Agentic AI is changing insurance - gradually
A September 2025 report from Economist Impact, sponsored by SAS, signals a shift from small efficiency gains to measurable outcomes as agentic AI becomes practical inside insurers.
These systems can handle defined tasks end-to-end - responding to emails, updating databases and transforming data - with controls that keep them inside guardrails.
What changes and what stays human
Insurer workforces will look different. As Jodie Wallis, global chief analytics officer at Manulife, notes, insurer teams will become "hybrids of human employees and agents collaborating closely, with some agents working largely independently under human oversight".
The model is clear: agents do the high-volume work; people handle judgement, exceptions and relationships.
Early productivity, slower cost takeout
Early adopters are seeing the biggest productivity improvements. But cost savings lag until firms wire agents into full workflows - policy admin, claims, finance and compliance - not just single tasks.
That requires orchestration, secure connectors, data quality, audit trails and human-in-the-loop checkpoints.
Use cases you can run now
- Claims intake and triage: classify FNOL, extract entities from emails and forms, route to the right queue, draft responses.
- Subrogation and leakage: scan adjuster notes, repair invoices and photos to flag recovery and payment anomalies.
- Underwriting prep: pre-fill submissions, summarize broker emails, surface missing data, draft endorsements for review.
- Policy servicing: address simple endorsements, billing inquiries and proof-of-insurance requests with human approval on exceptions.
- Regulatory reporting: assemble narratives and evidence packs from logs and case systems for compliance teams to sign off.
Expert view from SAS
Andrew Pollard, insurance specialist at SAS UK and Ireland, explained: "Agentic AI has the potential to fundamentally change how insurers operate - not by replacing expertise, but by complementing it.
"These agents will accelerate repetitive and data-intensive work and when orchestrated together, they can unlock new levels of productivity and insight. But the real value emerges when insurers combine them with trusted, domain-specific solutions and human expertise to tackle complex challenges like risk modelling or regulatory compliance."
Pollard added: "Agentic AI in particular is moving us beyond pilots into enterprise-wide transformation. These agents can operate autonomously or in concert with people, delivering faster service, reducing leakage and improving customer experiences.
"Most importantly, they free human teams to focus on creativity, judgement and relationship-building - the areas where people make the biggest impact."
Controls regulators and boards will expect
- Clear role definition: copilot vs. autonomous agent, with approval thresholds and stop conditions.
- Human-in-the-loop on high-risk actions (payments, coverage decisions, denials, regulatory filings).
- Audit logs: who did what, when and why, including source data, prompts, outputs and overrides.
- PII safeguards: redaction, tokenization and data minimization across email, documents and chat.
- Model risk management aligned to frameworks like the NIST AI RMF, including bias checks and scenario testing.
- Monitoring and rollback: output quality, error rates, drift and an exit plan for bad behavior.
Build this before you scale
- Agent orchestration layer: task queueing, retries and dependency handling across systems.
- Secure connectors: email, claims, policy admin, CRM, data lakes and document stores.
- Knowledge sources: underwriter guidelines, playbooks and regulatory rules with version control.
- Guardrails: policy-centric prompts, retrieval, rate limits and safety filters.
- Observability: dashboards for latency, success, overrides and leakage.
- Change management: role design, training, comms and a feedback loop with adjusters and underwriters.
Metrics that matter
- Claims: cycle time, LAE per claim, leakage, straight-through-processing rate, rework rate.
- Underwriting: time to quote, submission touch count, hit ratio impact, referral quality.
- Service: first response time, resolution time, CSAT/NPS and escalation rate.
- Quality and risk: override rate, exception accuracy, compliance findings.
Pragmatic 90-day path
- Weeks 1-2: Pick one narrow, high-volume use case with clear guardrails and a single system of record.
- Weeks 3-4: Stand up a prototype with human review. Define success criteria and red lines.
- Weeks 5-8: Pilot with 10-20 users on real volume. Instrument logs, measure baseline vs. pilot.
- Weeks 9-12: Automate approvals for low-risk cases, expand to 20-30% volume, finalize runbooks and controls.
Bottom line
Agentic AI is moving from experiments to everyday work. The firms that win will pair agents with domain rules, strong supervision and the plumbing to run end to end.
Upskill your team
If you are building internal capability for agentic workflows, see AI learning paths by role at Complete AI Training.
