🇪🇸 ES

Resources

Frequently Asked Questions

Straight answers organized by where you are in your AI journey.

I'm Skeptical

No. AI will not replace insurance professionals, but it will fundamentally reshape what they do and how they do it.

What AI does well in insurance: AI excels at processing large volumes of data, identifying patterns in claims history, automating routine document review, and generating first drafts of policy language. It can analyze thousands of claims in minutes, flag potential fraud indicators, and streamline underwriting workflows that previously took hours.

What AI cannot do: Insurance requires professional judgment that AI simply cannot replicate. Assessing the nuance of a complex commercial risk, building trust with a policyholder after a devastating loss, navigating ambiguous regulatory requirements, and making ethical decisions about coverage and claims — these remain deeply human capabilities.

The real shift is augmentation, not replacement. Accenture’s task-level estimate for insurance sales agents is that only 5% of their tasks are expected to be fully automated, 47% will remain unchanged, and the rest will be augmented. What that frees is time for higher-value work: client relationships, complex risk assessment, strategic advising, and creative problem-solving.

The insurance professionals most at risk are not those who will be replaced by AI. They are those who will be outperformed by professionals who use AI effectively. The role is evolving from data processor to strategic advisor, and AI is the catalyst making that transition possible.

The executives running the function read it the same way. In Accenture’s survey of 430 senior underwriting executives across 11 countries, 81% said AI and gen AI will create new roles; in Deloitte’s survey of 200 US insurance executives, 76% had already implemented gen AI in one or more business functions. The message is clear: learn to work with AI, and your value increases.

Sources

This is a fair question. The technology industry has produced genuine hype cycles before — blockchain in insurance promised transformation but delivered modest results. But the evidence suggests AI is fundamentally different, and here is why.

The investment is real and accelerating. On Gallagher Re’s tracking, cumulative global insurtech investment “crossed $60 billion in mid-2025, with roughly a quarter of that going to AI-focused firms” — and in the first quarter of 2026 AI-focused companies took a record 95.2% of all insurtech funding, $1.55 billion across 68 deals, up from 77.9% the quarter before. Unlike previous technology waves, the investment is coming from established carriers and reinsurers, not just venture capital. Lloyd’s, Swiss Re, Munich Re, and major US carriers are building AI capabilities internally — not experimenting, but deploying.

The outcomes are measurable in places — and unmeasured in others. In Accenture’s own documented health-claims implementation, work that took 11.5 minutes fell to three: “a 74 percent reduction in the claims settlement time”, with the machine learning “able to process health claims with 80 percent accuracy.” That is one client case, not an industry average, and we have found no published, verifiable industry-wide figure for AI’s effect on loss ratios. Read the case studies as existence proofs, not as benchmarks — and be wary of anyone who quotes you a range.

The technology has crossed a capability threshold. Previous AI waves struggled with natural language and unstructured data — the lifeblood of insurance. Large language models have solved that problem. AI can now read policy documents, interpret claims narratives, analyze medical records, and generate human-quality correspondence.

The regulatory environment confirms this is permanent. When the NAIC issues a Model Bulletin on AI governance and state departments of insurance create AI-specific compliance requirements, this signals institutional permanence, not passing trends.

The question is no longer whether AI will transform insurance. It is how quickly, and whether you will be prepared.

Sources

AI does not “understand” insurance the way a seasoned professional does — but it does not need to in order to be profoundly useful. The key is knowing what AI handles well and where human expertise remains essential.

What AI does remarkably well:

  • Document analysis: AI can process and summarize lengthy policy wordings, endorsements, and regulatory filings with high accuracy. It excels at identifying specific clauses, exclusions, and conditions across hundreds of pages.
  • Pattern recognition: AI detects patterns in claims data, loss histories, and risk factors that human analysts might miss. It can identify emerging trends across thousands of data points simultaneously.
  • Regulatory research: AI can rapidly scan regulatory updates across multiple jurisdictions and flag relevant changes for specific lines of business.
  • Routine drafting: First drafts of standard correspondence, policy summaries, and claims acknowledgment letters are well within AI capabilities.

Where AI falls short:

  • Professional judgment: Assessing whether a borderline claim should be covered requires contextual understanding that AI lacks.
  • Relationship nuance: Reading a client’s emotional state during a catastrophic loss, or understanding the unspoken concerns in a renewal negotiation, remains human territory.
  • Novel situations: Emerging risks — cyber, climate, pandemic — often lack the historical data AI needs to perform reliably.
  • Ethical reasoning: Decisions about fairness, equity, and social impact require moral reasoning AI cannot perform.

There is a real division of labour here, and the published task-level research draws it more modestly than the slogans do. Accenture’s estimate for insurance sales agents is that only 5% of their tasks are expected to be fully automated and 47% will remain unchanged, with the remainder augmented rather than removed. The professionals who will thrive are those who leverage AI for the routine and reserve their expertise for the complex.

