AI Agents and Reviews
AI agents and apps can hold access that nobody on your team has looked at. The two AI Access pages list what the collected data shows, ask a named person to take responsibility for each agent, and keep a record of what was decided and whether a later collection shows the change.
This page is for IT managers, security leads and compliance managers. It covers what the pages show, how to name an owner, how to run a review campaign and what the evidence pack does and does not prove. It does not cover how the data is collected; during the pilot AutoCISO does that with you.
Before you start
- AI Access turned on for your organisation (see the note above). Both pages are under Access & Risk in the sidebar.
- A completed collection of data. An AutoCISO operator runs the first one during your pilot; until then the inventory shows No collection yet.
- The right role. Owners and Managers assign owners, create, start, verify and close campaigns, and export evidence packs. Contributors and Members can view the inventory and record decisions on review items they are asked to decide. Auditors can view campaigns. The full list is in Roles & Permissions.
What the inventory shows
Click Access & Risk → AI agents to open AI agents and apps. Each row is one Actor: an AI agent, app or service account that has access in the latest collection. People are not listed.
The table has the columns Actor, Owner, Grants, Flags, Observation, Last observed and Actions. Use Search by name, the filters Ownership, Observation, Source and Flag, or one of the presets:
| Preset | Use it to |
|---|---|
| All actors | See everything in the latest collection |
| Without accepted owner | Find agents nobody has taken responsibility for |
| Leaver-linked (investigate) | Find agents tied to someone who has left; check them first |
| Sensitive app | Find apps marked as touching sensitive resources |
| Source unavailable | Find agents where the source could not be read |
A banner at the top tells you about the collection itself: still in progress, incomplete (for example “14 of 20 objects captured”), or the source being unavailable. When a collection is incomplete, items that were not seen are shown as Not observed in last collection.
How to read “Unknown” and “Not observed”
AutoCISO only states what a source showed it.
- Unknown means the source did not give us the data. It says nothing about the access.
- Not observed in last collection describes our collection, not the access. The access may still exist.
- Source unavailable means the source could not be read at that time.
None of these means the access was removed, unused or inactive. For the reasoning, see Absence of Evidence Is Not Revocation.
Open an agent
Click a row to open the agent’s detail page. Its panels are Ownership, Who created and consented, Attention flags, Grants, Credentials and Owner history. Grants lists each permission with its Resource, Type, Reach, Scopes and Last authentication. A grant with no sign-in data shows “Unknown — sign-in data not available”.
Name an owner
- In AI agents and apps, find the agent.
- Click Assign owner (or Suggest owner if you only want to propose one) in the row.
- Choose an employee, add an optional Reason, and save.
The owner badge then reads Assigned — awaiting acceptance. Assigning an owner does not record who approved the access.
The owner confirms in one of two ways:
- Accept link. Select agents that share one assigned owner (up to 50), click Send accept link, choose how many days the link is valid and click Send link. The owner opens the link, signs in and clicks Accept, or Not mine if the agent is not theirs. The owner needs a console login: invite them as a member first.
- Recorded confirmation. If the owner confirmed outside the console, click Record confirmation and enter the date. The badge reads Assigned — confirmation recorded by admin, and the record shows that an administrator entered it.
After the owner accepts, the badge reads Accepted by owner. Every ownership change appears under Owner history. A row marked Not yet in audit trail can be retried with Retry.
Run a review campaign
Click Access & Risk → AI agent reviews to open the list of Access review campaigns. A campaign asks owners to confirm what each AI agent can access and keeps the record. It covers non-human access only; people are not included.
- Click New campaign.
- Enter a Name, the Review period and a Due date.
- Check the Baseline collection. It must be a complete collection. If you see “Run a complete collection first”, wait for one.
- Click Create campaign, open it and click Start campaign.
Starting creates the review items and shows how many were created. The tiles Items, Undecided, Decided, Action claimed, Verified, Not verified, Cannot verify and Audit pending show progress.
Decide
Expand an item to see its grants. For each grant the reviewer chooses Keep, Reduce or Revoke and clicks Record decision.
- The proposal chip (for example Proposed: Keep) is a suggestion from the collected data only. Keep — no adverse signal in window means we saw nothing worrying; it does not mean the access was approved or is needed.
- A Keep on a grant flagged as consequential needs a reason of 10 to 1000 characters.
- Reduce and Revoke need a recorded action. Record what was done and, for Reduce, which permissions stay.
- An owner or manager can record a decision for the reviewer. The banner “You are recording the reviewer’s decision” appears, a source (email, meeting, call or document, a reference and a date) is required, and the record shows who entered it.
Other access paths of the same agent (other principals, credentials, grants) are not part of the decision; the item says so under Residual paths.
Verify
After the change is made in the source system, run another collection and click Verify, then pick the collection to Compare against. The result for each decision is one of:
| Result | Meaning |
|---|---|
| Verified | The later collection shows the change in effect |
| Not verified | The later collection still shows the access, or wider access than decided |
| Cannot verify | The comparison is not trustworthy: the collection was incomplete, the collections cannot be compared, or it is too soon after the change |
Cannot verify is not a failure of the change; it means we cannot tell.
Close
Click Close campaign. The dialog lists likely blockers (items still open, grants without a decision, records not yet in the audit trail). The server makes the final decision; if it refuses, the dialog shows the reasons. Closing records the campaign summary and ends the review.
Export an evidence pack
On a started campaign, owners and managers can click Export evidence pack, choose SOC 2 or ISO 27001, set the Period start and Period end (up to 400 days) and click Generate. You can close the dialog while it runs. When it is ready, click Download; earlier packs are listed under Recent exports until they expire.
The pack is labelled Complete or Partial, and Signed or Unsigned. Read the limits before you hand it to an auditor:
- A signature shows that AutoCISO produced this pack. It does not show that the contents are accurate.
- The pack has no external timestamp.
- The pack states that the campaign excludes people.
- Only English is available for now.
If something does not work
| You see | What it means |
|---|---|
| Not available for your organization | AI Access is not turned on for your organisation. See the note at the top of this page |
| Your role cannot view this page | Your role does not include this page |
| AI access is starting up. Try again in a minute. | The service is still initialising |
| The latest collection run is incomplete. Wait for a complete run before starting a review. | Start the campaign after a complete collection |
| Some items are still open. Decide or reassign them before closing the review. | Finish or reassign the open items, then close |
| Evidence export is turned off for your organization. | Ask your AutoCISO contact |
Was this page helpful?