Integrations
Identity provider integrations
Use Okta, Microsoft Entra or Google Workspace as the source of truth for who a person is, and tie every AI action back to one of them.
OpsAI is not an IdP
It reads your directory and never maintains a copy. A governance product with its own user list is wrong about leavers within a week, and “who approved this” gets answered from stale data.
People
6
User lists OpsAI keeps
0
Where availability is real
Every directory exposes its groups and its grants differently.
Which makes this per-provider work. Sign-in is the easy half and it is not the interesting one — what varies between providers is how you enumerate group membership and how you list the third-party app grants discovery depends on.
- available
- Governed today, with the action vocabulary of that system.
- in progress
- Being built. Partially usable, and the gap is named on the entry.
- not yet
- Not started. Reachable only through a generic connection, if at all.
- Okta
Okta
availablePeople, groups and the OAuth grants discovery reads to find AI nobody registered.
- Microsoft Entra
Microsoft Entra
availableSame reads. Group membership is how an approver is resolved rather than a list maintained in OpsAI.
- Google Workspace
Google Workspace
availablePeople and third-party app grants, which is where a shadow copilot usually shows up first.
- SAML or OIDC
SAML or OIDC
availableFor a directory not named above. Console sign-in works; automated discovery of grants does not.
IllustrativeAn illustrative catalogue for the OpsAI sample estate. §54: two-letter marks rather than logos.
What OpsAI reads
Two reads from the same integration, and the second is the surprising one.
People and groups is what you would expect: it resolves an approver and an accountable human. The other read is OAuth grants, and it is how discovery finds AI that nobody registered.
- people and groupsWho a person is
- An approver is resolved from group membership rather than from a list maintained in OpsAI. The sample estate has 6 people across 6 teams, and every accountable assignment points at one of them.
- supportfinanceprocurerevopsclaimslegal
- oauth grantsWhat somebody connected
- OAuth grants to third-party AI applications, and which employee authorised each one.
- Blind to: Anything reached with a personal account or an API key rather than a corporate sign-in.
From a person to an action
Every action carries the chain back to a human.
This is what a directory integration is ultimately for. An action attributable only to a service account answers none of the questions asked after an incident, and adding the name afterwards is not the same as having recorded it.
A person in your directoryyour source of truth
Read from the IdP, never maintained in OpsAI. If they leave, they leave — there is no second user list to fall out of date.
Named accountable for an AI systemrequired
One human per system, required at registration. Resolved from the directory, so a leaver is visible as a gap rather than as a name that still looks fine.
The AI system attests its own identityattested
Not a claim in a header. mTLS, a workload token or a platform-signed JWT, depending on how the system runs.
The action carries the whole chainon every action
From the accountable human down to the acting system, on every attempt. This is what makes an action attributable to a person rather than to a service account.
How a workload proves itself
- orchestrator
- mTLS client cert
- retrieval
- mTLS client cert
- conversational
- OIDC workload token
- scheduled worker
- Platform-signed JWT
An asserted identity makes every downstream control decorative. If anything can claim to be the refund resolver, its bounds constrain nothing.
In the canonical trace · 1 ms
Every hop holds a subset of the hop before it. A sub-agent cannot be talked into authority it was never issued.
People
6
read from the directory
Teams
6
owning policies and systems
Actions without a named human
0
the chain is required
User lists OpsAI maintains
0
your IdP is the truth
IllustrativeIllustrative figures for the OpsAI sample estate. How accountability is modelled.
What happens when somebody leaves
They disappear from the directory, and every AI system they were accountable for is immediately visible as unattributed rather than continuing to show a name that looks fine. That is the practical reason not to keep a second user list: a copy does not go stale loudly, it goes stale quietly, and the failure surfaces at the worst possible moment.
The accountable human for the canonical action is Priya Nair of Support Ops, resolved from the directory at the time the action was attempted and recorded in the evidence entry.
Where to start
Connect the directory first. It does two jobs before you have written a policy.
It gives every action a person to attribute to, and it starts surfacing the AI somebody signed up for without telling anyone. Neither needs a rule in place to be useful.