Sources

This is one of the most important questions in AI adoption, and the honest answer is: it depends on how you use it and which tools you choose.

When cloud AI is risky: General-purpose AI tools like ChatGPT, Claude, and Gemini process data on external servers. Entering policyholder names, policy numbers, claims details, medical information, or Social Security numbers into these platforms creates real confidentiality and compliance risks. Most AI providers’ terms of service state that input data may be used for model training — a direct conflict with insurance data protection obligations.

When cloud AI is appropriate: These same tools are perfectly safe for non-confidential work: summarizing public regulatory documents, drafting template language, brainstorming coverage concepts, or analyzing anonymized, aggregated data. The key is ensuring no personally identifiable information (PII) or protected health information (PHI) enters the prompt.

Enterprise and local AI options exist: Enterprise versions of major AI platforms (ChatGPT Enterprise, Claude for Business) offer data processing agreements and commitments that input data will not be used for training. Local AI models running on your own hardware — such as Ollama with open-source models — keep all data on-premises, eliminating cloud exposure entirely.

A practical protocol:

  1. Classify your data before using any AI tool: public, internal, confidential, or restricted.
  2. Never input PII or PHI into consumer AI tools.
  3. Use enterprise or local AI for confidential insurance work.
  4. Establish a firm-wide AI data policy that all team members understand and follow.

The NAIC Model Bulletin explicitly addresses data privacy in AI use, requiring insurers to maintain appropriate governance over how AI systems handle consumer data. Compliance is not optional — it is a regulatory obligation.

Sources

I'm Curious

Start small, start safe, and start today. The most effective approach is to pick one low-risk task you already do regularly and try using AI for it.

Step 1: Choose a free AI tool. ChatGPT (OpenAI), Claude (Anthropic), or Gemini (Google) all have free tiers. You do not need to spend money to begin learning. Create an account with any of them — it takes two minutes.

Step 2: Pick a low-risk, non-confidential task. Good starting points for insurance professionals include:

  • Summarizing a publicly available regulatory update
  • Drafting a template email for policy renewal reminders
  • Simplifying complex policy language into plain English
  • Creating a checklist for a claims review process
  • Brainstorming questions for a risk assessment

Step 3: Write a clear prompt. Tell the AI your role, the task, and what format you want the output in. For example: “You are an insurance professional. Summarize the key changes in [specific regulation] in bullet points, focusing on implications for commercial property insurers.”

Step 4: Evaluate the output critically. AI will produce confident-sounding text that may contain errors. Check every fact, verify regulatory references, and never use AI output without review.

Step 5: Iterate and improve. If the first result is not what you need, refine your prompt. Add more context, specify constraints, or ask for a different format.

The Quick Wins section on Ariski provides ready-made exercises with tested prompts designed specifically for insurance professionals. Each takes minutes to complete and teaches you a practical AI skill without requiring any technical background.

Starting small is our own recommendation, not a research finding — we have found no published study that quantifies it, and we would rather say so than attach a number to it. What the research does show is that the constraint in insurance is organisational rather than individual: Accenture reports that 92% of workers want generative AI skills while only 4% of insurers are reskilling at the required scale. Starting small is how you get moving without waiting for a programme that may not arrive.

Sources

The best tool depends on your role, your budget, and what you need to accomplish. Here is a practical overview organized by category.

General-Purpose AI Assistants (Start Here):

  • ChatGPT (OpenAI): The most widely adopted AI assistant. Strong at drafting, summarizing, analysis, and brainstorming. Free tier available; Plus ($20/month) adds GPT-4 access.
  • Claude (Anthropic): Known for handling longer documents and more nuanced analysis. Excellent for policy review and detailed regulatory research. Free tier available; Pro at $20/month.
  • Gemini (Google): Integrated with Google Workspace. Strong at research tasks with access to current web information. Free tier available.

Insurance-Specific AI Platforms:

  • Shift Technology: AI-powered fraud detection and claims automation used by major carriers globally.
  • Tractable: Computer vision AI for auto and property claims — assesses damage from photos.
  • Unqork / EIS: No-code platforms for insurance that incorporate AI into underwriting and policy administration.
  • Zywave: AI-enhanced analytics for brokers, including market intelligence and client insights.

Document and Data Analysis:

  • Microsoft Copilot: AI integrated into Office 365 — useful for analyzing spreadsheets, drafting in Word, and automating workflows.
  • NotebookLM (Google): Excellent for analyzing and cross-referencing multiple insurance documents.

Key recommendations:

  • Begin with a general-purpose tool for internal, non-confidential work.
  • Graduate to insurance-specific platforms as your needs become clearer.
  • Always check data privacy policies before inputting any client or policyholder information.
  • Explore the Ariski Tool Directory for independent, unbiased evaluations of these and other tools.

