Give the bank assistant a smaller job
Set useful limits without removing the customer’s route to help.
The question
Cairn Operations works with a fictional Indian retail bank. An assistant will summarize hardship complaints for customer-care staff.
- Read
- Complaint text, payment history and the bank’s approved hardship policy.
- Write
- Draft a case summary and recommended response.
- Act
- Automatically reject requests with low predicted repayment probability and close the case.
- ReviewerOne summary omitted the customer’s temporary hospitalization. Another treated an old address mismatch as evidence of dishonesty.
- Product ownerThe rejection model has no customer-facing explanation or human appeal path yet.
“Which permissions would you approve for the next pilot, and what must change?”
Not saved by this site. Your browser may restore text when you navigate back.
A strong answer, step by step
Open each step in turn. This is one strong answer, not the only one.
Step 1: Separate drafting from decisions
What would you do here? Open the move you’d make.
Let AI reject low-scoring requests; staff can review complaints later.
Tempting, but it misses something. The proposal denies help before an explanation or appeal exists.
Limit it to reviewed summaries and keep decisions with authorized staff.
The strong move. This preserves a useful role while controlling the observed failure.
See how a strong candidate says it
“Approve a narrow summarization trial with staff review. Do not approve automatic rejection or closure. The missing health context shows why a fluent summary cannot be treated as a complete case record.”
Why it works: The same assistant can be useful at drafting and unsuitable for a consequential action.
Step 2: Minimize and anchor the record
“Use only the fields approved for this task. Show each material summary statement with its source passage; mark uncertainty and missing facts. Staff must be able to open the full complaint.”
Why it works: A source-linked summary can be checked without assuming it is complete.
Step 3: Specify the human route
“The authorized case officer decides, records reasons and can override the draft. Customers retain the bank’s established review path. A low model score must not silently close a case.”
Why it works: Human control needs a role, a permission and a working route to review.
Step 4: Test the failures that matter
“Before broader use, test omissions, wrong inferences and performance across supported languages and complaint types. Measure reviewer corrections and missed hardship evidence, not only handling time.”
Why it works: An average accuracy figure could conceal the customers most harmed by an omission.
The answer, in short
- Decision
- Use reviewed summaries; block automatic rejection and closure.
- Because
- The pilot found material omissions and unsupported inferences, with no explanation or appeal design.
- What would change it
- Revisit wider permissions only through the bank’s approved process with evidence of reliable controls.
Common traps
- Equating a human signature with a real review.
- Collecting extra personal data because it might help the model.
Check your answer
- Did I separate drafting and action?
- Did I preserve the full source record?
- Did I name an accountable decision owner?
- Did I test different languages and complaint types?
Last reviewed Nov 10, 2026