How insurance companies can use AI to automate workflows

A denied insurance claim is one of the few business decisions that arrives with its own statute attached. Most carriers are piloting AI somewhere in the claims or underwriting pipeline right now, and the demos look sharp. What’s usually missing is the part that lets a model do real work in a regulated shop: a process that routes its output past a named adjuster before anything reaches the policyholder.

So here’s the answer before the detail. AI fits insurance as an extraction-and-triage layer that feeds a human who owns the call. It reads the submission, pulls the fields off the loss report, classifies the claim by severity, and drafts the correspondence a reviewer will send. What it doesn’t do is decide coverage, approve the payout, or sign a denial, because those are the actions a regulator or a court can pull a file on, and a reconstruction needs a person’s name attached.

That one line is the whole playbook.

The rest of this post is the map: which insurance workflows to hand a model first, and where the law has already told carriers to keep someone in the chair. It’s the evergreen version, the one that holds up after the regulatory news moves on. We’ve written the broader financial-services AI playbook and the general case for fixing the insurance workflow before you automate it, plus where AI is heading across regulated work. Here it’s carrier-specific.

Which insurance work is safe to hand a model?

The document-heavy, deadline-bound work, and none of the coverage decisions. A model laid over a broken claims process doesn’t fix it; it just produces inconsistent decisions faster and bills you for the tokens. Point that same model at the reading and drafting inside a process that logs what it did, and it does real work without ever touching a coverage call.

The split that matters in insurance isn’t claims versus underwriting. It’s the file versus the call. Working the file, reading the submission, extracting the loss details, classifying severity, drafting the letter, is where a model saves hours, and a wrong draft there costs a reviewer a minute. The call, whether to bind, what to pay, whether to deny, stays with a licensed person who can answer for it. Put the model on the file. Keep an adjuster or underwriter on the call. Get that boundary right and you’ve settled where AI is safe at your carrier, before anyone argues about model quality.

No model release moves where that line sits.

So write the off-limits list into the process, plainly, because a playbook that only sells AI isn’t one a state examiner will trust. Autonomous denial of a claim is out. So is binding a policy or releasing a payout with no human sign-off. So is any letter telling a policyholder no that no adjuster read first. The model can prepare every one of those; it can’t be the thing that issues them. A vendor pitching “AI that auto-adjudicates claims” is selling you a market-conduct finding dressed up as efficiency.

The claims leaders we hear from worry about exactly this shape of risk: a fast-but-unreviewed denial that a market-conduct exam later reads as a pattern rather than a one-off. The fix isn’t a more cautious model. It’s a step the run can’t skip.

The intake-heavy work to automate first

Start with the work your team already runs on a deadline: a defined entry, checks along the way, and a sign-off somebody owns. These five fit that shape.

New-business and underwriting intake. The model extracts the submission, runs the completeness and appetite checks, and routes a clean package to the underwriter while the broker’s email is still open. It flags the missing exposures and the inconsistencies. The underwriter prices and binds. The reading is the grind, and the model takes it off the desk.

Policy issuance. Once a risk is bound, the model assembles the issuance checklist, generates the policy documents from the rated inputs, and runs a QC pass against the quote before anything goes out. A person clears the exceptions and authorizes the issue.

FNOL and claims triage. At first notice of loss, the model captures the report, classifies the claim by severity and complexity, and routes it to the right queue. A minor glass claim and a suspicious total loss stop looking alike at intake, instead of three days into the file.

Claims handoff and escalation. The model classifies the documents in the file, drafts the status correspondence, and tracks the statutory clocks so nothing ages past a deadline. The coverage call, and any escalation to the SIU, stays human.

Renewal review. The model summarizes the exposure changes, drafts the renewal packet, and surfaces what moved since last term. An underwriter signs the renewal terms.