Sources

  • AI Tools Landscape for Insurance Professionals 2024 — Celent (2024-09-01)
  • The InsurTech AI Tools You Should Know About — Coverager (2024-07-20)

The cost ranges from zero to millions, depending on your approach. The good news is that meaningful AI adoption can start for free and scale gradually based on proven value.

Free Tier (Individual Start): ChatGPT, Claude, and Gemini all offer free versions with significant capabilities. An individual insurance professional can begin using AI today at no cost, handling tasks like document summarization, draft correspondence, regulatory research, and brainstorming. This is the recommended starting point.

Professional Tier ($20-50/month per user): Premium versions of AI assistants (ChatGPT Plus, Claude Pro, Gemini Advanced) offer more powerful models, longer context windows, and priority access. Microsoft Copilot for business adds AI across Office 365 for approximately $30/month per user. For a small team of five, expect $100-250/month.

Specialized Tools ($500-5,000/month): Insurance-specific platforms like claims analysis tools, underwriting assistants, and compliance monitoring systems typically run $500-5,000/month depending on scale and features. These often require annual contracts.

Enterprise AI ($50,000-500,000+/year): Full enterprise deployments — custom models, API integrations, on-premises installations, and workflow automation — represent significant investment. Large carriers and reinsurers are committing budgets at this level.

ROI expectations: Be sceptical of return multiples, including the ones this page used to carry. We found no published, verifiable benchmark from McKinsey or any peer firm for a return multiple on AI in insurance, and the industry’s own measurement is thin: Capgemini’s World Property & Casualty Insurance Report 2026 finds that 55% of P&C insurers note the absence of clear ROI on AI initiatives and 42% track no AI metrics at all, with 72% of AI investment going to technology and infrastructure against only 28% to change management.

The documented gains that do exist are operational and case-specific — Accenture’s health-claims implementation reports a 74 percent reduction in claims settlement time. Measure your own baseline. Do not buy a multiple.

The best advice: start free, prove value, then invest. Document your time savings and quality improvements with free tools before committing budget. The business case will make itself.

No. You do not need coding, data science, or technical engineering skills to use AI effectively in insurance work. The most important skill is one you can learn in an afternoon: prompt engineering.

What is prompt engineering? It is the practice of writing clear, structured instructions (prompts) that guide AI to produce useful output. Think of it as learning to communicate effectively with a very capable but literal-minded research assistant. The better your instructions, the better the results.

The key skills for insurance professionals using AI:

  1. Clear communication: Describe your task, provide context, specify the output format, and set constraints. Example: “Summarize this policy exclusion clause in plain English suitable for a commercial policyholder. Limit to 150 words.”
  2. Critical evaluation: Assess AI output for accuracy, completeness, and relevance. This is where your insurance expertise becomes invaluable — you know what a correct answer looks like.
  3. Iterative refinement: Learn to adjust prompts when the output is not quite right. Add context, narrow the scope, or provide examples of what you want.
  4. Data awareness: Understand what information is safe to share with AI and what must remain confidential.

What you do NOT need:

  • Programming or coding skills
  • Understanding of machine learning algorithms
  • Technical setup or configuration expertise
  • A computer science background

The Institutes note that prompt engineering is rapidly becoming a core professional competency alongside traditional insurance skills. Ariski’s Prompt Engineering guide provides a comprehensive, insurance-specific learning path that requires zero technical background.

Sources

  • Skills for the AI-Enabled Insurance Professional — The Institutes (CPCU Society) (2024-03-01)
  • Prompt Engineering as a Professional Competency — Harvard Business Review (2024-05-15)
I'm Using It

Verification is non-negotiable. Every piece of AI-generated content in insurance must be reviewed before it is used, shared, or relied upon. Here is a practical verification checklist.

Regulatory Accuracy Check:

  • Cross-reference any regulatory citations against official sources (state DOI websites, NAIC publications, federal registers).
  • Verify statute numbers, effective dates, and jurisdictional applicability.
  • AI frequently cites regulations that do not exist or conflates requirements from different jurisdictions.

Policy Language Review:

  • Compare AI-drafted policy language against approved ISO forms, proprietary wordings, and your organization’s style guide.
  • Check for unintended coverage grants or exclusion gaps.
  • Ensure defined terms are used consistently and correctly.

Claims and Coverage Analysis:

  • Verify that coverage determinations reference the actual policy provisions at issue.
  • Confirm that reserves, damage estimates, or liability assessments align with established guidelines.
  • Check that AI has not overlooked relevant endorsements, amendments, or sublimits.

Source Verification:

  • If AI cites case law, industry reports, or statistics, verify each citation independently.
  • AI is known to fabricate convincing-sounding citations — a phenomenon called “hallucination.”
  • Use primary sources, not the AI’s summary, for any external-facing content.

