Straight-Through Small Commercial: The 75% Threshold Proposal
Keystone Commercial has moved 70% of its small business volume, BOP and workers' compensation, to a straight-through process: agents submit via API, and a rules engine plus a predictive model returns bindable quotes in under two minutes when a submission scores below a risk threshold, referring the rest to underwriters. Since launch, the straight-through rate is 58%, quote-to-bind is up nine points, and time-to-quote has fallen sharply. But the underwriting quality team found that 6.3% of straight-through BOP policies carried a class code inconsistent with the business description, and declines are concentrated among businesses under two years old and a set of industry codes. The referral team also noticed, in a sample of 40 files, that agents re-submitted declined accounts with slightly altered business descriptions to get a bindable score. Loss experience is not yet credible. The chief underwriting officer wants to raise the straight-through threshold to 75% to hit the growth target.
1. Assess whether the straight-through program is actually performing. What metrics and tests would you require?
Read the worked answer
The reported numbers are throughput metrics, not quality metrics. A 58% straight-through rate and faster quoting say nothing about whether the decisions are right.
Required measures:
Referral overturn rate. How often does an underwriter materially change a straight-through decision on audit? A low rate supports the model; a high rate means the gate is mis-set.
Class-code accuracy. 6.3% is a premium-leakage and coverage-mismatch number, not a rounding error. Convert it into dollars of mispriced premium and identify whether errors cluster in particular industries or in submissions that came through a particular agent channel.
False-decline cost. For declined accounts that were later written elsewhere or re-submitted successfully, what was the lost premium? Concentration among new ventures and specific industry codes can be legitimate hazard differentiation or can signal thin training data, a bad SIC-to-class mapping, or reliance on inputs that proxy for something else. Test which.
Outcome cohorts. Early loss ratio and cancellation rates by straight-through versus referred business, tracked by month, with a caveat that credibility is low.
Gaming indicators. Near-duplicate submission rates by agency and by applicant identity.
Present these as a scorecard. A program that grows premium while misclassifying 6.3% of policies is buying growth with future claims disputes and market-conduct exposure.
2. The re-submission pattern is confirmed in the sample. What does it indicate and how should Keystone respond?
Read the worked answer
The pattern indicates the model is highly responsive to free-text business description and that agents have learned the trigger vocabulary. That is a rational response to an incentive Keystone created: a decline is a dead end, so agents optimize the input. The consequences are adverse selection on the retained book and a data-integrity problem that contaminates the model's own training data.
Response, in order:
Close the loop. Lock and hash the description field after first submission; any change requires a documented reason and triggers referral. Detect near-duplicates by FEIN, address, owner name, and phone within a rolling window and route them to a human queue. Require agent attestation on the business description with a stated consequence for misrepresentation.
Verify independently. Use an external source to confirm class code: NAICS lookup, business website review, or inspection for higher-hazard classes. Do not rely on the applicant's own text for the rating classification.
Fix the incentive. Make the referral path fast and predictable, with a two-hour service standard and a clear explanation of what would make the account eligible. Agents game gates that feel arbitrary; they use gates that feel fair.
Monitor at the agency level. Track submission alteration rates and referral overturn rates by agency, and address outliers through the appointment agreement rather than through silent score adjustments. Also re-audit the 6.3% class-code error rate to see how much of it traces to gamed submissions.
3. How would you respond to the CUO's proposal to raise the threshold to 75%?
Read the worked answer
Recommend against raising the threshold now, and separate the growth target from the risk gate.
Reasoning: the straight-through rate and the error rate are not independent. If 6.3% of policies are misclassified at a 58% rate, lifting the gate to 75% will admit submissions the model handles least well, so the error and leakage rates should rise, and they will concentrate in the segments already showing disproportionate declines. Because loss experience is not credible, there is no way to know whether the additional volume is profitable. The proposal trades an unmeasured future loss for a near-term premium number.
Alternative path:
Grow through capacity, not through the gate. Add referral underwriting capacity and shorten referral turnaround; that raises bound premium without lowering the quality bar.
Controlled experiment. If the CUO wants evidence, run a champion/challenger: raise the threshold for a defined segment only, with pre-agreed stop triggers such as class-code audit error above 4%, or early loss ratio exceeding plan by a set margin.
Revisit on credibility. Set a date when 12-18 months of cohort experience is available.
Decision rights. If the CUO still wants to proceed, that is a risk-appetite decision requiring documented acceptance by the chief risk officer, with the residual risk stated in dollars, and a board or committee notification. The conversation should be about which risks the company is choosing, not about whether 75% sounds achievable.