Will AI Replace Underwriters? A Straight Answer for Insurance Professionals

Short answer: standard-risk underwriting is largely already automated and the projections show a real decline, while complex, novel and disputed risk remains firmly human. The US Bureau of Labor Statistics projects insurance underwriters to decline 4 percent between 2025 and 2035, a loss of about 4,800 positions from a 2025 base of 125,600, with 2025 median pay of $81,370. Against 3.5 percent growth projected across all employment, that is a genuinely contracting occupation. It is also a small one, and the contraction is concentrated in a specific type of work.

Automation Arrived Here Long Before Generative AI

Underwriting has been rules-driven for decades, which makes it unusual among the occupations in this debate: most of the automation has already happened.

Personal lines, meaning motor and household insurance, moved to automated straight-through processing years ago. An application is scored against a rating model, priced and bound without a person seeing it. That transition removed a large volume of underwriting work and it is essentially complete.

What generative AI adds is the ability to read unstructured material at scale. Submissions arriving as documents, surveys, engineering reports, financial statements and correspondence can now be summarised and extracted automatically, which extends automation into commercial lines where the input was previously too messy to process.

So the honest framing is that this occupation is in the late stages of a long automation process rather than at the beginning of a new one, which is why the projected decline is 4 percent rather than something dramatic.

What Underwriting Actually Is, Below the Rating Engine

Concentrate on two rows. Novel risk is the clearest case: a model trained on historical losses is worthless for a risk with no historical losses, and the interesting parts of the insurance market are consistently the ones without history. Cyber, climate-affected property, emerging liability exposures and new business models all require someone to price with judgement and to be wrong in a bounded way.

The second row is “deciding when the model is wrong”. A rating model produces a number for every submission, including ones outside the population it was built on. Recognising that this particular risk does not resemble the training data is an underwriting skill, and it is the one that prevents a portfolio from accumulating exactly the risks the model misprices.

The adverse selection problem nobody automates away

There is a structural reason underwriting cannot become fully automatic, and it is more fundamental than any capability question.

If your pricing is mechanical and known, the market will find its errors. Brokers place business where it is cheapest relative to true risk, so any systematic underpricing in your model attracts a disproportionate share of exactly those risks. This is adverse selection, and it is the central problem of the entire industry.

The defence is a human who reviews the mix, notices that a particular class or region has grown unusually fast, and asks why. That question is not answerable from within the model, because the model is the thing being exploited. It requires someone who understands the market as a competitive system rather than as a dataset.

This is why insurers with sophisticated automation still employ underwriters, and why portfolio management roles have grown while transactional roles have shrunk.

A risk that shows the boundary

Take a submission that arrives on a commercial desk: a mid-sized manufacturer wants property and business interruption cover. The site is a converted industrial building, they have added a new production line in the last year, and there is a loss from four years ago described briefly in the broker’s summary.

Automated extraction handles the mechanics well. It pulls the sums insured, the construction details, the occupancy classification and the loss record, and produces a technical price from the rating model in seconds. That price is a genuine improvement on a person spending an hour doing the same arithmetic.

Then the underwriting starts. The new production line changes the fire load in a way the occupancy code does not capture, and the sprinkler design predates it. The business interruption sum insured looks low relative to turnover, which usually means either an accounting definition problem or a client who does not understand the exposure. The prior loss is described in one sentence, and the important question, whether the root cause was corrected, is not in the file. And the broker has marketed this account widely, which raises the question of why the incumbent insurer is not simply renewing it.

Answering those four questions requires a conversation with the broker, a request for a survey, a judgement about whether the price adequately reflects an exposure the model cannot see, and a decision about whether to write it at all.

None of that is in the submission. It is inferred from the shape of the submission by someone who has seen several hundred like it, and it is the reason experienced commercial underwriters remain in demand while the transactional roles around them disappear.

What to Know Before You Draw Conclusions

Regulation constrains automated decisions. Insurance pricing and declinature are regulated activities in most jurisdictions, with requirements around explainability, fairness and appeal. That shapes what may be automated regardless of capability.

Model risk management is a growth area. Someone qualified has to validate rating models, monitor drift and document governance. This is underwriting knowledge applied to a new object.

Exposure measures are not employment forecasts. BLS published AI exposure categories with the 2025-35 projections and states plainly that exposure “does not imply job loss, productivity gains, automation probability, or wage effects.” Financial roles score high on task overlap, and this occupation’s projected decline of 4 percent is real but modest.

The pipeline problem is severe. Personal lines underwriting was how people learned the discipline. With that gone, the industry has a shrinking supply of people who can handle complex risk in ten years, which several insurers openly regard as their largest talent problem.

Specialty markets behave differently. Lloyd’s-style specialty, reinsurance and large commercial risk remain relationship and judgement businesses where automation supports rather than substitutes.

Where Underwriting Careers Are Growing

• Complex and large commercial risk. Structuring, negotiating and pricing placements that no model handles.
• Specialty and emerging lines. Cyber, climate, parametric and new liability exposures where history is thin.
• Portfolio and accumulation management. Watching the shape of the book rather than individual risks.
• Model governance and validation. Applying underwriting judgement to the models themselves.
• Reinsurance. Concentrated judgement about tail risk, with limited data and high consequences, and a market where relationships still determine access to business.

A Decision Framework for Underwriters

1. Personal lines or standard commercial. The most exposed position, and the exposure is largely already realised. Move toward specialty, complex commercial or model governance within the next two years.
2. Commercial underwriter with broker relationships. Reasonably secure. Your protection is the relationship and your growth is in complexity, so seek the risks colleagues avoid.
3. Specialty or reinsurance underwriter. Strong position. Your risk is analytical currency rather than relevance.
4. Entering the profession. The traditional training route is contracting. Target insurers with genuine specialty books, because that is where you will be taught to underwrite rather than to check machine output.

The test that applies across all four: how often does your work involve a risk the rating model cannot price? That frequency is your security, and increasing it is the whole strategy.

It is worth adding that the decline in this occupation is small in absolute terms, roughly 4,800 positions over a decade, which is a fraction of the normal turnover in a workforce of this size. For an individual underwriter that means the practical risk is not redundancy but gradual role change, with the transactional part of the day being taken away and replaced by referral volume and portfolio work. Preparing for that shift is a much more useful response than preparing for job loss.

Common mistakes right now

• Accepting a model price on a risk that does not resemble the model’s population.
• Treating automated extraction as complete without checking what the submission implied rather than stated.
• Assuming portfolio growth in a class is good news without asking why it is growing.
• Letting technical underwriting skill atrophy while managing exceptions.

Building the Fluency Model Governance Now Requires

Insurers are deploying models that price, triage and decline, and the people best placed to govern them are underwriters who understand both the risk and the method. That combination is scarce, and it is currently one of the better-paid moves available inside the industry.

It requires understanding how these systems generalise, why performance degrades as the population shifts, what monitoring actually needs to measure, and how to document a decision so a regulator accepts it. Learning that in a structured sequence gives you the concepts rather than one vendor’s dashboard, which matters because the tooling changes faster than the governance requirements. A certificate alongside your underwriting qualifications makes the capability visible when an insurer is choosing who leads validation. If you want a structured route in, explore Coursiv AI lessons and check current plan details on the official site.

FAQ

Your Next Step

Take the last month’s submissions and count how many you priced by accepting the model output and how many required you to depart from it and justify why. The second number is the part of your role the industry cannot automate and currently struggles to replace, and if it is small, the fastest fix is asking for the referrals and complex risks that other underwriters find inconvenient.