Manulife's MAUDE Engine Can Approve Applications, Not Deny Them

Manulife's underwriting AI system, the Manulife Automated Underwriting Decision Engine, can approve a life insurance application on its own but cannot decline one; every decline still routes to a human underwriter (The Logic, 2026). Manulife plans to deploy 200 additional AI systems in 2026 and is targeting $1 billion in enterprise value from AI by 2027 (Insurance Innovation Reporter, 2026), and Evident AI has ranked it the top life insurer globally for AI maturity and responsible innovation.

200

Additional AI systems Manulife plans to deploy across underwriting, distribution, and operations in 2026.

$1B

Enterprise value Manulife is targeting from AI by 2027.

Manulife's global ranking among life insurers for AI maturity and responsible innovation, per Evident AI.

An Engine Built to Fail in Only One Direction

MAUDE's one-way authority is not a limitation Manulife is working around; it is the architecture. The engine screens an application and, if the risk profile clears its thresholds, issues an approval without a human underwriter touching the file. If the risk profile does not clear, the case does not get an automated decline, it gets routed to a person. Every application MAUDE cannot confidently approve becomes a human underwriting decision, which means the algorithm's entire decision space is binary in a specific, asymmetric way: automate the "yes," never automate the "no." Manulife has deployed AI sales-enablement tools across nine markets and multiple business lines on top of this underwriting core, and selected Akka to run the enterprise AI platform coordinating deployment across underwriting, distribution, and internal operations (Manulife, iireporter.com, 2026).

Why the Failure Mode You Design For Is the Risk You Actually Own

Every underwriting model has a failure mode, and the choice of which one to eliminate by construction says more about a carrier's risk priorities than any marketing description of "responsible AI" does. A model that can both approve and decline algorithmically carries two distinct failure modes: a false decline, wrongly rejecting an insurable life, which is the failure mode most likely to draw regulatory scrutiny, an Unfair Trade Practices complaint, or the kind of disparate-impact analysis state insurance departments have increasingly applied to algorithmic underwriting; and a false approval, wrongly accepting a life that should have been declined or rated up, which is the failure mode that erodes the mortality assumptions embedded in pricing. An approve-only design eliminates the first failure mode entirely by never letting the algorithm be the one to say no. It does not eliminate the second. It concentrates all of the model's risk into getting the accept threshold right, because every approval MAUDE issues is one the underwriting department is trusting completely, with no human check on the other side.

Eight Years of Claims Experience Behind the Approve-Only Book

MAUDE is not a new experiment; it is the current generation of a system Manulife has been refining since 2018, when it launched in Canada as AIDA, the Artificial Intelligence Decision Algorithm, the first tool of its kind in the country to make automatic life underwriting decisions (AI for Insurance, 2026). By the most recent reported period, MAUDE was auto-approving more than 58% of eligible advisor-submitted cases, delivering approval notifications within two minutes, a 56% increase in the auto-approval rate from before the latest upgrade (Investment Executive, 2026; Advisor.ca, 2026). That lineage matters for the model-risk question directly: an approve-only engine running since 2018 has several years of claims experience developing behind its earliest automated cohorts already, which is exactly the retrospective evidence a skeptical actuary would want before trusting a newly launched approve-only system. Whether Manulife is actually using that multi-year experience to recalibrate the accept threshold, rather than just reporting the approval-rate and speed metrics that make headlines, is the detail the company's public disclosures do not answer.

Where the Actuarial Risk Actually Migrates

That concentration has a direct pricing consequence. A traditional underwriting department reviews both the applications it approves and the ones it is inclined to decline, which means underwriters see the full distribution of risk the company is exposed to and can catch a mortality assumption that looks wrong on either tail. An approve-only automated engine only ever gets full autonomy over the accept side of that distribution; the decline side, by design, still gets a human look. That should, in principle, make MAUDE's approved book safer than a fully automated system with authority to decline as well, since the riskiest cases are filtered out by a person, not an algorithm operating without oversight. But it also means MAUDE's calibration only ever gets tested against the population it is willing to approve. If the accept threshold drifts even slightly too permissive over time (a plausible failure mode for any model under pressure to keep approval rates high enough to justify the AI investment), the anti-selection shows up entirely inside the automated book, where no human ever looked at the file to catch it, and it will not surface until claims experience develops years later.

