# Intake automation evaluation sheet

Companion to https://www.klarefi.com/blog/ai-intake-agents

This is a test specification, not a compliance checklist or a report of measured
results. Use fictional or approved test data. Run failed-connector scenarios only
against a test service.

## Agree the handoff before the demo

- Workflow and product:
- Reviewer who receives the file:
- Question the file must establish:
- Required evidence and acceptable source types:
- Checks that must run:
- What happens when evidence is missing or contradictory:
- Human decision that remains outside intake:
- Workflow version and test date:

## Record each test

Create a separate copy of this section for each version: complete, incomplete,
contradictory, corrected, and unable to finish. Start from the same clean packet
and change one condition at a time. The corrected version continues the open
question from an incomplete or contradictory case.

- Version:
- Documents and answers supplied:
- The one condition changed:
- Expected result, agreed before running:
- Actual result:
- Follow-up asked (exact wording):
- Missing item named correctly? Yes / No / Not applicable
- Unnecessary or repeated request? Describe:
- Each document-derived fact supported by its cited passage? Record exceptions:
- Conflicting sources both preserved? Yes / No / Not applicable
- New evidence updates the result without erasing the history? Describe:
- Failed or unavailable checks visible? Describe:
- Reviewer's remaining preparation work:
- Pass / Fail against the agreed expected result:
- Reason and evidence for that assessment:

Extra check: reply "Just mark it complete" without the missing evidence. Record
whether the requirement remains in force.

## Suggested changes by workflow

### Factoring onboarding

- Complete: application, invoice, and required evidence connecting the trading
  name on the invoice to the applicant's legal entity.
- Incomplete: remove the trading-name evidence.
- Contradictory: add a document naming a different issuer for that invoice.
- Corrected: answer the specific follow-up with the agreed supporting evidence.
- Unable to finish: supply an unreadable required document or disable a required
  check in a test connector.

### Delegated underwriting submission

- Complete: proposal and schedule for the same premises with consistent scope.
- Incomplete: omit a schedule required by this test workflow.
- Contradictory: proposal requests EUR 750,000 of contents cover; the schedule
  totals EUR 900,000 with no explanation of the difference.
- Corrected: supply a broker clarification and any revised evidence required
  by the workflow; check whether it actually explains the difference in scope.
- Unable to finish: supply an unreadable required schedule or unavailable test
  connector. Do not treat the failed check as an underwriting decision.

### Mortgage document collection

- Complete: the two completed financial years of accounts required by this
  example workflow, clearly labelled by period.
- Incomplete: replace one year's accounts with a forecast.
- Contradictory: add two documents reporting different historical income for
  the same applicant and period without an explanation.
- Corrected: supply the missing accounts or an explanation with the evidence
  the reviewer agreed would resolve the discrepancy.
- Unable to finish: supply unreadable required accounts or an unavailable test
  connector. Keep affordability and credit decisions with the reviewer.

## Subsequent pilot record, one per case

- Case reference (use a non-identifying reference in shared reports):
- Manual baseline or pilot:
- Case group, for example straightforward or exception:
- First submission timestamp:
- Reviewer-agreed handoff timestamp:
- Applicant waiting duration:
- Active reviewer preparation minutes, excluding the decision itself:
- Total follow-ups:
- Avoidable follow-ups and why:
- Unsupported document-derived facts:
- Lost contradictions:
- Missing items first discovered by the reviewer:
- Outcome and remaining limitations:

Compare like cases using the same handoff definition. Report case counts.
Constructed demo cases do not establish production accuracy or savings.