still valid.
Sources
Facts about other products come from their public pages, checked on the dates shown. Products change; tell us if something is out of date.
- nist.gov/itl/ai-risk-management-framework (checked 9 November 2026)
- genai.owasp.org/llm-top-10 (checked 9 November 2026)
See Diakrin with your own cloud
Let your team build internal apps with the AI tools they already use, while IT keeps control of identity, data, approvals and spend.