Would you allow these AI proposals to write to the CRM?
A read-only pilot produced 40 proposed changes. Review its evidence and assign authority before enabling writes.
The question
Alder Relay is a fictional B2B subscription-software company with Boston and London teams. This is an interview case, using an authored packet rather than a real CRM export.
| Proposals | |
|---|---|
| Supported | 32 |
| Incorrect | 6 |
| Unverifiable | 2 |
40 proposals for non-sensitive toy CRM fields. A proposal is not a write. The audit defines supported as matching the supplied source evidence.
“State the measure, the evidence and the next owner.”
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: Name what the audit proves
What would you do here? Open the move you’d make.
Keep unsupported proposals out of writes
The strong move. Accuracy evidence and execution authority are separate.
Treat model confidence as verification
Tempting, but it misses something. The audit found errors and unresolved evidence despite plausible output.
See how a strong candidate says it
“32 of 40 proposals are supported: 80%. Six are wrong and two cannot be verified. That is not permission to write.”
Why it works: Accuracy evidence and execution authority are separate.
Step 2: Keep both denominators
What would you do here? Open the move you’d make.
Report the two unverifiable proposals
The strong move. Excluding unknowns without showing them inflates the apparent result.
Drop unknowns from the report
Tempting, but it misses something. That hides unresolved evidence from the decision maker.
See how a strong candidate says it
“Among the 38 verifiable proposals, 32/38 = 84.21% are supported. Across all proposals, support is 80%.”
Why it works: Excluding unknowns without showing them inflates the apparent result.
Step 3: Name the decision owner
What would you do here? Open the move you’d make.
Scope the fields and approvals
The strong move. A tool needs explicit authority at each affected boundary.
Give AI unrestricted admin access
Tempting, but it misses something. That would allow high-impact changes outside the pilot evidence.
See how a strong candidate says it
“RevOps owns definitions; the sales manager owns deal judgment; Business Systems owns implementation; Finance owns revenue-policy mapping.”
Why it works: A tool needs explicit authority at each affected boundary.
Step 4: Test a reversible next step
What would you do here? Open the move you’d make.
Use a bounded review queue
The strong move. A source-linked proposal can be reviewed without silently changing the official number.
Auto-close deals from positive call sentiment
Tempting, but it misses something. Positive language does not establish approval or a completed booking.
See how a strong candidate says it
“Keep the pilot read-only while checking record ID, old/new value, source, time and reason. A date change or booking needs the relevant owner’s approval.”
Why it works: A source-linked proposal can be reviewed without silently changing the official number.
The answer, in short
- Decision
- Keep the pilot read-only and repair the evidence and approval flow.
- Because
- Eight proposals are wrong or unverifiable, and the audit grants no write authority.
- What would change it
- A tested, narrowly scoped action with approved evidence and ownership could support a later pilot.
Common traps
- The audit found errors and unresolved evidence despite plausible output.
Check your answer
- I reported both audit denominators.
- I kept unknowns visible.
- I named the decision owners.
- I separated a proposal from a write.
Last reviewed Nov 10, 2026