Professional Judgment Layer:

  • Ask yourself: does this output reflect what a competent insurance professional would produce?
  • Consider edge cases, exceptions, and nuances that AI may have oversimplified.
  • Never let AI output override your professional expertise.

The NAIC model bulletin puts the responsibility on the insurer, not on the tool: insurers are expected to “develop, implement, and maintain a written program (an “AIS Program”) for the responsible use of AI Systems”, with that responsibility vested in “senior management accountable to the board or an appropriate committee of the board”. Verification is not just best practice — it is a regulatory and professional obligation.

Sources

The short answer: transparency builds trust, and in many contexts, disclosure may be legally required. Here is a framework for navigating this question.

When disclosure is mandatory or strongly advised:

  • Automated decision-making: If AI is used to make or significantly influence underwriting decisions, claims determinations, pricing, or coverage eligibility, most regulatory frameworks require disclosure. The NAIC Model Bulletin explicitly addresses this.
  • Adverse actions: If an AI-influenced decision results in a coverage denial, rate increase, or claims denial, the basis for that decision must be explainable and disclosed to the consumer.
  • State-specific requirements: Several states (Colorado, Connecticut, and others) have enacted or proposed AI-specific disclosure requirements for insurance. Check your jurisdiction.

When disclosure is recommended but not required:

  • Using AI to draft correspondence that you review and personalize.
  • Leveraging AI for internal research, analysis, or workflow efficiency.
  • Employing AI tools as part of your professional toolkit, similar to using policy management software.

Best practices for transparency:

  1. Be proactive, not reactive. If clients ask, never deny AI use. Dishonesty about tools erodes trust far more than the tools themselves.
  2. Frame it as quality enhancement. “We use AI-assisted tools to enhance our research and analysis capabilities, and every output is reviewed by our professional team.”
  3. Document your AI use. Maintain records of when and how AI was used in client-facing work.
  4. Follow your organization’s disclosure policy. If your carrier or agency lacks one, advocate for creating one.

J.D. Power’s AI Insurance Experience Study — 8,352 customer evaluations fielded in June and July 2026 — found that “nearly one-third (29%) of auto and home insurance customers are using artificial intelligence (AI) tools” in their insurance dealings, and that among the most common reasons customers give for not using AI is a lack of trust in the results. We have found no published figure for how much comfort disclosure itself buys. What the study supports is narrower and still worth acting on: trust is the binding variable, and disclosure is how you build it rather than spend it.

Sources

AI hallucinations — instances where AI generates confident, plausible-sounding but factually incorrect information — are one of the most significant risks in insurance AI use. Understanding why they happen and how to detect them is essential.

Why hallucinations happen: AI models generate text by predicting the most statistically likely next words based on training data. They do not “know” facts — they produce patterns. When the model lacks sufficient data on a specific topic, or when the prompt is ambiguous, it fills gaps with fabricated but convincing content. This is particularly dangerous in insurance because the output looks authoritative.

High-risk areas in insurance:

  • Regulatory citations: AI frequently invents statute numbers, regulation names, or compliance requirements that do not exist.
  • Case references: AI may cite nonexistent court decisions or attribute real holdings to wrong cases.
  • Statistical claims: Numbers, percentages, and industry benchmarks are often fabricated with false precision.
  • Policy interpretation: AI may describe coverage or exclusions that do not appear in the actual policy language.

Detection strategies:

  1. Verify every citation. If AI references a regulation, case, or statistic, check the primary source.
  2. Watch for excessive confidence. Hallucinations are often presented with the same confidence as accurate information.
  3. Cross-reference with known sources. Compare AI output against your professional knowledge and trusted references.
  4. Test with known answers. Ask AI questions you already know the answer to in order to calibrate its reliability in your domain.

Mitigation protocols:

  • Use AI for first drafts, never final products.
  • Implement a mandatory human review step before any AI output reaches clients, regulators, or the public.
  • Ask AI to cite its sources, then verify each one — this catches a significant percentage of hallucinations.
  • Prefer AI tools that provide source citations and confidence indicators.

Ernst & Young recommends treating AI output like junior staff work: it may be competent and efficient, but it requires senior review before it leaves the office.

Sources

  • Hallucination in Large Language Models: Causes, Detection, and Mitigation — Stanford Institute for Human-Centered AI (2024-03-01)
  • Managing AI Risk in Insurance Operations — Ernst & Young Insurance Advisory (2024-05-15)

Based on documented cases and industry research, here are the five most common and consequential mistakes insurance professionals make with AI.

1. Trusting AI output without verification. This is the single most dangerous mistake. AI generates confident text regardless of accuracy. Professionals who treat AI output as final product rather than first draft expose themselves to errors in coverage analysis, regulatory compliance, and client communications. Every AI output requires human review.

