Intelligent automation for insurers, beyond the pilot.

Which insurance process should you automate first? Nearly every guide to intelligent automation in insurance answers the same way, with a list: claims intake, underwriting, fraud detection, policy administration, compliance. The list is accurate. It is also the reason so many insurers own a successful pilot and an unchanged operation.

A list tells you where automation is possible. It says nothing about which routine will still be running, unattended and correct, a year after the demo. That second question is the one the industry data keeps flagging, and it has a different answer for a 200-person broker or MGA than for a global carrier with an innovation lab.

So this article holds back the list until the evidence has had its say. It covers what intelligent automation means in an insurance operation, how far insurers have actually got, where the hours go, a way to pick the first process, and what the research still cannot tell you.

What is intelligent automation in insurance?

Intelligent automation in insurance is the combination of rule-based software that moves data between systems with AI that reads what rules cannot: emails, scanned forms, loss descriptions, medical notes. The rules do the repeatable steps. The AI handles the unstructured input and hands a structured result back to the rules.

The older half of that pair is RPA, software robots that click through screens the way a clerk would. Insurers adopted it early because so much of their work lives in policy admin and claims systems that are decades old and have no clean way in. The newer half is language models, which can read a first notice of loss written in a hurry and pull out the date, the location and the type of damage.

The difference matters in practice. A bot that copies a policy number from one screen to another breaks the day the screen changes. An AI step that reads a broker's email can misread it. Our comparison of AI agents and RPA covers why the first kind of failure is loud and the second is quiet, which is the more dangerous of the two in a regulated business.

How far have insurers actually got with AI?

Further than most industries on adoption, and not far at all on results. Insurers try AI early and widely, and very few have turned it into a changed operation.

BCG's 2025 study is titled "Insurance Leads in AI Adoption," and it means it. It also found that only 7% of insurers surveyed have brought their AI systems to scale, with about two-thirds still in the piloting stage. BCG attributes 70% of the problems in scaling to people, organization and process, and only the remainder to the technology.

Capgemini Research Institute's World Property and Casualty Insurance Report 2026, built on 344 executives, 809 employees and 1,113 policyholders, lands in the same place by a different route. It found 60% of insurers still in exploration or proof of concept and 42% tracking no AI metrics at all. Nearly half of employees with access to AI tools, 47%, said their workday was unchanged.

The two studies count different things. BCG asked whether AI systems had reached scale, while Capgemini graded overall AI capability and found 10% of insurers in its advanced group. The definitions differ and the conclusion does not: somewhere between seven and ten insurers in a hundred are getting the operational result the other ninety are paying to pilot.

Where the hours go in an insurance operation.

The hours sit in three places: underwriting preparation, claims handling and the service desk between policyholder, broker and carrier. In each, most of the cost goes into gathering, retyping and chasing the information a decision needs, before anyone makes it.

Underwriting: preparation before judgment

Accenture's underwriting research, summarized on its insurance blog in 2025, found that more than a third of an underwriter's time still goes to non-core activities such as data collection and administration. That third is the work between a submission arriving and an underwriter making a call: extracting exposure data from a broker's spreadsheet, checking it against prior years, requesting what is missing.

That work is mostly rules. The fields are known, the checks are known, and the request for missing data follows a template. The AI part is narrow: reading a submission that arrives as a PDF, an email body and an attachment named "final_v3".

Claims: the wait between steps

J.D. Power's 2026 U.S. Property Claims Satisfaction Study, published in March 2026, found customers waited an average of 40.7 days for final payment, down 3.4 days from the prior year, and that 38% reported their first notice of loss digitally. A repair took 29.6 days on average. A good part of the remaining time is documents and photos waiting in a queue for someone to open them.

Bain's 2024 analysis of generative AI in property and casualty claims projects that the technology could cut loss-adjusting expenses by 20% to 25% and leakage by 30% to 50%. Those are projections, not measured outcomes, and Bain presents them that way. They describe the size of the prize if the work gets done; nobody has yet published evidence of collecting it at that scale.

Service and distribution: the middle of the chain

For brokers, MGAs and travel insurers, a large share of the routine work sits between companies rather than inside one. Certificate requests, mid-term changes, renewal reminders, monthly reports to the carrier, claim status questions from policyholders who already called twice. None of it shows up in a carrier's AI strategy deck, and all of it is repetitive, rule-bound and done by hand in most mid-size firms.

This is also where the search results for the term go quiet. Most of the first page is written for carriers, and the 50 to 500 person intermediary, whose margins depend on exactly this work, barely appears.

Which insurance process should you automate first?

Score each candidate on four questions and start with the one that scores best on all four. The winner is usually a dull, high-volume routine, which is also the kind most likely to survive its first year.

