AI Underwriting: How Insurers Use It and a 90-Day Plan

2026-10-29 · Tommaso Maria Ricci

AI underwriting is the most quoted and least understood use case in insurance. Accenture's survey of 434 US underwriters, run with The Institutes, found that up to 40% of an underwriter's time goes to non-core and administrative work, a drag the firm valued at between $85 billion and $160 billion over five years. That is the real prize. Not a robot that prices risk on its own, but a desk where the people who understand risk finally spend their day on risk.

This guide explains how AI is used in insurance underwriting today, where it pays back, where it breaks, what regulators in the US and Europe now expect, and how to run a first program in 90 days without betting the book. I write it from the operator's side. I have spent more than twenty years building and running companies, and the pattern I see in underwriting is the same one I see everywhere: the technology is ready long before the process, the data and the governance are.

If you lead underwriting, operations, or a transformation office at a carrier, an MGA, or a broker with delegated authority, this is the practical version.

What AI underwriting actually means in 2026

Strip away the vendor language and AI underwriting is four families of tools working on one decision: should we take this risk, on what terms, at what price.

Document and data intake. Models that read submissions, ACORD forms, loss runs, schedules of values, medical records, inspection reports, and broker emails, then turn them into structured data. This is where large language models changed the economics after 2023.
Enrichment. Systems that pull third-party data automatically: property characteristics, aerial imagery, business registries, vehicle data, prescription histories in life underwriting, cyber posture scans for commercial accounts.
Risk scoring and triage. Predictive models that estimate loss propensity, flag appetite fit, and route each submission to the right queue: straight through, light touch, or full review.
Decision support. Assistants that summarize the file, compare it with guidelines and similar past risks, draft referral notes, and suggest terms, while the underwriter decides.

The phrase "AI underwriting" covers all four, which is why conversations about it get confused. A carrier that has automated intake has done something real. It has not "automated underwriting". It has removed the part of underwriting nobody was hired to do.

How AI is used in insurance underwriting, line by line

The mix differs a lot by line of business:

Personal auto and home. Mostly straight-through processing for standard risks, with models on pricing, prefill, and fraud signals at quote. The human role shifts to exceptions and portfolio steering.
Life and health. Accelerated underwriting programs that replace fluids and paramedical exams for a share of applicants, using prescription data, medical records, and predictive scores. This is also where regulators look hardest.
Small commercial. The sweet spot for triage and enrichment: high volume, thin margins, lots of public data. Many carriers aim to quote a large share of small business submissions with little or no manual touch.
Mid-market and specialty. Here the value is speed to quote and quality of attention. AI reads a 300 page submission in minutes and tells the underwriter where the risk actually sits. The decision stays human.

If you also write credit-linked products, the logic overlaps with loan decisioning, which I cover separately in my guide to AI in lending use cases.

Why underwriting is where insurance AI earns its keep

Claims gets the headlines because the savings are visible. Underwriting is where the margin is decided. A loss avoided at the front door never has to be managed at the back.

There are three economic levers, and every serious business case uses at least two of them.

Lever one: capacity. If underwriters spend a large part of their week on rekeying, chasing documents, and building files, every hour you give back is an hour of judgment on the risks that matter. Accenture's breakdown is blunt: around 40% of time on administrative tasks, 30% on negotiation and sales support, and only about 30% on actual underwriting, according to the firm's analysis of why underwriters don't underwrite much. Even partial relief on the first bucket changes the capacity of a team without new hires.

Lever two: hit ratio and speed. In commercial lines, the broker sends the same submission to several markets. The carrier that answers first with a credible indication wins a disproportionate share of the business it actually wants. Triage that surfaces the best-fit submissions on day one is a growth tool, not a cost tool.

Lever three: loss ratio. Better data at the point of decision, prefilled property attributes, cleaner exposure values, earlier fraud signals, translates into fewer mispriced risks. This lever takes longest to prove because losses develop over years, so never build a first business case on it alone.

The paradox technology has created so far

The same Accenture research contains a number every transformation lead should keep on the wall: 64% of underwriters said technology had either increased their workload or made no difference. Carriers have bought systems for a decade. Many of them added screens rather than removing steps.

That is the warning for AI underwriting. If the model produces a score that the underwriter must then validate by doing the old work anyway, you have added cost. The design question is never "what can the model predict" but "which human step disappears when the model is right".

