Skip to main content
LLM Squared

AI scoping records

What sensitive information passes through AI?

Map the data an AI use can receive, reach or transmit, including access through connected tools.

What examiners look for

The Work Program covers customer, consumer, confidential and other sensitive data. The question extends beyond what an employee pastes into a chat box. It includes data processed, sent or accessible to the tool. Check data types and access paths for this use. Mark what remains unknown.

CSBS Examiner Work Program, initial scoping 8. Checked October 16, 2026.

Which records answer the question

  • Use data classifications and tool-use guidance together.
  • Keep provider records describing data handling and access.
  • Retain terms about use, retention, sharing and training.
  • Link access-control and monitoring evidence to the use at issue.

See all 19 requested records.

Where the answer directs review

A Yes can extend attention to governance, inventory, data controls, vendor access and consumer impact. Privacy, data-security and cybersecurity resources may also apply. Generative AI review is relevant when that capability is involved. An answer alone does not prove approved access. It also does not prove that safeguards were tested.

Core Examiner Guide, page 3. Checked October 16, 2026.

Two gaps to check

The policy forbids sensitive inputs, but an assistant can search a shared repository. Identify the documents and data the connection can reach. Ask for evidence that the approved access boundary matches the actual setup. A written ban alone does not demonstrate the behavior of a connection.

The vendor describes retention without addressing training or upstream sharing. Keep those questions separate and request the missing terms. Record the relevant product, service and contract version. An answer about one data practice does not answer every handling question.

How to prepare the record

Describe categories instead of collecting member records for this public preparation exercise. Ask who authorizes access, what the tool can retrieve and how access is removed. Include who built the model and who can stop an agent where relevant. Keep a returned access description separate from evidence of a tested limit. Record testing only when it actually occurred in an approved setting.

A row behind the question

Illustrative inputs, not a client.

An illustrative evidence row
UseWhere it runsData it can touchOwnerEvidence on handQuestion still open
Assistant connected to procedure documentsEnterprise service and document connectionReach of the connection not yet checkedAccess owner role not assignedPolicy and stated connection settingsWhich data can the tool reach beyond the approved procedure set?

What the reviewer should retain

The kept record should name the boundary, the evidence behind it and the unanswered handling questions. It should also identify the owner who can return more evidence. A new source connection, permission or provider term can reopen the review. Keep that new review visible while the institution decides what access to accept.

Prepare the next step

Return to the hub to assemble the records that apply to the institution's uses.