• How many times a week does it run? A routine that runs 400 times a week pays back faster than a clever one that runs four times.
• Could a new hire do it from a one-page instruction? If the answer is yes, most of it can be code, and only the reading step needs AI.
• What happens when it is wrong? A misfiled document costs a search. A wrong payment amount costs a complaint, a regulator's interest or both. Start where a mistake is cheap to catch.
• How many applications does it cross? Each one is a point where the routine can break when that system changes, so two is easier to keep running than six.

Applied honestly, this rule tends to rank first notice of loss intake, submission data extraction and certificate or document requests above fraud detection and pricing. Fraud and pricing are higher value, and they are also where a wrong answer is expensive and the data science is specialist work. They make better second or third projects, once the first one has shown the operation what a running automation looks like.

Many operations cannot list candidates at all, and that is a common place to start. Uplift's Brainstormer was built for that situation. Give it a role, say a claims handler at an MGA, and it proposes the routines in that job most worth handing over, based on a library of over 14,000 real automations.

Why do insurance automations stall after the pilot?

They stall because a pilot proves a routine can work once, on prepared data, with its builder watching. Production needs it to work every day, on real data, after the builder has moved on. BCG's 70% figure says the same thing in a consultant's vocabulary: most of what blocks scale is people and process, with the model a smaller share.

Three things go wrong most often. First, nobody owns the routine after launch, so when the claims system adds a field, it fails until someone notices. Second, the pilot chained several AI steps together, and the chain is weaker than any link: ten steps that are each 95% reliable succeed together only 59.9% of the time (0.95 to the power of 10). We work through that arithmetic in our piece on how agentic AI differs from traditional automation. Third, nobody measured anything, so there is no evidence to justify the next project, which fits Capgemini's 42% of insurers tracking no AI metrics.

Gartner expects the pattern to continue. Its June 2025 forecast is that over 40% of agentic AI projects will be canceled by the end of 2027, citing costs, unclear value and weak risk controls. In insurance, the risk controls carry extra weight, because a routine that touches claims or underwriting decisions has to show a regulator why it did what it did.

The reliable fix is to keep AI out of the steps that can be rules. Uplift builds to a fixed split, 95% tested code and 5% AI. On a claims intake routine, a model turns the policyholder's free-text description into fields, and from there code checks the policy number, the coverage dates and the amount. A wrong result then has one place to come from and one place to be caught, and a person approves anything that pays money or tells a customer no.

Should an insurer build it, hire it or buy the running result?

There are three ways to get intelligent automation into an insurance operation, and each one buys something different. The right one depends on whether you want to own an asset or a result.

Build it with a platform

Automation suites and no-code tools give your team the means to build. This is buying everyone a fishing rod, and in most operations the rods stay in the closet, because the underwriters and claims handlers who know the work have no time to also maintain software. It suits insurers with a real automation team who will own every routine for its whole life.

Hire a consultancy or development shop

Outside experts build well while they are engaged. When they leave, the knowledge of why each step exists tends to leave with them, and the first system change after handover lands on whoever is left. What you get in return is real: a codebase you own. If your company wants an asset on the balance sheet at the end, this is the better purchase, and no managed service will match it on that point.

Buy the running result

Uplift builds the routine from a plain-language description, runs it and keeps it working as the claims system, the carrier portal and the email format change. The analogy is the best rod on the market, kept in working order, together with someone who knows how to fish. Pricing is flat and the whole organization gets access, with no seats and no tokens, so there is no meter running and no surprise bill at the end of the month.

That last model has a cost, and it should be said in the same breath. When the plan ends, the routines stop running, and there is no source code sitting on your servers to restart them. A buyer considering any managed service, ours included, should ask the supplier directly what happens to the routines, the logic and the data if the contract ends, and get the answer in writing. Our breakdown of what intelligent automation services really sell lists the other questions worth asking before signing.

What the evidence cannot settle yet.

The research is clear that most insurers are stuck between pilot and scale. It is much less clear about what the few who got through are actually earning from it, and anyone selling you a number should be pressed on where it came from.

The BCG and Capgemini figures describe how far insurers have got, not what they gained. Bain's claims figures are projections. Capgemini reports that its advanced group sees higher revenue growth, but a correlation in a survey does not show that AI caused it; firms that are well run may simply adopt AI better and grow faster for the same reasons.

The workforce question is open too. The US Bureau of Labor Statistics projects that employment of claims adjusters, appraisers, examiners and investigators will decline 6% between 2025 and 2035, and says technology is expected to automate some of their tasks. Whether that means fewer people or the same people handling more complex claims is something no study can answer for your team.

What you can settle is narrower and more useful. Pick one routine with the four-question test, decide before it starts what number would count as success, and measure it for a quarter. One measured result in your own operation is worth more than any industry figure in this article, including ours.