What the leaders get back

McKinsey's 2025 analysis of AI in insurance put a number on the gap: over five years, the carriers it classed as AI leaders delivered about 6.1 times the total shareholder return of the laggards. Correlation is not causation, and leaders tend to be better run in general. But the direction matches what I see in every sector: companies that rebuild a core process around AI pull away from those that bolt tools onto the old one.

AI underwriting use cases that pay back first

Not every use case deserves to go first. The ones below share three traits: the data exists, the decision is frequent, and the human step that disappears is obvious.

1. Submission intake and clearance

The first win in almost every commercial carrier. A model reads the broker email and attachments, extracts the insured, the address, the class of business, the limits requested, and the loss history, checks for duplicates and clearance conflicts, and creates the record.

What disappears: manual rekeying and the back-and-forth to find missing documents.
What to measure: time from receipt to first touch, percentage of fields extracted without correction, clearance errors.
Typical trap: treating extraction accuracy as the goal. The goal is the underwriter not having to open the attachment to check.

2. Triage and appetite matching

Once the data is structured, a scoring model ranks each submission by fit with appetite and by expected profitability. Low-fit risks get a fast, polite decline. High-fit risks jump the queue.

What disappears: the underwriter reading files the carrier was never going to write.
What to measure: quote rate on top-ranked submissions, hit ratio, decline turnaround.
Typical trap: training on past bound business only. The model then learns what the carrier used to like, not what it should like now.

3. Data prefill and enrichment

For property and small commercial, external data fills the application before a human sees it: year built, roof type, distance to hydrant, occupancy, business age, revenue band. For personal lines, prefill cuts the questions a customer has to answer.

What disappears: questions to the broker, and the errors in self-reported data.
What to measure: questions removed per application, data discrepancies found, quote abandonment.
Typical trap: trusting third-party data blindly. Every external source needs a documented accuracy check and a way for the applicant to correct it.

4. Underwriting assistants for complex risks

In mid-market and specialty, a generative AI assistant summarizes the submission, maps it against the underwriting guidelines, lists missing information, compares it with similar past risks, and drafts the referral note for the authority holder.

What disappears: the first two hours of every complex file.
What to measure: time to quote, referral quality, guideline exceptions caught.
Typical trap: letting the assistant state facts without citing the page they came from. Every claim in a summary must link back to the source document, or the underwriter will rightly redo the work.

5. Accelerated life underwriting

Predictive models use prescription histories, medical information bureau data, motor vehicle records, and application answers to offer a share of applicants a decision without exams.

What disappears: paramedical exams and weeks of waiting for a share of applicants.
What to measure: share of applicants accelerated, mortality slippage against fully underwritten cohorts, placement rate.
Typical trap: going live without a monitoring plan for slippage. You only see the cost of a bad model years later, in the claims.

6. Portfolio monitoring and renewal underwriting

AI is not only for new business. Models that watch the existing book, flagging accounts where exposure, location risk, or loss trends have shifted, let underwriters focus renewal attention where it changes the outcome.

What disappears: blanket renewal reviews where nothing changed.
What to measure: rate adequacy at renewal, non-renewal of deteriorating risks, retention of good ones.

A worked example: one small commercial submission

Abstract use cases hide the mechanics, so here is a single file, end to end, the way a well-designed AI underwriting flow handles it.

A broker emails a submission for a small restaurant group: three locations, general liability and property, five years of loss runs as a scanned PDF, a schedule of values in a spreadsheet, and a two-line note.

Minute zero. The intake model reads the email and the attachments, extracts named insured, addresses, class codes, limits, and the five years of losses. It checks clearance and finds no conflicts.
Minute two. Enrichment pulls building characteristics for the three addresses, business registry data, and public inspection records. One address shows a building age that does not match the schedule of values.
Minute four. The triage model scores the account as a strong appetite fit with an expected loss ratio below target, and routes it to the light-touch queue, with one flag: the address discrepancy.
Minute six. The assistant drafts a summary for the underwriter: risk profile, loss history trend, the discrepancy, and the two guideline points that apply to commercial kitchens, each with a link to the source page.
Minute twenty. The underwriter reads the summary, checks the flagged address, asks the broker one question, and issues an indication the same day.

Without the flow, the same file sits in a queue for days and then takes a couple of hours of reading. The decision is the same person's decision. What changed is the hours around it.

This is the part I always insist on: the model never issued the quote. It removed the waiting and the reading. That is where the money is in year one.

The risks that sink AI underwriting programs

Most failed programs did not fail on the model. They failed on one of the five problems below.

