Every consultancy has published the “AI will transform actuarial work” piece. McKinsey says it will reshape insurance. Deloitte says it will redefine risk assessment. PwC’s 2025 Global Actuarial Modernization Survey — covering 200+ insurers — found that 87% are actively modernizing their actuarial functions (up from 75% in 2023), with 94% citing efficiency as the primary driver. And yet, the same survey found that actuaries still spend more than half their time on data preparation, and automation maturity across actuarial tasks averages just 2.5 out of 5. The ambition is there. The execution is lagging.
But if you’re an actuary who has actually tried to use AI on real work — preparing an exhibit, reviewing a competitor’s filing, responding to a DOI objection — you know the gap between those predictions and your Tuesday afternoon. The models are impressive at generating plausible text. They are not impressive at getting a loss development factor right to four decimal places.
This post is about the other side of that gap: the specific actuarial tasks where AI delivers finished, verifiable work product today, and the tasks where it doesn’t. No frameworks. No roadmaps. Just what works.
The Hype vs. Reality Gap
The consulting-firm version of AI in insurance sounds like this: AI will automate pricing, transform underwriting, and enable real-time risk assessment. These statements are technically directional. They are also useless to an actuary trying to decide whether to invest time learning a new tool or just keep doing things the way that works.
The reality is more specific and more interesting. AI doesn’t “transform actuarial work” as a monolithic category. It transforms individual tasks within actuarial workflows at different rates, with wildly different reliability levels. Some tasks are essentially solved. Others are years away. Knowing which is which is the difference between saving hundreds of hours per quarter and wasting weeks chasing hallucinated output.
The pattern is consistent: AI excels at tasks that are labor-intensive but procedurally well-defined, where the inputs are structured (or structurable) and the outputs can be verified against source material. It struggles with tasks that require genuine judgment under uncertainty — the work that makes actuarial science a profession, not a process.
What Actually Works Today
Filing Research and Precedent Discovery
Before AI: An actuary preparing a rate filing in a new state spends days reading through competitors’ filings, DOI bulletins, and prior objection letters to understand what the regulator expects. For a prior-approval state like California, this means navigating CDI’s 71-page filing instructions, reviewing precedent filings for acceptable methodologies, and identifying which exhibits are required versus optional. The work is critical and almost entirely manual.
With AI today: AI agents search across thousands of indexed filings, surface relevant precedent filings by line of business and state, extract the specific methodologies used (trend selections, credibility procedures, territorial definitions), and compile a research brief with citations to source filings. What took two to three days of reading takes under an hour. The actuary reviews the output against the cited sources rather than starting from a blank page.
Why it works: The underlying task is search and synthesis across structured documents. The filing data exists in SERFF. The documents are PDFs with consistent formatting conventions. The output — a research summary with citations — is directly verifiable.
Rate Table Extraction and Validation
Before AI: A carrier implementing a competitor’s rates (or its own, from a newly approved filing) manually reads rate pages, transcribes tables into spreadsheets, cross-references against prior filings for amendments, and validates the complete table set. For a multi-state personal auto program, this means processing hundreds of territory factor tables, class plans, tier structures, and exception rules across thousands of pages. Carriers pay consulting firms $20,000–$30,000 per state for this work, and it takes weeks to months.
With AI today: AI agents read the filing documents, extract every rate table, factor, and rule, reconcile amendments against the base filing, and produce structured output — validated against the source PDF, page by page. The output includes a verification packet: every extracted value is linked to its source location in the original document. An actuary can audit the extraction in hours instead of performing it over weeks.
Why it works: Rate tables are structured data trapped in unstructured documents. The extraction task is well-defined (identify tables, parse values, map relationships). The verification step is mechanical (does extracted value X match the value on page Y of filing Z?). This is the kind of labor-intensive, high-precision work where AI agents outperform manual processes by orders of magnitude.
Competitive Rate Analysis Across Carriers and States
Before AI: Running a market basket analysis — comparing your rates against competitors for a representative set of risks — has traditionally required purchasing competitor rate manuals, manually building each competitor’s rating algorithm, and running thousands of sample quotes. For commercial lines, this was functionally impossible at scale because the rate structures were too complex and the manuals too long.
With AI today: Because AI can extract and implement rating algorithms from public filings (see above), competitive analysis becomes a computation problem instead of a data-entry problem. An actuary defines a set of sample risks. The system rates each risk through multiple carriers’ filed algorithms and produces factor-by-factor comparisons: where you’re higher, where you’re lower, and which rating variables drive the differences. Analyses that previously required a dedicated team and a six-figure budget run in hours.
Why it works: This is a downstream application of rate table extraction. Once rating algorithms are digitized and executable, comparing them is straightforward computation. The hard part was always getting from PDF to working algorithm — not the comparison itself.
Objection Response Research and Drafting
Before AI: When a DOI issues an objection letter, the actuary needs to understand the specific concern, research how other carriers have addressed similar objections, and draft a response that directly addresses the regulator’s questions with supporting data. The research phase — finding precedent responses across other filings — is where most time is spent.
With AI today: AI agents identify the specific objection categories, search for filings where similar objections were raised and resolved, surface the response strategies that worked (and the ones that didn’t), and draft a response framework citing relevant precedent. The actuary rewrites and refines based on the specifics of their filing, but starts with a researched foundation rather than a blank document.
