What actually got automated in claims, and what still has not

If you read the insurance technology press this year, you would conclude that claims handling is close to solved.

The figures in circulation are striking. Insurers with full-scale AI adoption reportedly jumped from 8% to 34% in a single year. Straight-through processing rates are said to have moved from between 10% and 15% to between 70% and 90%. Lemonade's often-cited record is a claim settled in two seconds, with the large majority of first notices of loss taken with no human involved at all.

We build and operate a claims management platform for an investigation firm running roughly 25,000 cases a year across 28 legal entities. So we have a specific vantage point on those numbers, and it is worth being blunt about them.

Most are true in the narrow sense and misleading in the way they are read.

Read the denominator before you read the number

Almost every impressive claims automation statistic is measured on a deliberately favourable subset, and the subset is rarely stated.

Lemonade's two-second settlement is real. It is also a low-value personal lines claim: structured intake through an app, a policyholder answering standard questions in a fixed order, and a fraud model that can decline to auto-approve anything unusual. That is a genuinely excellent piece of engineering applied to the most automatable claim in existence.

It tells you almost nothing about a commercial liability claim in Monterrey, where the file arrives as photographs taken on a phone at a warehouse, the policy wording is disputed, three parties disagree about the sequence of events, and the carrier's requirement is a defensible written adjustment.

Lemonade's own published figures make the point better than we can. At the end of 2025 the company reported that 96% of first notices of loss were taken with no human involved, and that 55% of claims were automated end to end. Both numbers are impressive. They are also not the same number, and the difference is the entire subject of this article. The first measures intake, which is a chat form. The second measures resolution. Everything hard lives in the gap between 96 and 55, and almost every summary collapses the two.

The same applies to straight-through processing figures. A 90% STP rate on the claim types selected for STP is not a 90% STP rate on the book. When we have been able to see the underlying definitions, the improvement is usually real and usually about a third of what the headline implies.

None of this means the technology is oversold. It means the published numbers are answers to a question you are not asking.

What automates well, in our experience

Four things work reliably, and they are consistent with what we have written about where AI automation pays: high volume, low variance, a natural verification point, and a stable definition of correct.

Intake classification. Deciding what kind of claim has arrived, which entity and carrier it belongs to, and which queue it should enter. High volume, obvious when wrong, and the error surfaces within minutes because the next person to open it was expecting something else.

Document extraction into structured fields. Policy numbers, dates of loss, vehicle identifiers, amounts, party names. The case we built for tracked 93 fields per case in a spreadsheet before we replaced it. Pulling most of those out of an incoming document rather than typing them is the single largest time saving available in claims, and it is unglamorous.

Completeness checking. Comparing what is in a file against what this carrier requires for this claim type at this stage, and telling somebody what is missing. This is the most underrated automation in claims. It does not decide anything, it cannot really be wrong in an expensive way, and it removes an enormous amount of rework caused by files that go out incomplete.

First-draft correspondence. The letters, status updates and report sections that are written from the same facts every day. A person still reads and signs them. The saving is in typing, not in judgement.

Notice what these have in common. None of them decide the claim. Every one of them is checked by somebody who was going to look at the work anyway.

What does not automate, and keeps not automating

Coverage decisions on anything contested. The moment a claim's outcome turns on how a policy clause applies to a specific set of facts, you are in the business of interpretation, and interpretation is exactly what a model produces confidently and unaccountably. This is also where being wrong is most expensive and where discovery is most likely to surface your reasoning.

Anything where the file is the only evidence. In investigation work, the adjustment file is what a court sees two years later. An extracted figure that nobody re-read, feeding an invoice and a report and an audit trail, is not a productivity gain. It is a liability with a confidence score attached.

Multi-party fact reconstruction. Where the work is reconciling three accounts of the same incident, the value is in noticing what does not fit. Models are optimised for producing plausible narratives, which is the opposite of the required skill.

Exception handling, permanently. Every claims operation has a long tail: the wrong currency, the policy that was endorsed last week, the entity that handles this carrier differently. The tail does not shrink because you automated the head. It becomes a larger share of what humans do, which changes who you need to hire.

The regulatory turn is the real 2026 story

While the industry published throughput numbers, the ground moved underneath them, though not in the way most summaries claim.

Colorado's AI Act, SB 24-205, is still widely cited as having taken effect on 1 February 2026. It did not. Implementation was pushed to 30 June 2026, and the framework was then substantially rewritten by SB 26-189, signed in May 2026 and due to take effect on 1 January 2027.

Two things are worth taking from that, and neither is the headline.

The first is that insurance was treated as a special case rather than a general one. Under the original act, an insurer was deemed compliant if it was already subject to the insurance commissioner's rules on external consumer data, algorithms and predictive models. That pattern keeps repeating: legislators conclude that insurance already has a supervisory apparatus and should be governed through it. The requirements that actually bind you are far more likely to come from your insurance regulator than from a general-purpose AI statute.

The second is that the churn is itself the lesson. Two years of delays and rewrites in the most closely watched state means nobody can sensibly build to a specific text. What has survived every version is the same underlying demand: document what the system does, test it for disparate outcomes, and be able to name who is accountable for it.

That is the durable requirement, and it is a governance requirement rather than a technology one. Somebody senior has to be able to sign a statement about how decisions get made.

If you cannot currently produce, for a given claim, a record of which steps were automated, what the model saw, what it output, what confidence it carried and who reviewed it, you do not have a compliance gap in the future. You have one now, and it will be discovered at the worst possible moment.

This is why we build the audit trail before the automation, not after. It is the least interesting part of the system and the part that determines whether the rest of it is usable.

What we would build first

If you run a claims operation and want the benefit without the exposure, the sequence that works is not the one most vendors propose.

1. Make the case record complete first. Automation on top of correspondence living in one adjuster's inbox and evidence in a shared drive produces fast decisions on partial information. Fix the record, then automate.
2. Start with completeness checks. Highest ratio of value to risk in the whole category. Nothing is decided, and rework falls immediately.
3. Then extraction, with a threshold you own. Above the confidence line it files automatically; below it, a person looks. That number is a business decision, not a technical setting, and somebody on the business side should control it.
4. Then drafting, with signature retained. The person stays accountable for what goes out.
5. Leave coverage decisions alone until you can explain, in writing, to a regulator, how any given one was reached.

The honest summary

Claims is one of the best genuine fits for AI in any industry. The volume is real, the documents are repetitive, and the manual work is enormous.

But the automatable part is the clerical layer of reading, extracting, checking and drafting, not the adjudicative one. The firms getting durable value are automating the typing and keeping the judgement, and they are building the record that lets them prove which was which.

The ones buying the two-second settlement are usually buying a demo of somebody else's easiest claim.

If you want a view on which parts of your claims operation are actually automatable, that diagnostic is where we start, and the claims platform describes what the full lifecycle looks like when it runs on one system instead of six.