Proxy discrimination

Underwriting is, by definition, the business of treating risks differently. The legal line is between differences justified by risk and differences that track protected characteristics. A model fed with external data such as credit attributes, location, or purchasing behaviour can end up reproducing race, ethnicity, or disability through proxies, even if those fields are never used. Regulators now expect carriers to test for this before deployment and on a schedule after it.

Explainability at the point of decline

When a model contributes to a decline, a surcharge, or a coverage limitation, the applicant and the regulator will ask why. "The algorithm said so" is not an answer, and neither is "the vendor says it's proprietary". If you cannot produce specific reasons for an adverse decision, the model is not ready for production.

Data drift

Underwriting models are trained on history. When the world shifts, inflation in rebuild costs, new climate patterns, a new class of business, the model keeps answering confidently with yesterday's assumptions. Drift monitoring is not optional, and the monitoring must have an owner with the authority to pull the model.

Third-party data quality

External data is the fuel of modern underwriting and its weakest link. Records are outdated, addresses are mismatched, business data is stale. Every external source needs an accuracy assessment, a contract that allows audits, and a correction path for the customer.

Shadow AI on the underwriting desk

The newest risk is the simplest. Underwriters under pressure paste submissions into consumer chatbots to get a quick summary. That sends customer and broker data outside the carrier's control. The fix is not a ban. It is giving the desk an approved tool that is better than the unofficial one. I wrote a full playbook on this in my guide to shadow AI risk management.

What regulators expect from AI underwriting, in plain terms

The regulatory picture differs by jurisdiction, but the direction is the same everywhere: governance, testing, documentation, and a human who is accountable. Here is the map as it stands at the end of 2026.

United States: NAIC bulletin, New York, Colorado

The NAIC Model Bulletin. In December 2023 the National Association of Insurance Commissioners adopted a model bulletin on the use of AI systems by insurers. It expects each insurer to adopt and maintain a written AI program covering governance, risk management, internal controls, and oversight of third-party vendors. By March 2025, according to a Quarles review of state adoptions, 24 states had adopted it with little or no change. It does not create new law. It tells you what examiners will ask for.

New York DFS Circular Letter No. 7. Issued on July 11, 2024, it covers the use of AI systems and external consumer data in underwriting and pricing. The circular letter expects insurers to test for unfair discrimination before a system goes into production and regularly afterwards, to have board-level oversight, to stay responsible for vendor tools, and to give applicants specific reasons for adverse decisions. It states that an insurer may not rely on the proprietary nature of a vendor's algorithm to justify a vague explanation.

Colorado. Colorado's SB21-169 led to Regulation 10-1-1, which has applied governance and risk management requirements to life insurers using external consumer data since November 2023. An amendment effective October 15, 2025 extended those requirements to private passenger auto and health benefit plan insurers, with full compliance and a first report due by July 1, 2026, as summarized by Willkie Farr & Gallagher.

European Union: the AI Act and EIOPA

The AI Act. Annex III of the EU AI Act lists as high risk the AI systems intended for risk assessment and pricing in relation to natural persons in life and health insurance, as you can read in the text of Annex III. After the 2026 amendments, obligations for Annex III high-risk systems apply from December 2, 2027. They include risk management, data governance, technical documentation, logging, human oversight, accuracy, and robustness. Property and casualty underwriting is not on that list, but the general rules of the Act, and existing insurance law, still apply.

EIOPA. On August 6, 2025 the European insurance supervisor published an Opinion on AI governance and risk management. It covers the AI systems that are not high risk or prohibited under the AI Act and explains how existing rules, Solvency II and the Insurance Distribution Directive above all, apply to them: data governance, record keeping, fairness, cybersecurity, explainability, and human oversight, applied in a risk-based and proportionate way.

What this means in practice

Across all these frameworks, an examiner will want to see five things:

An inventory of every model used in underwriting and pricing, with an owner.
Documentation of data sources, training, validation, and known limitations.
Testing records for unfair discrimination, before deployment and on a schedule.
Human oversight with clear authority to override and a record of overrides.
Vendor controls: contracts with audit rights, and your own validation of vendor claims.

If you build these from the first pilot, compliance becomes a by-product. If you add them after go-live, you are rebuilding the program.

The economics of AI underwriting: building the business case

A credible business case for AI underwriting starts from the cost of today's process, not from the price of the software.

