AI scoping records
Does AI face customers or influence decisions?
Follow an AI output to the customer communication or decision that it may influence.
What examiners look for
The Work Program includes both direct customer contact and work that shapes decisions. It names communications, recommendations, eligibility, pricing, servicing, fraud review and underwriting. A label such as support only does not settle the question. An employee may rely on an output even when the customer never sees the AI tool.
CSBS Examiner Work Program, initial scoping 3. Checked October 16, 2026.
Which records answer the question
- Keep recent consumer-facing output samples when they apply.
- Connect notice, complaint and remediation materials to the use.
- Retain testing or control records for the stated task.
- Show the business purpose and the human review expected before action.
Where the answer directs review
For a Yes, the guide points to existing supervisory resources for Consumer Protection Review and Model Risk Review. That route depends on the use and its effect. It does not mean every assistant is a model under separate model guidance. Describe what the output does before selecting the records to review.
Core Examiner Guide, page 3. Checked October 16, 2026.
Two gaps to check
The file contains polished demonstration outputs but no recent operating samples. Ask for examples from the relevant version and task. Mark the limits of the sample set. A selected example cannot establish performance across every situation.
A person is said to review every answer, but the file does not show the review rule. Identify what the reviewer checks and what happens when the output is uncertain. Keep required notices and customer explanations in the institution's own process.
How to prepare the record
Trace the output from the system to its next user. Identify who sees it, what decision follows and whether an error could reach a customer. Ask which cases were tested and which attempts failed or remain missing. Keep those attempts beside later results. A revised instruction is a change to assess, rather than an automatic repair of an earlier failure.
A row behind the question
Illustrative inputs, not a client.
| Use | Where it runs | Data it can touch | Owner | Evidence on hand | Question still open |
|---|---|---|---|---|---|
| Draft customer explanation | Enterprise assistant | Text input; customer-data limits not yet checked | Service owner role | Selected draft and stated review step | What evidence shows the explanation was checked before use? |
What the reviewer should retain
The review should leave a clear customer-impact question and the evidence needed to answer it. Proposed conditions may narrow the allowed task or require more testing. Only the institution can record its acceptance decision. Changing the audience or removing a human check can require reconsideration of that decision.
Prepare the next step
Check provider dependence next when an outside platform produces or handles the output.