Two of those are already public templates you can clone: a claims-processing flow and a commercial-quote intake. Each ships with a defined entry and a sign-off step, so the AI work is narrow on purpose, drop a model into the reading parts and leave the human sign-off exactly where it sits.

The order is the point. The AI step sits before the adjuster’s review, so the model’s read never reaches the policyholder on its own, and the coverage decision always carries a name. The logging lives at the workflow level, so the trail outlives whatever model you’re running next year. None of it is exotic. It’s the ordinary discipline of an insurance file moving from intake to settlement, with a record of every step as the work happens and one reading step now handled by a model instead of a clerk.

Why an unreviewed denial is a regulatory problem

Because the denial is the one claims action with a statute waiting behind it. Most states run a version of the NAIC’s Unfair Claims Settlement Practices Act, and it’s specific about what a carrier can’t do. Among its defined unfair practices are “refusing to pay claims without conducting a reasonable investigation” and “failing in the case of claims denials or offers of compromise settlement to promptly provide a reasonable and accurate explanation of the basis for such actions.” A model that denies a claim and can’t explain the basis fails that second test on its face.

There’s a second set of teeth in the same act. A single slip usually isn’t the violation; the act bites when a practice is “committed with such frequency to indicate a general business practice.” That’s the exact risk profile of an automated denial step. One bad call is an error anyone can have. The same automated bad call repeated across a book of claims is a general business practice, which is precisely what a market-conduct exam is built to surface.

Sitting above the claims rules now is AI governance itself. The NAIC’s Model Bulletin on the Use of Artificial Intelligence Systems by Insurers, adopted in December 2023, reminds carriers that “decisions or actions made or supported by AI must comply with all applicable insurance laws and regulations” and sets out expectations for how insurers govern their AI use. In plain terms, pointing a model at a regulated decision doesn’t move the liability to the model. The carrier still owns the outcome, the explanation, and the trail behind both.

So what does an examiner actually want when a complaint lands on one denied claim? Not your model’s accuracy score. A score describes a population, and a complaint is about one policyholder on one day. The reviewer reopens that single file and asks a plain question: was this claim investigated, who made the coverage call, and does the denial letter give a basis a person can stand behind?

Here’s where a lot of AI deployments quietly fail that question. The model produced a number, the number drove a denial, and nobody can say what the model actually read or why it landed where it did. A confidence score isn’t a reason under the statute. “The model flagged it” isn’t an investigation. When the only artifact behind a denial is the model’s output, a carrier has automated its way straight into the practice the act describes, at the speed and volume the act treats as a general business practice.

A logged process turns that around. Once the loss report the model extracted, the severity it assigned, the adjuster who reviewed it, and the stated basis for the denial all live in one run history, the reasonable-investigation-and-explanation standard is satisfied by how the work happened. Put the AI extraction step before a blocking adjuster approval, and the defense is the record itself, not a memo asking adjusters to be careful.

Where to start, and where to bring help

The contained experiment is yours to run this quarter. Pick one workflow, FNOL triage or underwriting intake is the usual first win, drop a model on the extraction step, keep your existing adjuster or underwriter sign-off, and watch what it classifies correctly. Small blast radius, clear owner, easy to reverse, since a wrong draft just gets corrected before anyone acts on it.

Rolling AI across live claims adjudication is a different animal altogether, and it’s mostly a governance and evidence problem rather than a model one. Once a model touches coverage decisions, denials, or anything a market-conduct exam samples, you’re into the documented AI program the NAIC bulletin expects, plus the unfair-claims testing that proves the same claim gets the same handling every time. What carriers tell us, again and again, is that the rollout breaks on the process and the record, not on the model’s intelligence. That’s where an outside read on what to automate, and in what order, changes the odds, because in claims the sequencing is the whole risk.

A clever model behind no process is the thing that fails the exam. The same model inside a defined, logged workflow with a licensed adjuster on every call is the thing a carrier can defend, the way the wider move to workflow automation already carries work that has no AI in it at all.