Step one: measure the current desk

Before any vendor meeting, collect four numbers for one line of business:

Submissions received per month, and the share that gets quoted.
Average underwriter hours per quoted submission, split into intake, research, analysis, and negotiation.
Average days from receipt to indication.
Hit ratio on quoted business.

Most carriers do not have the second number. Getting it, even with a two week time study, is the most valuable thing you can do before spending money.

Step two: model capacity, not headcount

A first program should be framed as capacity released, not jobs cut. If intake automation takes a meaningful share of hours off each submission, the team can quote more of what it already receives, answer faster, and spend more time on complex risks. That is a growth story, which is easier to fund and easier to land with underwriters.

Step three: add the speed effect carefully

Faster indications usually improve hit ratio, but the size of the effect depends on the line and the broker relationships. Model it with a conservative range and validate it in the pilot.

Step four: leave loss ratio as upside

Loss ratio improvement is real but slow to prove. Present it as upside to be measured over several underwriting years, never as the number that pays for year one.

Where the cost sits

In every underwriting AI program I have seen, the budget surprises are the same: data cleaning, integration with the policy administration system, and the time of senior underwriters to label data and validate outputs. Software licences are rarely the biggest line. If you want a structured way to frame returns before you commit, my guide to AI ROI for business walks through the method.

A lesson from pricing outside insurance

The closest parallel in my own work is revenue management. In one hotel I worked with, applying AI to pricing, demand forecasting, and operational efficiency took annual revenue from 9 to 10 million. The mechanism is the same one underwriting faces: match price to the real risk or demand signal faster than the competition, and stop leaving margin on the table through slow, manual decisions. The difference is that a hotel sees the result next season, while an insurer sees it as losses develop. That is why the first underwriting case must stand on capacity and speed.

Self-assessment: is your underwriting operation ready for AI?

Score each statement from 0 (not true) to 2 (fully true). Be honest. The point is to find the bottleneck, not to pass.

Data

We can extract the last three years of submissions, quotes, binds, and declines for one line of business, with outcomes.
Our loss data can be linked to the underwriting file at policy level.
We know the accuracy of every external data source we buy.

Process

We have measured underwriter time per submission in the last twelve months.
Our underwriting guidelines are written down, current, and machine readable.
Referral and authority rules are clear and consistently applied.

Technology

Our policy administration system can receive data and decisions through an API.
Submissions arrive in a single intake channel we control.

Governance

We have an inventory of models already in use in underwriting and pricing.
Someone owns model risk, with authority to stop a model.
We test models for unfair discrimination before and after deployment.

People

Senior underwriters are willing to spend time labelling data and reviewing outputs.

How to read the score:

0 to 8: Fix the foundations first. Start with time measurement and a data inventory. A pilot now will fail for reasons that have nothing to do with AI.
9 to 16: Ready for a narrow pilot, ideally intake and triage on one line, with governance built in from day one.
17 to 24: Ready to scale. The constraint is now change management and integration capacity, not technology.

For a broader readiness view across the whole company, not only underwriting, see my AI readiness assessment guide.

If your score lands in the middle band and you want an outside view on where to start, this is exactly the kind of decision I help leadership teams make before they commit budget: a short diagnostic, one line of business, one measurable target.

A 30/60/90 day roadmap for AI underwriting

Ninety days is enough to prove or disprove a first use case, if the scope is narrow. Here is the plan I would run.

Days 1 to 30: diagnose and decide

Pick one line of business with volume and a clear pain point. Small commercial or a high-volume personal line is usually the right first choice.
Run a two week time study on the underwriting desk. Count hours per submission by activity.
Pull the data: one to three years of submissions, quotes, binds, declines, with outcomes where available.
Write the success metric before choosing the technology. For example: cut time from receipt to first touch by half, with no increase in clearance errors.
Set up governance: model owner, risk sign-off, testing approach for unfair discrimination, data protection review.
Decide build, buy, or partner. For intake and enrichment, buying usually wins. For triage on your own appetite, partnering or building on your data often wins. My framework for this choice is in build vs buy AI software.

Days 31 to 60: pilot in shadow mode

Run the model alongside the underwriters, not instead of them. They work as usual, and you compare the model's extraction, triage, and flags with what humans did.
Measure accuracy where it matters: fields that change the decision, not every field.
Collect underwriter feedback weekly. If they do not trust the output, find out why. Usually it is a missing citation or a data source they know is unreliable.
Run the first bias tests on the triage model and document them.

