Insurance AI programmes rarely fail on model choice. They fail because the organisation cannot retrieve what it knows, cannot prove how a customer-impacting decision was reached, and cannot show it learned from the outcome. We assess exactly that — and map it to what IRDAI already requires.
Most insurers and brokers already own the systems: portals, rules engines, policy feeds, agent networks, CRM, claims. AI nativity is not achieved by placing a chatbot over them.
It requires institutional knowledge that can be queried, verified signals from the ground, explicit decision rights, redesigned workflows, and a loop that learns from what actually happened. Each of those is an operating-model question before it is a technology question.
Why a proposal was declined, what the branch knew about a locality, which objection killed a renewal. The knowledge existed in a conversation and left with the person.
Agent-entered facts, third-party enrichment, last year's circular. The record exists and is treated as true, though nobody can say when it was checked or by whom.
The document is filed, the audit trail is complete, the evidence pack is somewhere. The underwriter making the decision still cannot query it under time pressure.
If any of the three cannot be answered with evidence today, the AI roadmap is being built on assumption.
Can an authorised user find the current approved exclusion wording, its effective date and its owner — without asking a person? Queryability is the first test of institutional intelligence.
For a customer-impacting recommendation: which inputs, which rule or model version, which reviewer, which override, and on what evidence. Reconstruction after the fact is the test.
Not from escalations and complaints alone — from verified field corrections, overrides, claims outcomes and drift, feeding back into rules, training and guidance.
Each dimension is scored 0–4 against defined maturity anchors. Weights are an insurance-specific v1 configuration, set before the assessment begins — never adjusted once the score is known.
| Dimension | Weight | What we look for in an insurer or broker |
|---|---|---|
| Strategy & leadershipS | 8% | An approved strategy tied to underwriting quality, policy suitability, customer outcomes and agent productivity — not to tool count. A funded owner for verified ground intelligence, and a named list of use cases deferred until evidence exists. |
| Workflow & decision redesignW | 10% | Needs discovery, suitability, carrier comparison, underwriting review, pricing exceptions and claims triage — each with documented decision points, owners and an exception register. Recontact and rework measured from case logs, not estimated. |
| Knowledge operationsK | 16% | Product rules, exclusions, circulars, SOPs and field learnings with a named owner, effective date, expiry and version history. Frontline corrections that actually change operating guidance. |
| Data quality & verificationD | 16% | Customer attributes traceable to method, time, field-agent identity and reviewer. Customer assertions, agent-entered facts, third-party data and inferences kept apart rather than merged into one confident record. |
| People, agents & capabilityP | 12% | Identity and credentials verified before production access. Competence tested rather than attendance recorded. Inactivity read with context — leave, seasonality, territory, tool failure — before any action. Misconduct handled through documented due process. |
| Governance & decision rightsG | 10% | Separate approval authorities for agent recommendations, underwriting, pricing and fraud investigation. Every material recommendation records inputs, evidence versions, rule or model version, reviewer and override. |
| Technology & integrationT | 8% | A register of AI, ML, rules engines, vendor models and chatbots by business owner, data used, output, criticality and supplier. Governed APIs, tested least-privilege access and tenant isolation, and connector failures that are visible rather than silent. |
| Trust, risk & customer protectionR | 10% | Disclosure when AI materially assists a decision, a working route to human review, screening for unfair proxies, minimisation of sensitive health and financial data, and complaint outcomes read systematically rather than case by case. |
| Value economicsE | 4% | A measured baseline — onboarding cost, training cost, rework, conversion, leakage, review time — with the cost of inaccurate information and mis-selling visible as its own line. |
| Learning & feedback loopsL | 6% | Corrections become owned tickets. Verified field observations enter a controlled refresh cycle. Outcomes update rules and training after bias checks. AI outputs evaluated on unseen, realistic cases. |
The dimensions are shared. The questions, the evidence and the remediation are not. The assessment branches before the first question is asked.
| Insurance broker | Insurer / carrier | |
|---|---|---|
| Core workflow | Recruit agents → train and refresh → discover customer needs → compare carrier APIs → advise and facilitate selection | Receive proposal → apply underwriting rules and review → price → accept risk or refer → service the policy |
| Where nativity breaks | Agent knowledge decays faster than training cycles. Recommendations cannot be explained against exclusions, affordability and carrier responses. Ground observations arrive as unstructured anecdotes. | Underwriters cannot see sources, discrepancies or confidence. Rule and model versions cannot be reconstructed after the fact. Exception logic lives inside two or three experienced people. |
| What the assessment adds | Agent credential and competency coverage, needs-discovery consistency, recommendation traceability, activity fairness, mis-selling complaint patterns. | Underwriting and pricing traceability, evidence retrieval at decision time, exception registers, model and rule change control, outcome-to-rule feedback. |
| First remediation pack | Agent training and continuity, grounded product comparison, structured ground-feedback capture. | Decision-support evidence layer, audit trail for overrides, governed knowledge operations. |
Twenty of the eighty questions are tagged to IRDAI control expectations and embedded inside the ten dimensions. They use the same evidence pipeline and the same maturity anchors. No second questionnaire, no duplicated evidence requests.
A single assessment therefore produces two outputs that are reported separately and never averaged together: an AI Nativity Index and a Regulatory Preparedness Report.
RG-13 — For underwriting and pricing recommendations, are model and rule versions, input evidence, explanations, reviewer decisions and overrides traceable?
RG-09 — Is public and third-party ground intelligence separated from personal policyholder records, with licensing, reliability, freshness and underwriting-relevance checks?
