Manulife

Life insurance underwriting has long struggled with the tension between thorough risk assessment and application speed. At Manulife, one of Canada's largest insurers, a significant portion of incoming applications were straightforward enough to approve algorithmically — yet they still consumed underwriter time. The original AI questionnaire collected excessive medical information regardless of relevance, degrading both the applicant experience and model accuracy. Meanwhile, new underwriters were being assigned routine cases that offered little professional development. The cumulative cost: slower approvals, underutilized talent, and an automation ceiling well below what the underlying data could support.

Manulife overhauled the questionnaire powering Maude (Manulife Automated Underwriting Decision Engine), its predictive ML platform first deployed in 2018. The redesigned intake uses an open-ended, conversational format that adapts dynamically to each applicant — adjusting questions based on age, requested coverage amount, and prior responses. Irrelevant medical questions are skipped automatically, reducing applicant burden by up to 40%. Agents can select from standardized lists for medications, existing conditions, hobbies, and travel, giving the predictive model cleaner, more consistent inputs. Applications that fall outside Maude's approval criteria — complex medical histories, affordability mismatches, or novel risk profiles — are automatically routed to human underwriters. The system currently underwrites term policies up to $2M and permanent products up to $500K for applicants aged 60 and under, with no external vendor disclosed.

Since the September 2025 questionnaire update, Maude's instant approval rate nearly doubled to 58% of all eligible applications processed — a material jump from the prior baseline. Qualifying applicants receive a decision in as little as two minutes. Key outcomes include:

• 58% instant approval rate across eligible applications (up from roughly half that)
• Up to 40% reduction in medical questions per application
• ~2 minutes end-to-end approval time for straight-through cases
• Human underwriters now focus exclusively on complex, higher-value cases
• New underwriters report more intellectually demanding workloads from day one, improving skill development and job satisfaction

• Questionnaire design is a first-order ML problem: removing irrelevant inputs improves model signal as much as adding training data.
• Standardizing free-text fields (medications, conditions, hobbies) into predefined lists directly increases predictive accuracy without changing the underlying model.
• Graduated authority — starting with narrow approval scope and expanding as the model proves itself — mirrors how human underwriters earn trust and is essential for regulatory and actuarial credibility.
• A hybrid routing architecture (AI approves standard cases, humans handle exceptions) preserves underwriter expertise for edge cases the model cannot yet be trained on.
• Product adjacency matters: data availability, not technical complexity, determines which lines to automate next.