Days 61 to 90: go live on a slice and measure

Switch on the model for a defined slice, for example one region or one class of business, with human review of every adverse outcome.
Track the success metric weekly against the baseline from the time study.
Write the scale decision memo at day 90: results, issues, cost to scale, governance status, and the next two use cases.

After 90 days you should have either a proven win or a clear explanation of why the first attempt fell short. Both are worth having. What you must not have is a pilot that runs for a year because nobody defined what success looks like.

How to choose AI underwriting vendors

The vendor market is crowded, and most demos look the same. Here is how to separate them.

Questions that reveal the truth

"Show me the output on our submissions, not your demo data." Insist on a proof of concept with a sample of your own files, including messy ones.
"Where does every extracted field come from?" The tool should cite the page and line for each value. If it cannot, underwriters will check everything by hand.
"How do you test for unfair discrimination, and will you give us the results?" Under the New York circular letter and the NAIC bulletin, you stay responsible for vendor models. A vendor who refuses to share testing is a liability.
"What happens when your model is wrong?" Ask about confidence thresholds, human routing, and how errors feed back into retraining.
"Can we audit you?" Contracts should include audit rights, data usage limits, model change notifications, and an exit plan.
"How does it integrate with our policy administration system?" An answer that starts with "export a CSV" means manual work will come back.

Red flags

Accuracy claims with no definition of what was measured.
"Fully automated underwriting" for complex commercial risks.
Training on your data with no clear limits on reuse for other clients.
No answer on explainability for adverse decisions.

The operating model: who owns AI underwriting

Technology projects in insurance fail when ownership is split between IT, underwriting, and actuarial with nobody accountable for the outcome. AI underwriting needs a clear operating model.

Business owner: the chief underwriting officer or line head. Owns the target, the guidelines, and the decision to scale.
Model owner: data science or actuarial. Owns performance, drift monitoring, and retraining.
Model risk and compliance. Independent validation, bias testing sign-off, regulatory documentation.
IT and architecture. Integration, security, data pipelines.
Underwriting champions. Two or three respected underwriters who test outputs, shape the workflow, and carry credibility with the rest of the desk.

The champions matter more than most leaders think. Underwriters trust other underwriters, not slides. In every transformation I have led, the adoption curve followed the people the team already respected.

How AI changes the underwriting team

The fear on most desks is simple: will this replace me? The honest answer is that it changes the job more than it removes it.

What shrinks

Rekeying and data entry.
Chasing missing documents.
Reading files that were never going to be written.
Building the first draft of referral notes and summaries.

What grows

Judgment on complex and unusual risks.
Broker relationships and negotiation.
Portfolio thinking: which segments to grow, which to exit.
Working with models: challenging outputs, spotting drift, explaining decisions.

What leaders must do

Retrain before you deploy, not after. Accenture's research showed the share of underwriters rating their technical training as superior dropped from 52% in 2013 to 34% in 2021. Adding AI to a team that already feels undertrained creates resistance. Give people the skills to supervise the model, and frame the program as giving them back the job they signed up for.

I have seen this in other sectors. At a medical centre I worked with, AI applied to scheduling and operations raised capacity by about 20% with the same staff. The staff did not resist because the tool took away the work they disliked, not the work they were trained for. Underwriting is the same: protect the judgment, automate the friction. My AI change management framework covers how to sequence that shift.

Common mistakes in AI underwriting programs

These are the patterns I see most often, and all of them are avoidable.

Starting with the hardest line. Specialty and large commercial look attractive because premiums are high, but data is thin and every risk is different. Start where volume is high and decisions repeat.
Measuring model accuracy instead of business outcomes. A model that is 95% accurate on fields nobody uses is worthless. Measure time saved and decisions improved.
Skipping the time study. Without a baseline, you cannot prove the win, and the program loses funding at the first budget review.
Treating governance as phase two. Under the NAIC bulletin, the New York letter, Colorado's rules, and the AI Act, governance is part of the product. Adding it later means rebuilding.
Ignoring underwriter trust. If the tool does not show its sources, underwriters will redo the work, and the business case disappears.
Buying a platform before proving a use case. Large platform contracts signed before a single use case is proven lock in cost without results.
Forgetting the broker. Faster internal processing means nothing if the broker portal still requires manual uploads. Map the flow from the broker's side too.

If you are looking at the wider insurance picture beyond underwriting, including distribution and operating model, my guide to AI for insurance companies covers the whole value chain, and the companion piece on AI claims processing goes deep on the back end.