2. Inputting confidential data into consumer AI tools. Entering policyholder PII, claims details, medical records, or proprietary underwriting data into ChatGPT or similar consumer platforms violates data privacy obligations and potentially exposes your organization to regulatory penalties. Use enterprise or local AI solutions for confidential work.

3. Over-relying on AI for complex judgment calls. AI excels at routine tasks but struggles with novel, ambiguous, or ethically complex situations. Using AI to make coverage determinations on complex claims, assess bad faith exposure, or evaluate emerging risks without substantive human judgment is a recipe for errors with significant consequences.

4. Ignoring regulatory requirements. Many professionals adopt AI tools without understanding the regulatory landscape. The NAIC Model Bulletin, state DOI guidance, and emerging legislation create real compliance obligations. Ignorance is not a defense.

5. Failing to document AI use. When AI contributes to a coverage decision, claims determination, or underwriting assessment, there should be a record of what tool was used, what prompts were given, and how the output was reviewed. Without documentation, defending those decisions becomes significantly harder.

The common thread: All five mistakes stem from treating AI as a replacement for professional judgment rather than a tool that amplifies it. For detailed guidance on avoiding these pitfalls, explore Ariski’s What Not To Do section.

Sources

  • AI Adoption Pitfalls in Insurance: Lessons from Early Adopters — Boston Consulting Group (2024-07-01)
  • When AI Goes Wrong: Insurance Industry Case Studies — Best's Review (AM Best) (2024-09-01)
I'm Leading

An AI governance framework is the foundational document that guides responsible AI use across your organization. Without one, adoption is ad hoc, risk is unmanaged, and regulatory compliance is uncertain. Here are the essential components.

1. AI Usage Policy: Define what AI tools are approved for use, what tasks they may be used for, and what restrictions apply. Specify which tools are authorized at each data classification level (public, internal, confidential, restricted). This is your most critical governance document.

2. Approved Tool Registry: Maintain a vetted list of AI tools that meet your organization’s security, privacy, and compliance requirements. Include version information, data processing agreements, and renewal dates. Unapproved tools should be explicitly prohibited.

3. Data Classification Framework: Establish clear categories for data sensitivity and map each category to appropriate AI tools. Policyholder PII and PHI require enterprise-grade or local AI only. Public regulatory information can be processed with consumer tools.

4. Training Requirements: Define minimum AI competency standards for different roles. Underwriters, claims adjusters, compliance officers, and customer service staff will have different training needs. Include both initial certification and ongoing education requirements.

5. Audit Trail and Documentation: Require documentation of AI-assisted decisions, especially those affecting policyholders. Record what tool was used, what inputs were provided, what outputs were generated, and what human review occurred.

6. Review and Update Cycle: AI capabilities and regulations change rapidly. Commit to quarterly reviews of your governance framework and immediate updates when significant regulatory changes occur.

The NAIC Model Bulletin provides a baseline expectation that insurers maintain governance frameworks proportional to their AI use. Oliver Wyman recommends that governance be a living document — not a shelf document — with active enforcement and regular updates.

Sources

Effective AI training for insurance teams is not about making everyone a data scientist. It is about building practical competency at the right level for each role. Here is a structured approach.

Level 1 — AI Literacy (All Staff, 4-8 hours): Every team member needs foundational understanding: what AI can and cannot do, data privacy obligations, your organization’s AI usage policy, and basic prompt engineering. This is non-negotiable — uninformed users create the most risk.

Level 2 — Practical Application (Active Users, 8-16 hours): Staff who will use AI regularly need hands-on training with approved tools. Cover prompt engineering techniques, output verification workflows, common pitfalls, and role-specific use cases. Include supervised practice sessions, not just lectures.

Level 3 — Advanced Integration (Power Users, 16-40 hours): Team leaders and AI champions need deeper skills: workflow automation, custom prompt libraries, quality assurance frameworks, and the ability to evaluate new AI tools. These individuals become internal resources for their colleagues.

Level 4 — Strategic Leadership (Executives, 8-16 hours): Leaders need to understand AI’s strategic implications: competitive landscape, regulatory trajectory, investment decisions, vendor evaluation, and governance responsibilities. Focus on decision-making frameworks rather than technical details.

Recommended learning paths:

  • Start with Ariski’s AI 101 for foundational knowledge.
  • Use Quick Wins for hands-on practice in real insurance scenarios.
  • Complete the Prompt Engineering guide for communication skills with AI.
  • Periodic refreshers as tools and regulations evolve.

Measuring competence: assess in practice, not on paper. Can the team member use AI to complete a relevant task accurately and safely? That is the competency standard we recommend — our own position, not a cited finding.