RG-07 — Can the origin, permitted purpose, date, field-agent identity, verification method and update history of each important customer attribute be established?
RG-10 — Can the business verify that agents and distribution personnel hold applicable permissions, required training, current authorisations and product-specific competence? RG-12 — Are inactivity and follow-up gaps measured fairly, with documented coaching and escalation?
RG-01 entity applicability and licensing · RG-02 maintained obligation register · RG-16 performance, failure and false-positive measurement · RG-17 adversarial testing and AI incident handling · RG-18 decision reconstruction · RG-20 named owner for regulatory change.
RG-03 inventory of AI, ML, rules and vendor tools · RG-04 executive ownership of information and cybersecurity obligations · RG-05 least-privilege access and anomaly monitoring · RG-06 vendor contracts, access boundaries and incident processes.
RG-08 lawful basis, retention and deletion controls · RG-11 detection of anomalies without presuming misconduct · RG-14 fairness of customer attributes and area-level features · RG-15 contest, correction, complaint and documented human review.
RG-19 — When ground agents contribute risk or customer knowledge, is it independently verified, scored for reliability, challenged where contradictory, and deprecated when wrong?
The control register is maintained against IRDAI instruments including the Information and Cyber Security Guidelines 2026; the Protection of Policyholders' Interests, Operations and Allied Matters of Insurers Regulations 2024; the Corporate Governance for Insurers Regulations 2024; and the Maintenance of Information by Regulated Entities and Sharing of Information by the Authority Regulations 2025, together with applicable intermediary conduct, registration, training and data-protection requirements. IRDAI's 2026 working-group mandate on AI governance is treated as a preparedness signal, not a mandate to achieve a score. Instruments, citations and applicability are verified against the official text and with insurance counsel before any engagement is reported on.
Every number in the report is traceable to a question response and an accepted piece of evidence. Nothing is inferred into a score.
The sequence does. Each phase is independently useful, and each one produces the evidence the next phase needs.
| Phase | Product | Service wrapped around it | Hard boundary |
|---|---|---|---|
| Phase 0Now | Readiness assessment platform | Leadership interviews, evidence validation, calibration, management readout | No production underwriting automation |
| Phase 1Next | Queryable insurance knowledge bank | Taxonomy design, data rights, knowledge ownership, sourcing and retrieval testing | Public and licensed material kept distinct from customer records |
| Phase 2Next | Agent feedback and broker enablement | Training redesign, feedback operations, supervised pilot | Existing quote and issuance APIs retained |
| Phase 3Later | Underwriting decision support | Workflow redesign, risk controls, model evaluation, reviewer workbench | Recommendation only — no unreviewed price, decline or denial |
| Long termVision | Integrated AI-native insurance operating system | Enterprise operating-model transformation | Not offered as an initial engagement |
Insurance budgeting runs on a hard boundary. What has not been assessed and costed before the year closes is rarely funded in the year that follows — and an implementation programme without a dated baseline cannot demonstrate that it worked.
One assessment cycle fits inside a quarter: scope, collect, verify, score, publish. Run it before the year closes, and the 90-day plan begins on day one of the next — not after a three-month discovery.
Structured records live in the platform. Interpretation, sampling and calibration are done by people — the platform carries the evidence, it does not invent the judgement.
Entity type and licence, lines of business, channels, regions, assessment period, respondent groups and the review owner. Sub-units — branch, zone, product, channel — are configured here, not improvised later.
Systems inventory, workflow map, workforce and agent counts, customer data sources, and an honest statement of what cannot be accessed. A refused source is recorded as a limitation, never extrapolated over.
Eighty questions distributed to distribution, underwriting, compliance, IT, L&D, operations and branch owners. Broker and carrier prompts diverge; shared prompts remain shared.
Documents, structured exports, screenshots and interview notes, each with source, date, owner and sensitivity. Optional read-only connectors where a client prefers them to uploads.
A reviewer samples, accepts or rejects, and records the reason. Self-reported ratings are kept alongside validated ones, and the disagreement between them is published rather than smoothed away.
Deterministic scoring, evidence confidence and gate rules. Critical blockers are raised as explicit No-Go conditions for specific use cases — they are never averaged into a comfortable number.
Interventions prioritised on impact, urgency and risk, dependency benefit, and feasibility. Legal and security blockers override the arithmetic and become prerequisites.
The report is locked to a version, permissions and an approval log. A reassessment cadence is agreed before the engagement ends, so the next cycle compares like with like.
An assessment product earns trust by being clear about the edge of its remit. These limits are contractual, not aspirational.
We do not certify compliance or file with the regulator. Status of applicable obligations is assessed separately by a qualified reviewer, and missing evidence is never treated as proof of a breach.
No automated underwriting, pricing, rejection, claim or fraud decision is built or operated in this programme. Assessment first; decision support only after gates are cleared.
Irregular activity, rework and documentation anomalies are investigative signals requiring context and human review. The platform produces no individual finding of misconduct.
Pseudonymised or synthetic samples by default. Purpose limitation, retention schedules, access logs and contractual controls apply to any real customer or agent data.
Benchmarks require matched product, territory and customer cohorts with sufficient sample sizes and explicit consent. We do not publish comparative insurer scores.
Number of AI tools deployed is not a dimension and does not raise a score. Capability is judged by queryability, evidence, decision rights and closed loops.
The snapshot is a 30-question self-assessment that returns an indicative score, your weakest dimensions and the blockers that would stop a use case today. It takes about fifteen minutes and it is not the full assessment — it is a reason to have the conversation.
Take the readiness snapshot Book a 30-minute diagnostic