New risks to underwrite: AI itself

There is a second side to AI underwriting that many carriers miss: AI is not only a tool for underwriters, it is a new exposure to underwrite.

Businesses deploying chatbots, agents, and automated decision systems face losses when those systems fail: wrong answers to customers, discriminatory decisions, errors in automated processes. Traditional technology errors and omissions policies were not designed with this in mind.

The market has started to respond. In May 2025, insurers at Lloyd's began offering cover developed by the start-up Armilla for losses caused by underperforming AI tools, backed by Chaucer and Axis Capital, as Armilla describes in its launch announcement. The trigger is performance: the AI must fall short of the level it was expected to reach.

For underwriters this creates a new problem. How do you assess an AI system's risk? The answer looks familiar: governance, testing records, monitoring, human oversight. The same controls regulators expect from carriers are becoming the underwriting criteria for their clients. A carrier that has built good AI governance internally is better placed to underwrite it externally.

A practical checklist before you start

Before you sign any contract or launch any pilot, make sure you can answer yes to these questions:

Have we chosen one line of business and one measurable target?
Do we have a baseline from a recent time study?
Do we know which data we have, where it sits, and how reliable it is?
Is there a named business owner and a named model owner?
Have compliance and data protection reviewed the use case?
Do we have a testing plan for unfair discrimination?
Will the tool cite sources for every extracted value?
Do underwriters have a voice in the design?
Is there an exit plan if the vendor or model fails?

If any answer is no, fix that first. It is cheaper than fixing it after go-live.

If you want help structuring a first AI underwriting program, from choosing the line of business to setting up governance that will satisfy examiners, a focused conversation at the start saves months of drift later. That is the work I do with leadership teams that want results inside a quarter, not a two year roadmap.

FAQ

How is AI used in insurance underwriting today?

AI is used in insurance underwriting mainly in four ways: reading submissions and documents into structured data, enriching applications with third-party data, scoring and triaging risks by appetite fit and expected profitability, and assisting underwriters with summaries and guideline checks on complex files. In personal lines and small commercial, standard risks can move straight through with little human touch. In mid-market and specialty, AI speeds up the work around the decision, while the underwriter still makes the call on terms and price.

Will AI replace insurance underwriters?

AI is unlikely to replace underwriters on complex risks, but it changes the job. Surveys of US underwriters have found that a large share of their time goes to administrative and non-core work, and that is the part AI removes first. What grows is judgment on unusual risks, broker negotiation, portfolio decisions, and supervising models. Teams that retrain early tend to gain capacity rather than lose roles. Straight-through processing does reduce manual review for standard personal lines risks.

How long does it take to implement AI underwriting?

A focused first use case, such as submission intake and triage on one line of business, can be piloted and measured in about 90 days: one month to measure the current process and prepare data, one month running the model in shadow mode alongside underwriters, and one month live on a defined slice. Scaling across lines and integrating fully with policy administration systems usually takes twelve to twenty-four months, depending on data quality and legacy systems.

Is AI underwriting legal under US and EU rules?

Yes, but it comes with obligations. In the US, roughly half the states have adopted the NAIC model bulletin requiring a written AI program, New York's Circular Letter No. 7 sets testing and transparency expectations for underwriting and pricing, and Colorado regulates life, private auto, and health insurers using external data. In the EU, AI for risk assessment and pricing in life and health insurance is high risk under the AI Act, with obligations from December 2, 2027.

What data do AI underwriting models need?

AI underwriting models need historical submissions, quotes, binds, and declines, ideally linked to loss outcomes at policy level, plus current underwriting guidelines. Most programs also use third-party data such as property characteristics, business registries, vehicle records, or prescription histories in life insurance. Every external source should have a documented accuracy check and a way for applicants to correct errors, because regulators hold the insurer responsible for the data it uses.

How do you prevent bias in AI underwriting?

Bias prevention starts with testing models for unfair discrimination before deployment and on a regular schedule afterwards, looking for disproportionate adverse effects on protected groups and for less discriminatory alternatives. Carriers should review external data sources for proxies of protected characteristics, document the business rationale for each variable, keep humans in the loop for adverse decisions, and require vendors to share their testing. Regulators in New York and Colorado now expect this documented.

What does AI underwriting cost?

Costs vary widely with scope, but software licences are rarely the largest line. The main costs are data preparation, integration with the policy administration system, governance and testing, and the time of senior underwriters to validate outputs. A narrow first use case on one line of business costs a fraction of a platform rollout, which is why starting small and proving value before scaling is the safer financial path.