We publish no hour count for this, because we have found no research that establishes one. Accenture’s insurance workforce research locates the problem elsewhere: 92% of workers want generative AI skills, and only 4% of insurers are reskilling at the required scale. The binding constraint is whether the programme exists at all — and after that, the quality of instruction and its relevance to the actual work, not the volume of content.

Sources

Measuring AI ROI in insurance requires tracking both quantitative efficiency gains and qualitative improvements. Here are the metrics that matter most, organized by function.

Time Efficiency Metrics:

  • Processing speed: Measure time-to-completion for AI-assisted tasks vs. manual baselines. Claims processing, underwriting review, and document analysis are the clearest comparisons.
  • Throughput increase: Track volume of policies reviewed, claims processed, or documents analyzed per period.
  • A caution on benchmarks: the time-savings percentages that circulate for claims and underwriting trace to no publication we could verify, and this page used to carry a pair of them. Use your own pre-AI baseline as the benchmark. Accenture’s documented health-claims implementation (a 74 percent reduction in settlement time) shows what one well-scoped deployment achieved — not what yours will.

Quality Metrics:

  • Error reduction: Compare error rates in AI-assisted work vs. pre-AI baselines. Track policy drafting errors, claims coding mistakes, and compliance oversights.
  • Consistency: Measure variation in outputs across similar cases. AI tends to produce more consistent first drafts than manual processes.
  • Rework rate: Track how often AI-assisted deliverables require significant revision.

Financial Metrics:

  • Cost per transaction: Calculate the all-in cost of processing a claim, underwriting a policy, or handling a customer inquiry with and without AI.
  • Loss ratio improvement: Track whether AI-assisted underwriting correlates with improved loss ratios over time.
  • Compliance cost: Measure reduction in regulatory penalties, remediation costs, and audit preparation time.

Client and Stakeholder Metrics:

  • Client satisfaction: Survey clients on response times, communication quality, and overall service experience.
  • Employee satisfaction: Track team engagement and satisfaction with AI-assisted workflows.
  • Retention metrics: Monitor whether AI tools correlate with improved policyholder retention rates.

Practical measurement approach: Start with a baseline before AI implementation. Track three to five key metrics for 90 days pre-AI and 90 days post-AI. Be honest about confounding variables. Capgemini’s World Property & Casualty Insurance Report 2026 shows why that discipline matters: 55% of P&C insurers report no clear ROI on AI initiatives and 42% track no AI metrics at all, while 72% of AI investment goes to technology and infrastructure against 28% to change management. The measurement gap, more than the technology, is what leaves the value unproven.

The regulatory landscape for AI in insurance is evolving rapidly. Here is what you need to know about current obligations and emerging requirements.

NAIC Model Bulletin (United States): The NAIC’s December 2023 Model Bulletin is the primary US guidance. Key requirements include:

  • Insurers must maintain governance frameworks for AI use.
  • AI-driven decisions must comply with existing unfair discrimination laws.
  • Outcomes-based testing is expected to ensure AI does not produce unfairly discriminatory results.
  • Insurers are responsible for third-party AI models and vendor tools.
  • Documentation and audit trails are required for AI-assisted decisions.

State Department of Insurance Guidance: Individual states are implementing their own requirements. Colorado’s SB 21-169 specifically addresses algorithmic discrimination in insurance. Connecticut, New York, and California have issued or proposed AI-specific guidance for insurers. Monitor your state DOI for jurisdiction-specific obligations.

EU AI Act (International Operations): The EU AI Act classifies insurance AI applications as “high-risk” when they influence underwriting, claims, or pricing decisions. Requirements include:

  • Mandatory risk assessments for high-risk AI systems.
  • Transparency obligations for AI-driven decisions affecting consumers.
  • Human oversight requirements.
  • Record-keeping and documentation mandates.
  • Administrative fines of up to 7% of total worldwide annual turnover for the prohibited practices of Article 5, up to 3% for breaches of most other operator obligations, and up to 1% for supplying incorrect or misleading information (Regulation (EU) 2024/1689, Article 99).

Key documentation requirements across frameworks:

  1. What AI tools are being used and for what purposes.
  2. How models were tested for bias and fairness.
  3. What human oversight processes are in place.
  4. How consumer complaints about AI-driven decisions are handled.
  5. Vendor management and third-party AI oversight protocols.

The practical imperative: Regulatory enforcement is accelerating. The cost of compliance is far less than the cost of penalties and remediation. Build governance now, not after an enforcement action.

Sources

I'm Deciding

The build vs. buy decision depends on your organization’s size, technical capacity, and specific needs. Here is a practical framework.

