Insurers Are Concerned About AI Bias, But Accept It's The Future
Every data source carries assumptions. That's fine until a model starts making calls on pricing, risk, or claims. One pattern in a cohort doesn't define an individual. Precision is the point of AI, but precision without control creates new exposure.
What the new EIP research says
Fresh research from embedded insurance provider EIP surveyed 250 senior insurance professionals across the UK and Europe. The signal is clear: concern is high, adoption is higher.
- Bias risk: 87% are concerned about bias or unfair outcomes in AI-driven processes.
- Near-term automation: 90% expect end-to-end claims administration to be AI-managed within 24 months. Respondents also cited an average expectation of 15 months.
- Top risks: Data security and privacy (23%), regulatory non-compliance (21%), system reliability and errors (21%), job displacement/staff resistance (20%). For larger firms, regulatory risk jumps to 33%.
- Least comfortable to automate: Claims submissions (40%), underwriting recommendations (39%), customer interactions (35%).
- Governance levers that help: Transparent algorithms and decision logs (39%).
- Impact confidence: 100% expect sizeable impact; 58% expect major efficiency gains; 33% expect full transformation.
Where insurers are already investing
- Customer service and chatbots (45%)
- Risk modelling and data analytics (43%)
- Claims management (41%)
- Underwriting and pricing (40%)
- Fraud detection (40%)
- Marketing (38%)
Human in the loop stays non-negotiable
- 99% want some level of human oversight for AI-driven outcomes.
- Selection drivers: ease of integration (27%) and partnership/support (27%) top the list.
- Cost/ROI is a weak barrier: only 10% say it strongly influences selection.
Quotes from the industry
Ross Sinclair, Founder and CEO of EIP: "Insurance is, by necessity, a heavily regulated industry, which makes introducing AI complicated. This research shows that the development of industry-specific capabilities could help to overcome challenges and accelerate the adoption of AI right across the industry, whether that's for customer service, risk modelling or even claims management and decisioning."
On claims: "Claims decisions need to be consistent, predictable, transparent and auditable, which makes them unsuitable for the subjective and opaque nature of AI tools that are based on probabilities. This is where an insurance-specific framework can make a real difference. The industry needs to utilise rules-based decisioning systems configured by insurers, that adhere to their specific policies and compliance requirements. This will allow them to create systems that are enhanced by AI but ultimately leave the final judgement to be made by a human."
David Mitchell-Dawson, Director of Product at EIP: "It is important to remember that bias is an issue for insurers regardless of whether they are using AI or not. It has always been a challenge for the industry. A rules-based engine addresses this because it is not subject to any inherent biases."
On pace: "We are pleased to see that the industry is overall bullish on the power of AI... Regulators too are moving fast. Insurers told us they're expecting their full end-to-end claims administration to be handled by AI within the next 15 months on average, so frankly if you're not doing anything now, you're already too late."
What this means for carriers and MGAs
AI is coming deeper into the stack. Bias, privacy, audit, and reliability will decide who benefits and who gets flagged by customers and regulators. The play is a controlled architecture: rules for determinism, models for speed and scale, humans for judgement.
Practical steps to de-risk adoption now
- Stand up model governance: Define roles, approval gates, and documentation standards. Log datasets, features, versions, prompts, and overrides for full audit trails.
- Build a hybrid decisioning layer: Pair rules-based engines for policy compliance with ML for triage, document extraction, fraud scoring, and next-best-action. Keep final decisions human for higher-risk outcomes.
- Measure fairness and accuracy: Set monitored thresholds for disparate impact, error rates, and override frequency. Trigger review when thresholds are breached.
- Instrument explainability: Provide reason codes, feature attributions, and decision logs visible to handlers and, where appropriate, customers.
- Data security by design: Minimise PII, control vendor access, use redaction and synthetic data for training and testing.
- Control the blast radius: Start in low-risk flows (FNOL intake, document classification, subrogation cues). Graduate to pricing recommendations and claims decision support once controls are proven.
- Vendor diligence: Require model cards, bias testing evidence, security attestations, uptime SLAs, and rollback plans.
- Train the frontline: Upskill claims handlers, underwriters, and QA on reading AI outputs, challenging recommendations, and recording overrides.
- Track business impact: Monitor cycle time, settlement accuracy, leakage, complaint rates, fairness metrics, and loss ratio movement.
Claims deserves extra care
Use AI for classification, verification, fraud signals, and document routing. Keep adjudication rule-driven, with clear thresholds for human review. That preserves consistency while still reducing cycle time and admin load.
Compliance signals to watch
- EU AI Act: risk-based obligations, transparency, and governance expectations.
- ICO guidance on AI and data protection: lawful basis, fairness, and explainability for UK operations.
If you're starting now
- Pick two use cases with clear ROI and low customer harm. Ship in 90 days with governance in place.
- Stand up a unified decision log across underwriting and claims. You'll need it for audits and complaint resolution.
- Publish an internal AI policy with guardrails. Make exceptions rare and reviewable.
- Create a cross-functional review board (actuarial, claims, legal, compliance, security). Meet monthly.
Resources
The industry sees the risk. It also sees the upside. Build for control, ship small, and keep humans in the loop. That's how you get the gains without the backlash.
Your membership also unlocks:
