Platform
Find the AI your organization doesn't know it has
Surface unmanaged AI, shadow AI, unknown agents, unapproved models, external AI APIs and unregistered tools — then bring them under governance.
Found, not declared
Corroborated
5
Can change state
4
Against 11 declared systems. The comparison is the point — a discovery count on its own says nothing about coverage.
Why an inventory is not enough
The systems that matter most are the ones nobody registered.
A copilot that arrived with a seat licence. A scheduled job that quietly gained a model call. A SaaS feature that turned up in a release note. None of them passed through a build review, so none of them are on any list — and every one can change something.
This estate declares 11 AI systems and discovery has surfaced 5 more that were corroborated by at least two independent signals. 4 of the findings can change something in a system of record, which is the number that should set the order of work.
The honest framing is that this is not a scanning product finding secrets. It is a join across systems you already run, made useful by asking one question of every result: can this thing act?
All 5 sources are systems you already run.
Source
What it can see
What it is blind to
- Identity provider
- OAuth grants to third-party AI applications, and which employee authorised each one.
- Anything reached with a personal account or an API key rather than a corporate sign-in.
- Egress proxy
- Outbound requests to known model provider endpoints, by source workload and volume.
- Traffic that leaves through an unmanaged network, and the contents of any request.
- SaaS admin consoles
- AI features enabled per tenant, and seat licences assigned to them.
- Features enabled by default in a product update that nobody was notified about.
- Code host search
- Model SDK imports and provider base URLs across every repository.
- Whether the code is deployed, and whether it runs against production.
- Card and procurement data
- Recurring charges to AI vendors, including ones below any approval threshold.
- Free tiers, trials, and anything expensed as something else.
One signal is a lead. Two is a finding.
An OAuth grant proves somebody connected an app. Egress proves something called a model. A repository import proves code exists. None of those alone proves an ungoverned system is taking actions — which is why the headline number here is corroborated findings rather than raw detections.
What was found
Sorted by whether it can change something, not by how many seats it has.
Most shadow-IT reporting counts licences, which produces a long list ordered by spend. The column that should drive the work is whether the thing can reach a system of record.
Discovery findings
7 findings · 2 unclaimed| What was found | Kind | Appears to belong to | Seen by | Can change state | Status | First seen |
|---|---|---|---|---|---|---|
| Vendor copilot enabled in the CRM tenant | embedded AI | RevOps | SaaS admin consolesIdentity provider | yes | triaged | 2 Sep 2026 |
| Meeting notes app with an AI summariser | copilot | Sales | Identity providerCard and procurement data | — | triaged | 18 Aug 2026 |
| Scheduled script calling a model API directly | AI API | unattributed | Egress proxyCode host search | yes | unclaimed | 11 Sep 2026 |
| Helpdesk macro suggestions | embedded AI | Support Ops | SaaS admin consoles | — | onboarded | 4 Jul 2026 |
| Document OCR with a model behind it | application | Finance | Card and procurement dataEgress proxy | yes | onboarded | 22 Jun 2026 |
| Consumer chat assistant used with work documents | model | unattributed | Egress proxy | — | removed | 9 Aug 2026 |
| Reporting automation with a model call added | automation | Ops | Code host searchEgress proxy | yes | unclaimed | 13 Sep 2026 |
Findings
7
across 5 sources
Can change state
4
triage these first
Unattributed
2
nobody to accept them
Brought under governance
2
registered and owned
IllustrativeThe OpsAI sample estate, alongside 11 declared systems. The declared inventory.
What happens next
Two outcomes, and a catalogue entry is not one of them.
A finding is either brought under governance or shut off. Leaving it discovered is the failure mode that makes discovery tooling feel like paperwork — a longer list every quarter and nothing changed.
01
Can it change anything?
The first question, and the one that sets the order. A summariser that reads is a different problem from a copilot that edits records, and counting seats tells you neither.
02
Whose is it?
Attribution is usually the slow part. An unattributed finding cannot be onboarded, because there is nobody to accept it — which is why the unattributed count matters more than the total.
03
Govern it or remove it
Registered with an owner and a rule at the action, or switched off at the source. Anything reached through a browser rather than an API has no boundary to sit in front of, so removal is often the only honest option.
Some things cannot be governed, and blocking them is the honest answer.
A consumer assistant reached through a browser has no API boundary to sit in front of, so it cannot be governed the way a connected agent can. Pretending otherwise would be the dishonest option. It is handled at the data boundary instead — which is a real control, just a different one.
Where to start
Ask your identity provider which AI apps have an OAuth grant.
It is a five-minute query against a system you already run, and it is usually the first time anyone sees the real number. Nothing about it needs OpsAI — which is rather the point about how available this information already is.