When to buy (most organizations):

  • You are a small to mid-size agency or carrier. Off-the-shelf AI tools and SaaS platforms are almost always more cost-effective than custom development. General-purpose tools like ChatGPT, Claude, and Microsoft Copilot handle the large majority of everyday tasks. (We used to put a percentage here. There is no study behind one.)
  • You need fast time-to-value. Commercial AI tools are ready today. Custom development takes 6-18 months minimum.
  • You lack in-house technical talent. Building AI requires data engineers, ML specialists, and ongoing maintenance. Most insurance organizations do not have and should not build these teams.

When to build (large carriers and reinsurers):

  • You have proprietary data advantages. If your organization has unique datasets that could create competitive advantage (decades of claims data, proprietary risk models), custom AI may unlock value that generic tools cannot.
  • You need deep integration. When AI must embed seamlessly into existing policy administration, claims management, or underwriting systems, custom integration may be necessary.
  • Regulatory requirements demand it. Some jurisdictions may require AI systems that are fully auditable and explainable — custom builds offer more transparency and control.

Vendor evaluation criteria: These are not only commercial questions. The NAIC model bulletin expects an insurer’s AI programme to address “[d]ue diligence and the methods employed by the Insurer to assess the third party and its data or AI Systems acquired from the third party to ensure that decisions made or supported from such AI Systems that could lead to Adverse Consumer Outcomes will meet the legal standards imposed on the Insurer itself”, and, where appropriate, contract terms that “[p]rovide audit rights and/or entitle the Insurer to receive audit reports by qualified auditing entities”. Buying does not transfer the responsibility.

  1. Data privacy: Where is data processed and stored? What are the contractual commitments?
  2. Insurance expertise: Does the vendor understand insurance-specific use cases and regulatory requirements?
  3. Integration capability: Can the tool connect with your existing technology stack?
  4. Total cost of ownership: Include licensing, training, integration, maintenance, and opportunity costs.
  5. Exit strategy: What happens to your data and workflows if you switch vendors?

The hybrid approach: start with commercial tools for general tasks and evaluate custom development only for high-value, proprietary use cases where off-the-shelf solutions demonstrably fall short. This minimizes risk while keeping options open. That is our own recommendation — this page previously attributed it to a vendor-research firm whose published version of it we could not retrieve, and we would rather own the advice than borrow an authority for it.

One cost that both paths share and most business cases omit: Capgemini finds P&C insurers putting 72% of AI investment into technology and infrastructure and only 28% into change management. Whatever you buy or build, the adoption work is the part that gets underfunded.

AI adoption creates new liability exposures that insurance organizations must understand and actively manage. Here are the primary risk categories.

Errors and Omissions (E&O) Exposure: If AI-generated advice, coverage analysis, or claims determinations prove incorrect and cause policyholder harm, the insurer or agency faces E&O liability. The standard of care has not changed — using AI does not diminish the professional’s obligation to deliver accurate, competent work. AI errors are your errors if you rely on them without verification.

Unfair Discrimination and Bias: AI models trained on historical data may perpetuate or amplify existing biases in underwriting, pricing, and claims handling. Regulators are increasingly scrutinizing AI-driven decisions for disparate impact on protected classes. The NAIC Model Bulletin explicitly requires insurers to test AI systems for unfairly discriminatory outcomes.

Regulatory Penalties: Non-compliance with emerging AI regulations carries real financial consequences. Fines for inadequate AI governance, failure to document AI-driven decisions, or using AI in ways that violate consumer protection laws are becoming more common. The EU AI Act’s ceilings run to 7% of total worldwide annual turnover for its prohibited practices and 3% for breaches of most other operator obligations (Regulation (EU) 2024/1689, Article 99).

Reputational Risk: Publicized AI failures — biased underwriting algorithms, incorrectly denied claims, data breaches from AI systems — can damage brand trust and market position. In insurance, trust is the product. Reputational damage from AI misuse can be harder to recover from than financial penalties.

Mitigation strategies:

  1. Implement robust governance before deploying AI in production.
  2. Test for bias regularly using diverse datasets and outcome analysis.
  3. Maintain human oversight for all AI-assisted decisions affecting policyholders.
  4. Document everything — the tool used, the input, the output, and the human review.
  5. Review your own E&O coverage to ensure AI-related exposures are addressed.

The regulators make the same point in their own words. The NAIC model bulletin warns that AI “can present unique risks to consumers, including the potential for inaccuracy, unfair discrimination, data vulnerability, and lack of transparency and explainability,” and that insurers “should take actions to minimize these risks.” EIOPA’s 2025 Opinion on AI governance and risk management reads existing European insurance legislation as already requiring proportionate governance of AI systems, rather than waiting for new rules. On both readings, the exposure comes from governing AI badly — not from using it.

Sources

AI adoption across the insurance industry is accelerating unevenly, creating both competitive risks and opportunities. Here is where the industry stands and where leading organizations are gaining advantage.

