This piece is by Greg B. Davies, Head of Behavioural Finance, Oxford Risk
The efficiency gains AI brings to the insurance industry are meaningful: faster underwriting, automated claims, precision fraud detection. But while the industry optimises its toolsets for institutional convenience, it is missing AI’s most critical opportunity – and its greatest risk.
Insurance decisions are fundamentally behavioural, not actuarial. They are made under stress, time pressure, cognitive overload, and unfamiliarity with complex products. The customer buying cover during a house move, the policyholder filing their first claim after damage, the business owner reviewing liability under cash-flow strain — this is not laboratory decision-making.
They are also inherently probabilistic. Customers are asked to judge low-probability, high-impact events using statistics that are often counter-intuitive. Even educated individuals misinterpret percentages, relative risk increases, and small probabilities. A “20% increase in risk” sounds alarming, even when the absolute change is small. A 1-in-200 chance feels negligible, even when the consequences are catastrophic.
Insurance decisions are prone to a predictable asymmetry. Customers over-insure small, affordable risks — extended warranties, add-on rental cover, gadget protection — while under-insuring the risks that could permanently damage financial security: life cover, critical illness, longevity risk. Small losses feel vivid and immediate. Catastrophic risks feel distant and abstract.
Behaviourally grounded AI can help correct this imbalance by translating relative risks into absolute terms, highlighting financial impact rather than emotional salience, and drawing attention to protection gaps that truly matter.
Yet AI implementations often focus on automating transactions without questioning whether customers understand their options, whether choices align with actual needs, or whether timing amplifies vulnerability.
Missed Signals
AI systems may collect interaction data: response times, navigation patterns, question frequency. But data do not equal understanding.
Behavioural context means knowing how this particular customer typically responds to risk and complexity. Do they become anxious under pressure? Do they over-weight worst-case outcomes? Do they delay decisions when overwhelmed? Without that context, the same signal can mean opposite things.
A customer repeatedly viewing policy details without progressing, is that diligence or paralysis? A long pause before accepting terms – thoughtful reflection or cognitive overload? Even numeracy can mislead. Customers react very differently to “a 20% increase in risk” versus “from 5 in 100 to 6 in 100,” though the outcome is identical. AI that repeats percentages may amplify confusion; AI that reframes clearly can improve decision quality.
When systems ignore these distinctions, foreseeable harm is the result.
Dynamic Vulnerability
Vulnerability is not a static label. It is situational.
A financially sophisticated business owner may evaluate investments confidently, yet struggle with life insurance decisions following a health scare. Under stress, people simplify. They over-weight recent events, fixate on worst-case scenarios, or default to inaction. In insurance, this can mean over-insuring immediately after a scare, under-insuring to cut short-term costs, or cancelling protection at precisely the wrong moment.
Claims moments are inherently vulnerable. Loss creates distress. Unfamiliar processes create cognitive burden. Outcome uncertainty generates anxiety. When systems misread these states, the consequences are visible: complaints, cancellation spikes, underinsurance, erosion of trust.
Engagement-Sensitive Design
Effective AI must adjust the balance between protection and autonomy based on real-time signals of capability.
When indicators such as confusion markers, decision paralysis, stress-driven timing suggest vulnerability, systems should introduce friction: cooling-off prompts, staged decisions, escalation to human support.
When indicators suggest engagement and comprehension, processes should streamline.
Consider cancellation during financial distress. A behaviourally safe system pauses, clarifies the protection gap, and explores alternatives. The same action from an informed, engaged customer proceeds smoothly. This is not about removing choice. It is about recognising when structure protects better than speed.
Governance: Rules, Not Black Boxes
Behaviourally grounded AI requires a clear separation between what gets decided and how it gets communicated.
Deterministic rules must define what is fair, suitable, and compliant — coverage eligibility, claims criteria, pricing boundaries. These are explicit and auditable. Speed and fluency are not evidence of suitability. If reasoning cannot be traced, validated, and explained, it will not be trusted by regulators or customers.
AI should orchestrate communication within those constraints: adjusting explanation to comprehension signals, pacing decisions appropriately, and escalating to human judgment where required.
“Computer says ‘no’” is not an explanation.
Consumer Duty and Competitive Reality
The FCA’s Consumer Duty requires good outcomes and avoidance of foreseeable harm. AI optimising purely for institutional efficiency cannot satisfy that standard.
AI that identifies inadequate coverage under pressure, that reframes risk clearly, that adapts to comprehension and vulnerability, operationalises Consumer Duty. It demonstrates that firms have considered understanding, capability, and harm, not merely throughput.
Insurance already suffers from scepticism about fairness and transparency. AI that prioritises speed over clarity deepens that scepticism. Each opaque interaction erodes trust incrementally.
The alternative is AI as trust infrastructure: systems designed to recognise when customers need support, communicate clearly, protect against foreseeable poor decisions, and align institutional incentives with customer outcomes.
Pricing and products increasingly converge. Differentiation will come from how firms behave when customers are under strain.
AI in insurance is inevitable. The question is not whether it will be used, but whether it will be used responsibly. Systems built for speed will scale confusion. Systems built for behavioural safety will scale trust. The difference will define the industry.
Be the first to comment