So you're about to buy claims AI. Read this first.

A first notice of loss lands at 4:58 on a Friday. In a lot of claims operations it waits in a shared inbox over the weekend, gets keyed into the claims system on Monday, sits for triage on Tuesday, and reaches an adjuster who is already carrying a hundred and fifty open files. None of that is the hard part of a claim. All of it is where the days go.

J.D. Power put the average cycle time for a repairable auto claim at 19.3 days in its 2025 study, down from 22.3 the year before. The distance between a good claims operation and an average one is rarely the adjuster's skill. It is how fast the file moves before and after a human touches it.

Bring us one claims file and we will show you where the days are hiding

Insurers know this, and they are buying. In Conning's 2025 survey, 90 percent of insurers are somewhere in generative AI evaluation and 55 percent report early or full adoption. What happens next is the problem. Bain, surveying 160 insurers, found 78 percent of property and casualty carriers have adopted generative AI, but only 4 percent have scaled it across claims, and just 27 percent are pursuing a comprehensive claims transformation. Adoption is easy. Getting the order right is the part almost nobody does.

What a claims operation should automate first

Automate the parts of a claim that repeat and follow rules, in the order the file moves. First, intake: read the first notice of loss, in whatever form it arrives, and turn it into structured data. Second, document extraction: pull the policy number, loss date, coverage, and amounts off the PDFs and photos so no one retypes them. Third, triage and segmentation: score each claim for complexity, then route the simple, well-understood ones onto a fast track and send the rest to the right adjuster. Fourth, decision support: put coverage, claim history, and comparable losses in front of the adjuster before they open the file. Do not automate the decision itself. Those first three steps are where the calendar days hide. They are low-judgment, high-volume, and they carry the least regulatory risk. Start there, measure the cycle time you win back, then earn the next step.

The order we would run it

When we map a claims week, we do not start with a model. We start with the file's journey, from the first notice of loss to the closed claim, and we time every hop. The automation follows the friction, it does not lead. The order we would run it:

1. Intake and first notice of loss. Turn every channel, phone, email, portal, and PDF, into one structured record the moment a claim arrives, so nothing waits for a person to key it in.
2. Document and photo extraction. Read policy numbers, loss dates, estimates, and coverage off the attachments. This is the single biggest source of retyping and re-keying errors in most operations.
3. Triage and segmentation. Score complexity and severity, fast-track the clean claims, and route the rest to the adjuster who should own them. A simple auto glass claim and a disputed water-damage claim should never sit in the same queue.
4. Adjuster decision support. Assemble coverage, history, and comparable losses so the adjuster starts informed instead of hunting. The person still decides.
5. Fraud and leakage flags. Surface the signals for a human to weigh. A flag is a prompt to look closer, not a verdict.

Notice what is not on the list: deciding the claim. That line is deliberate, and it is where the regulator is looking.

What to leave to a human, and why the examiner cares

There is a version of claims AI that looks like a win and is a trap: letting a model deny or adjudicate claims on its own. Every state runs an unfair claims practices act, and a denial a model cannot explain is a market-conduct complaint waiting to be filed. When the Illinois Department of Insurance, or any other regulator, asks why a claim was denied, "the system decided" is not an answer that survives the exam. Keep a person on coverage decisions, denials, and anything a policyholder could contest, and keep a written record of who decided and why. Explainability is not a nice-to-have here. It is the difference between an audit trail and a consent order.

Read Bain's 4 percent that way. Part of it is a technology gap. The rest is carriers being right to move carefully on the decision itself while they move fast on the plumbing around it. Caution about adjudication is correct. The error is letting that caution freeze the intake, extraction, and triage work, which carries none of that risk and all of the wasted days.

Illinois makes the point concrete. It is one of the country's insurance capitals: Zurich North America, CNA, Kemper, Old Republic, and Ryan Specialty all run out of Chicago or its suburbs, alongside the brokerages and third-party administrators that handle claims for everyone else. The claims desks in this city are not short on AI pilots. They are short on the sequencing that turns a pilot into fewer days per claim. It is the same mistake we see when banks reach for the model before the process.

How to buy it without the pilot graveyard

Two questions keep a claims AI purchase honest. First, which specific step does this shorten, and how will we measure the days before and after? If the vendor cannot name the step or the metric, it is a demo, not a system. Second, what does it hand to a human, and what does it decide on its own? The right answer to the second question is that it decides nothing a policyholder could dispute. An intake tool that files a clean record, an extraction tool that stops the retyping, a triage model that routes correctly: each one pays for itself in cycle time and never ends up in front of an examiner. Buy those first, prove them on one line of business, then move up the ladder. That sequencing is the work we do across finance and insurance operations, and it starts by mapping how a claim actually moves before any model is chosen.

How to know it actually worked

The reason most claims AI cannot prove its value is that nobody measured the file before the tool arrived. Before we automate a single step, we baseline the cycle time hop by hop: how many hours from first notice of loss to a structured record, how many from record to triage, how many from triage to the first adjuster touch, how many from touch to close. Most operations have never seen those numbers, and the first surprise is always where the time really sits. It is almost never the adjuster's decision. It is the hours a file spends waiting in a queue no one owns. That baseline is not a spreadsheet exercise. It is the honest before a claims leader can point to when finance asks what the tool actually bought, and the number the next automation has to beat.

Once the baseline exists, every automation earns its place against it. Intake automation should collapse the intake-to-record hop from days to minutes. Extraction should cut the re-keying corrections that send a file back a step. Triage should shrink the share of simple claims that ever reach a senior adjuster's desk. Report each one as a before and after on the same claims, on a single line of business, before rolling it wider. A cycle-time number that moves on a scoreboard everyone can see is what keeps a claims AI program funded past the pilot. A vague promise of efficiency is what parks it in the graveyard next to the last three.

The insurers who win the next few years will not be the ones with the most models. They will be the ones who put automation where the days actually go, kept adjusters on the calls that need judgment, and can show a regulator exactly who decided what. Start with the file's first mile. The decision can wait for a person.