Why it works: Objections follow patterns. State regulators raise similar concerns across filings in the same line of business. The research task — finding those patterns across thousands of filings — is exactly what AI does well. The drafting assistance is useful as a starting point, though it requires heavy actuarial editing to match the filing’s specific context.
Exhibit Generation from Raw Data
Before AI: Building filing exhibits — loss development triangles, trend analyses, rate indications — from raw loss and premium data involves structured actuarial calculations, formatting to DOI specifications, and cross-checking every number. It is time-consuming but formulaic.
With AI today: AI agents take raw data inputs, apply standard actuarial methodologies (selected development factors, trend assumptions, credibility weighting), and produce formatted exhibits. The actuary reviews the methodology selections and validates key outputs rather than building exhibits cell by cell.
Why it works (with caveats): The calculations are well-defined. Producing a loss development triangle from raw data is a deterministic process once the methodology is specified. The caveat is that selecting the methodology — choosing development factors, deciding on trend periods, applying credibility — still requires actuarial judgment. AI handles the computation and formatting. The actuary owns the decisions.
What Doesn’t Work Yet
Actuarial Judgment Calls
Credibility weighting, trend selection, development factor selection — these decisions require understanding not just what the data shows but what it means in context. Is a spike in frequency a one-time event or a trend? Should you weight industry data more heavily because your book is too small, or trust your own experience because the book is unique? AI can surface the data and present the options. It cannot make the judgment call, and it shouldn’t.
An AI system that confidently selects a 3.5% trend factor without understanding why the last two years deviate from the prior eight is not a tool. It is a liability.
Novel Risk Assessment Without Historical Data
Emerging risks — new product types, new coverage territories, climate-driven loss patterns — lack the historical data that makes AI reliable. An actuary pricing cyber coverage for a new industry segment or homeowners coverage in a region with a changing catastrophe profile is doing genuinely novel analytical work. AI can assist with analogous data and scenario modeling, but the core pricing judgment is human.
Regulatory Negotiation and Relationship Management
The back-and-forth with a state DOI — understanding an examiner’s concerns, knowing when to push back versus concede, managing the political dynamics of a rate increase in a sensitive market — is a relationship and judgment skill that AI does not replicate. AI can prepare you for the conversation. It cannot have it for you.
The Verification Question
The central question for any actuary evaluating AI tools is not “can it do the work?” but “can I verify what it did?” Actuarial work carries professional and regulatory accountability. An actuary who signs a filing is personally responsible for its accuracy. No AI system changes that.
The viable model is verification-first: every AI-generated output includes a complete audit trail linking each result to its source data. A rate table extraction includes page-level citations to the source filing. A competitive analysis includes the specific filed factors used in each calculation. An exhibit includes the raw data, methodology selections, and intermediate calculations.
This is not optional. It is the only way AI tools can integrate into workflows where professional accountability exists. Actuaries should evaluate any AI tool by asking: If a DOI examiner questions this number, can I trace it back to its source in under 60 seconds?
What a Verification Packet Looks Like
For a rate table extraction: every extracted value links to a specific page and location in the source PDF. The actuary can click through to the original document and confirm.
For a competitive analysis: every quoted rate links to the filed factor tables used to calculate it, with filing tracking numbers and effective dates.
For an exhibit: every calculation step is transparent — selected factors, applied weights, intermediate results — with the raw data available for independent recalculation.
What to Look for in AI Tools
If you are evaluating AI tools for actuarial work, here are the criteria that matter:
- Source citation on every output. If a tool produces a number without telling you where it came from, it is not ready for actuarial use.
- Domain-specific, not generic. A general-purpose chatbot that can “answer insurance questions” is not the same as a system that understands how to read a filed rate manual, extract an amendment chain, and produce an executable rating algorithm. The difference is years of domain engineering.
- Handles real document complexity. Insurance filings are not clean datasets. They are thousands of pages of PDFs with exceptions, amendments, cross-references, and state-specific variations. Evaluate tools on your hardest filings, not your simplest ones.
- Integrates into existing workflows. AI tools that require you to rebuild your workflow around them will fail. Tools that accelerate the workflow you already have will succeed.
- Honest about limitations. Any vendor that tells you AI replaces actuarial judgment is either lying or doesn’t understand the work. Look for tools that are explicit about what they automate and what they leave to you.
Where This Is Heading
In the next 12–18 months, three trends will accelerate:
Broader task coverage. AI will move deeper into the actuarial workflow — from research and extraction into reserving analysis, predictive model validation, and real-time competitive monitoring. Each new capability will follow the same pattern: automate the labor-intensive, verifiable components while leaving judgment calls to the actuary.
Higher verification standards. As AI-generated work product enters more filings, regulators will develop expectations around AI audit trails. Actuaries who adopt verification-first tools now will be ahead of this curve.
Compound knowledge systems. AI systems that learn from every filing they process, every objection pattern they identify, and every rate structure they extract will develop institutional knowledge that exceeds what any individual actuary carries. The actuary’s role shifts from doing the analytical labor to directing and verifying a system that knows more than any one person can.
The profession is not going away. The profession’s relationship to its tools is changing — the same way it changed when spreadsheets replaced columnar pads and statistical software replaced hand calculations. The actuaries who engage with what actually works today, rather than waiting for the hype to settle, will define how the profession uses these tools for the next decade.
Effective AI offers a platform of specialized AI agents that do the research, extraction, and exhibit assembly actuaries spend most of their time on — with a verification packet on every output so you can confirm correctness without retracing the entire analysis. See how it works →