Industry adoption, from the published surveys:

We do not publish a single “share of insurers using AI” number: the surveys measure different populations and different things, and averaging them would invent a figure none of them reports.

Use cases by insurance sector:

  • Claims processing: Automated first notice of loss intake, damage assessment using computer vision, fraud detection pattern analysis, and reserve estimation. Accenture’s documented health-claims implementation reports a 74 percent reduction in claims settlement time — one client case, not an industry average.
  • Underwriting: AI-assisted risk assessment, automated data extraction from submissions, portfolio analysis, and pricing optimization. Some carriers use AI to triage submissions and prioritize human review.
  • Customer service: AI chatbots for policy inquiries, automated renewal processing, personalized communication, and 24/7 claims reporting.
  • Compliance: Regulatory change monitoring, automated compliance checking, and audit trail documentation.
  • Marketing and distribution: Personalized product recommendations, lead scoring, and agent performance analytics.

Competitive advantage patterns: BCG’s Build for the Future 2025 survey sorts companies into three tiers. Note that this segmentation is cross-industry, not insurance-specific (BCG, 30 September 2025):

  1. Future-built (5%): “at the forefront of AI innovation, systematically building cutting-edge AI capabilities across functions and consistently generating substantial value.”
  2. Scalers (35%): “scaling up their efforts and beginning to generate value, but many of them admit that they could be moving faster.”
  3. Laggards (60%): “report minimal revenue and cost gains and don’t yet have the proper capabilities for scaling AI in place.”

BCG reports that the future-built group expects “twice the revenue increase and 40% greater cost reductions than laggards in the areas where they apply AI.” The insurance-specific counterpart is Capgemini’s finding that only 10% of the industry is successfully scaling AI — on either reading, the top tier is small.

The competitive gap between tiers is widening. Organizations that defer AI adoption risk falling further behind as early adopters compound their efficiency and data advantages. However, rushing into AI without governance creates its own competitive risks through regulatory exposure and reputational damage.

The strategic imperative is not to adopt AI fastest, but to adopt it most responsibly and effectively.

AI is already changing insurance. The question is how deep and how fast the transformation goes. Here is a realistic timeline based on current trajectories and industry analysis.

Near-term (2025-2026) — Productivity transformation: This is where we are now. AI becomes a standard productivity tool for insurance professionals. Expect widespread adoption of AI assistants for drafting, research, analysis, and customer communication. Claims processing times drop significantly. Underwriting workflows accelerate. The professionals who resist AI begin to fall measurably behind their peers. Regulatory frameworks solidify, giving organizations clearer compliance guardrails.

Medium-term (2027-2029) — Workflow reimagination: AI moves from assisting individual tasks to reshaping entire workflows. Agentic AI — systems that can execute multi-step processes autonomously — begins handling routine end-to-end processes: standard claims adjudication, straightforward policy renewals, and basic underwriting for well-understood risks. Human professionals shift toward oversight, exception handling, and complex cases. Insurance roles evolve significantly, with new specializations emerging (AI governance, algorithmic auditing, human-AI workflow design).

Long-term (2030+) — Industry restructuring: AI enables fundamentally new insurance models. Parametric insurance expands dramatically as AI monitors trigger conditions in real-time. Predictive risk prevention supplements traditional indemnification. Autonomous underwriting handles standard commercial and personal lines with minimal human intervention. The industry’s employment structure shifts from processing-heavy to judgment-heavy, with fewer but higher-skilled professionals managing larger portfolios.

What this means for decisions today:

  • AI investment is not optional — it is a competitive necessity.
  • Governance frameworks built now will serve you through the entire transformation.
  • Team training should begin immediately; the skills gap widens over time.
  • Technology choices should prioritize flexibility and interoperability over any single vendor.

Be careful with the value figures that circulate. McKinsey’s published insurance estimate is smaller and differently shaped than the ones you will hear: generative AI could generate “between $50 billion and $70 billion in additional insurance industry revenue, with particular impact across marketing, customer operations and software engineering”. No 2030 horizon attaches to it, and it is revenue rather than “value”. We cite Reinsurance News’ report of the estimate because McKinsey’s own pages could not be retrieved during this review; the primary is owed and logged.

What is firmer is the industry’s own expectation of itself: the underwriting executives in Accenture’s survey put their use of AI technologies at 14% today and 70% within three years, while Capgemini finds only 10% of the industry successfully scaling AI right now. The organizations that capture the value are the ones making strategic decisions about AI today — not waiting for the future to arrive.

Still Have Questions?

The best way to answer questions about AI is to experience it firsthand. Try a Quick Win, explore our AI 101 learning path, or dive into real-world applications to build your own informed perspective.

Try a Quick Win Start AI 101

Ready for structured learning? Explore the Learning Program →

Comments

Loading comments...
Ctrl+K

No results found