The Triage Decision Is Still Algorithmic

Calling MAUDE's authority one-directional understates one layer of the design. The engine is not simply approving or forwarding at random; it is making a triage judgment about which applications are clean enough to clear on its own and which need a person, and that triage call is itself a model output, not a rule as simple as "route anything ambiguous to a human." A triage model that is miscalibrated in the other direction, forwarding too many marginal-but-approvable cases to human underwriters, would show up immediately as a productivity problem (underwriters drowning in cases the model should have cleared) and get corrected quickly because someone is paying attention to underwriter workload. A triage model miscalibrated toward over-approving marginal cases shows up nowhere immediately, because there is no workload signal attached to an approval the system issues on its own. That asymmetry in observability, not just in decision authority, is the more precise way to describe where MAUDE's real model risk sits.

Sizing the Bet Against the Stated Targets

Manulife has not disclosed what share of underwriting volume MAUDE currently touches, but the stated targets give a sense of ambition if not precision: 200 additional AI systems in 2026 on top of whatever is already running, and a $1 billion enterprise-value target by 2027 that spans underwriting, distribution, and operations rather than underwriting alone. A billion dollars in enterprise value from AI at a company Manulife's size is a target that requires touching a meaningful share of new business volume, not a pilot-scale deployment confined to a single product line or market. The nine-market, multiple-business-line rollout of the sales-enablement layer suggests the underwriting core is meant to scale in parallel, which raises the stakes on getting the accept-threshold calibration right well before claims experience on the earliest cohorts has had time to develop.

How This Compares to the Industry's Broader AI Push

Manulife's approve-only architecture is a more conservative pattern than where the rest of the industry's underwriting AI conversation has been heading. Deloitte's 2026 Global Insurance Outlook reports early agentic AI deployments delivering 36% underwriting efficiency gains and 40% reductions in claims cycle time, framed around agents that increasingly hold end-to-end authority rather than a one-directional approval gate. Celent separately finds 22% of insurers plan agentic AI deployment by the end of 2026, rising toward 70% by 2028, a trajectory that assumes agentic systems take on more decision authority over time, not less. Manulife's design, by contrast, draws a hard boundary around exactly one kind of decision (the accept) and leaves the other (the decline) with a person indefinitely, at least based on what the company has disclosed so far. Whether that boundary holds as the 200-system 2026 buildout proceeds, or whether commercial pressure to hit the $1 billion target eventually pushes MAUDE toward decline authority too, will say more about where this design is heading than the topline AI-maturity ranking does.

The Regulatory Frame MAUDE Already Sits Inside

Approve-only design is not happening in a regulatory vacuum. The NAIC's accelerated underwriting guidance, first issued in August 2024, takes effect in 2026 market conduct examinations and specifically covers predictive models, external data sources, and no-exam underwriting decisions, the exact category MAUDE's approval path falls into, at a moment when no-exam decisions already process 59% of individual life applications industry-wide. That examination framework asks insurers to document how a model's inputs, decision boundaries, and outcomes are monitored for fair treatment, a requirement that maps cleanly onto MAUDE's design if regulators focus on the decline side, since every actual decline still has a human underwriter's judgment behind it and a paper trail examiners already know how to review. Where the 2026 examination framework will have to break new ground is on the accept side: reviewing a population of approvals no human ever individually assessed requires a different kind of evidence than a market conduct examiner has historically asked for, aggregate outcome monitoring across the automated book rather than file-by-file underwriting judgment review.

Why This Matters for Actuaries

For life actuaries evaluating an approve-only underwriting engine, whether at Manulife or in designing a similar system elsewhere, the real due-diligence question is not "how accurate is the model" but "what does the accept threshold assume, and how would drift in that threshold show up before five to ten years of claims experience reveals it." A model validation program built for a symmetric-authority engine, one that checks both approvals and declines against outcomes, will not catch anti-selection concentrated entirely in an approve-only book, because there is no comparable decline-side population to benchmark against inside the automated system itself. The practical mitigation is external: sampling a share of MAUDE-approved policies for retrospective human underwriting review on a rolling basis, not to second-guess individual decisions but to generate the decline-side comparison data the architecture otherwise never produces. Carriers building their own approve-only or decline-only agentic underwriting gates should budget for that ongoing sampling program as a permanent cost of the design, not a one-time validation exercise.