AI is already in claims, and it is spreading fast. But the useful question is not whether it arrives — it is what it does when it gets there. On the evidence and the law, the answer is augment, not replace: AI reads and assembles the file in seconds; the human still decides, negotiates, and inspects. The real risk is not the machine taking the judgment. It is augmentation used as an excuse to pile more work onto fewer people.
The AI-in-claims debate is loud and mostly held by people who do not work claims. The people who do draw a specific, useful line: the software is very good at reading and assembling a file, and it cannot negotiate a split, stand in a fire loss, or tell fresh damage from years of wear. That line is not wishful thinking — it shows up in the adoption data, the failure modes, and, increasingly, the regulation.
Adoption is real — and it’s mostly the routine parts
AI runs the simple, high-volume claims. It stalls exactly where judgment starts.
The gains are real and concentrated. Leading motor carriers report straight-through processing north of 60% on simple claims under a severity threshold; document extraction and triage cut cycle time on paperwork-heavy files. Aviva ran 80-plus AI models in claims and used them to assist — cutting complex-case liability-assessment time by 23 days rather than removing the human from the call. The pattern is consistent: automation clusters in low-complexity work and thins out fast as the file gets harder.
Where it still can’t go
The parts that resist automation are the parts that were always the actual job.
Where AI has been trusted with judgment on physical damage, adjusters have watched it fail in an expensive direction — accepting everything, buying every roof, because it cannot reliably separate sudden-and-accidental damage from ordinary wear. The division of labor that works is not subtle:
| What AI does well | Where it still fails |
|---|---|
| Reads a policy and summarizes a file in seconds | Negotiates a liability split with a claimant |
| Extracts fields from documents at scale | Tells fresh damage from years of wear |
| Triages and routes by severity | Judges an ambiguous, contested story |
| Runs simple, high-volume claims end to end | Handles the litigated, fatality, and large-loss files |
The law already put a human in the loop
Even where AI could decide alone, regulators say it can’t.
This is not only a capability question; it is now a compliance one. Under GDPR Article 22, a claimant has the right not to be subject to a solely automated decision with significant effect — and to obtain human intervention and contest it. In the U.S., the NAIC’s AI Model Bulletin, adopted in more than twenty states, requires insurers to run a documented AI governance program with human accountability for decisions, and Colorado has gone further with binding rules. The human in the loop is not a nice-to-have; on significant claims decisions it is the law.
The real threat is the business model, not the model
The danger isn’t AI taking the judgment. It’s AI used as permission to pile on files.
The useful precedent is the ATM. It did not end bank tellers; it moved them off counting cash and onto the work that needed a person. AI will do the same to the adjuster’s day — if the industry lets it. The danger is the opposite move: treating AI as license to push more files per adjuster or ship the work somewhere cheaper, so the tool that should have created room instead removes it. Adjusters are less worried about the model than about the business model layered on top of it.
A different AI thesis for claims
The useful version of AI does not replace the adjuster’s judgment — it captures and compounds it.
Mysa is built on the opposite of black-box automation. It automates the assembly — reading the file, chasing what is missing, checking coverage — and hands the adjuster a case that is ready to decide. The call stays human. But the reasoning behind that call gets recorded as part of the decision, so it becomes visible, structured, and learnable instead of locked in one person’s head. The machine does the gathering, the adjuster makes the decision, and the system keeps the why. That is augmentation that makes adjusters more valuable, not more disposable — and it is exactly the model the regulators are asking for.
The people doing the job are right that AI will not replace them. The open question is whether the industry uses AI to give them better work, or just more of it.
Sources
- Deloitte — Generative AI in insurance (2024): ~76% of insurers have implemented gen AI in one or more business functions, claims among the highest-adoption areas.
- McKinsey — "The future of AI in the insurance industry": 60%+ straight-through processing on simple motor claims under a severity threshold; Aviva ran 80+ claims AI models and cut complex-case liability-assessment time by 23 days.
- EU GDPR Article 22 — right not to be subject to solely automated decisions with legal or significant effect, including the right to human intervention and to contest.
- NAIC — Model Bulletin on the Use of Artificial Intelligence Systems by Insurers (adopted Dec 2023; 20+ states): documented AI governance and human accountability for decisions.
- U.S. Bureau of Labor Statistics — Claims Adjusters, Examiners, and Investigators: ~5% projected employment decline (2024–2034) with AI framed as improving productivity, ~21,600 openings a year.
Common questions
Will AI replace claims adjusters?
No — not for the work that carries the risk. The U.S. Bureau of Labor Statistics projects only about a 5% dip in adjuster employment this decade and frames AI as a productivity tool, not a replacement. AI reads documents, summarizes files, and runs simple high-volume claims; adjusters keep the negotiation, inspection, and judgment on complex, litigated, and large-loss files.
Are insurers allowed to let AI decide claims on its own?
Not for significant decisions. GDPR Article 22 gives claimants the right to human review of solely automated decisions, and the NAIC AI Model Bulletin — adopted in 20-plus U.S. states — requires documented governance and human oversight of insurer AI. A human in the loop on meaningful claims decisions is effectively mandated.
What can’t AI do in claims handling?
It cannot negotiate a liability split, inspect damage in person, or reliably distinguish fresh damage from long-term wear — in tests it has tended to accept everything as sudden and accidental. It also cannot handle the human side of a claim, such as guiding a claimant through a serious loss.
Tiago is Co-Founder and COO of Mysa, where he works with claims teams on how liability, subrogation, and leakage decisions actually get made — and how to keep the reasoning behind them from walking out the door.