How to use AI in insurance: claims and policy admin
In insurance, AI helps most in the back office: it opens the claim from an emailed notification, extracts data from estimates and reports, checks for missing documents and cites the relevant policy clauses. Coverage, quantum, fraud and pricing decisions stay with people. Start with first notification of loss on a single line of business.
In insurance, most of the work that can be automated is not in the actuarial team but in the back office: loss notifications arriving by email, claim files with dozens of documents, policy changes requested by brokers in free text, complaints to classify. This guide is for the operations, claims or digital transformation lead at an insurer or broker who wants to know where AI helps today and where it has no business.
How do you use AI in insurance?
You use AI to read, extract and prepare, and the decision stays with the claims handler or underwriter. The best starting points are the ones where a person reads documents and keys them into the claims or policy system. The table sums up the problems that show up at almost every insurer.
| Problem | What AI can do today | What stays with a person |
|---|---|---|
| First notification of loss (email, PDF, photos, broker files) | Opens the claim record, extracts the data, suggests a priority | Confirmation and assignment |
| The claim file (repair estimates, medical reports, police reports, invoices) | Extracts the fields and checks which documents are missing | Assessment and quantum |
| Checking against the policy wording | Finds the relevant clauses and cites them | The coverage decision |
| Complaints and correspondence | Classifies, summarises, drafts a reply, keeps the trail | The final reply |
| Policy changes requested by brokers | Turns the email into a structured request in the policy system | Approving the change |
1. First notification of loss
Notifications arrive on every channel: an email with three photos, a PDF from a broker, a half-completed form. Someone reads them, opens the claim and assigns it. An AI system can do the first part: it reads the message and attachments, extracts the policy number, date and place of the event, type of loss and contact details, opens the record in the claims system and suggests a priority. The handler confirms.
What it cannot do: decide whether the event is covered, or reject a notification. An automatic rejection would be a GDPR problem (see below), not only a quality one.
2. The claim file and missing documents
A motor or health claim file collects repair estimates, assessments, police reports, medical reports and invoices, each in a different format. AI extracts the fields that matter and runs the checklist: which documents are there, what is missing, where they contradict each other (a different date on the estimate and on the police report, for example). The handler receives an ordered file and a list of questions, not a pile of PDFs. The technical pipeline is described in how to automate document processing with AI.
What it cannot do: set the quantum or conclude that a claim is fraudulent. It can flag an inconsistency; judging it is the person's job.
3. Checking against the policy wording
"Is X covered under product Y, the 2021 version?" is a question handlers and the call centre ask every day, and the answer sits in a 60-page PDF. An AI assistant built on the policy wordings finds the relevant clauses and cites them with a page reference, for each product version. That is how ai-aflat.ro, our assistant over Romanian legislation, works too: it answers only from the source and shows the source.
What it cannot do: give the coverage verdict. The assistant shows the text; a person interprets it and signs off.
4. Complaints and correspondence
Complaints have to be logged, classified and resolved on time, with a clear trail. AI can classify each complaint by type and product, summarise the customer's history and draft a reply. Every step stays in the log, which also helps with reporting to the supervisor.
What it cannot do: send a complaint reply on its own. A person reads and approves it.
5. Policy changes requested by brokers
Brokers request changes in free-text email: "please add a second driver to policy X from 1 November". AI turns the email into a structured request in the policy system and puts it in the approval queue. Older policy systems with no API can be reached through extracts, file drops or, if needed, automation over the interface. The difference between an agent and classic RPA is covered in how to choose between an AI agent and RPA.
What we do not automate in insurance
There are things we do not build, however often we are asked, because the decision must stay with people and because the rules are strict:
- Pricing and underwriting: we do not build models that set the price or accept a risk. In life and health insurance, risk-assessment and pricing systems for natural persons are high-risk under the AI Act.
- Credit scoring: we do not offer it.
- Automatic claim rejections or fraud accusations.
What should an insurer automate first?
Start with first notification of loss, or with extraction from the claim file, on a single line of business. The volume is high, every file already has a person checking it, and today's cost is easy to measure: time from notification to open claim, minutes per file, documents requested twice. The policy-wording assistant is a good second step, because it works only with texts you already own.
A hypothetical example: an insurer with a motor own-damage line would pick the notifications brokers send by email, measure today's time to an open claim, then run a pilot in which a handler checks every record the AI creates. That is the only way to see the real gain: measured in your own pilot. The steps are in how to run an AI pilot.
Data and compliance: GDPR, the AI Act, DORA
A few things to know before the pilot. This is not legal advice; for your case, talk to your lawyers, your DPO and your compliance function.
- GDPR: claim files often contain health data, a special category (Article 9 of the GDPR). And Article 22 limits decisions based solely on automated processing that produce legal effects for a person. Both are good reasons for AI to prepare and a person to decide. More in how to stay GDPR-compliant when using AI.
- AI Act: risk-assessment and pricing systems for natural persons in life and health insurance are listed in Annex III (Regulation (EU) 2024/1689), with obligations from 2 December 2027 under Regulation (EU) 2026/1744. Claims triage and document extraction are not, as a rule, high-risk. An assistant that talks directly to customers has to say it is AI (Article 50, applicable since 2 August 2026).
- DORA: since 17 January 2025, Regulation (EU) 2022/2554 applies to insurers too. An AI supplier becomes an ICT third-party service provider: it goes into the register of information, and the contract has to state where the service is provided and where data is processed. DORA does not require data to stay in the EU, but it does require that this is written down clearly.
- In Romania insurers are supervised by the ASF; its rules on complaints and outsourcing still apply when part of the process is done by AI.
How do I start?
- Tech Call, free: a short conversation where you show us the claims or policy flow and we tell you honestly whether it is worth automating. Book it here.
- Tech Audit: a short, paid audit on real files (anonymised where needed), with the claims system and the people who use it. You get in writing the use case, the data it needs, the AI Act classification and what "working" means in numbers.
- Pilot: one line of business, with handlers checking every result.
- Production: integration with the claims or policy system, an audit trail, monitoring and a person approving. The code and documentation stay with you.
What we can show: an AI assistant that answers from 220,000+ legislative acts with the source cited (ai-aflat.ro) and a contract-analysis project that identifies critical clauses, the same techniques that working with policy wordings needs. What we build is on the AI automation and AI agents pages.
Sources
- Regulation (EU) 2024/1689 (AI Act), Annex III point 5 and Article 50
- Regulation (EU) 2026/1744, the new dates for high-risk systems
- Regulation (EU) 2022/2554 (DORA), Articles 28 and 30
- Regulation (EU) 2016/679 (GDPR), Articles 9 and 22
Under the AI Act, risk assessment and pricing for natural persons in life and health insurance are high-risk systems (Annex III, point 5), with obligations from 2 December 2027. Claims triage and document extraction are not on the list.
Frequently asked questions
Can AI decide on its own whether to pay a claim?
It should not. AI can open the claim, extract the data and check the documents, but the coverage decision and the quantum stay with the handler. Beyond the quality risk, Article 22 of the GDPR limits decisions based solely on automated processing that have legal effects on a person.
Is AI claims triage a high-risk system under the AI Act?
As a rule, no. Annex III covers risk assessment and pricing for natural persons in life and health insurance. Notification triage and document extraction are not on the list, but classification depends on the specific system. This is not legal advice.
Do you build pricing or underwriting models?
No. We do not build models that set a policy price, accept a risk or calculate a credit score. We work on the processes around those decisions: notifications, claim files, documents, correspondence, where a person remains the one who decides.
What does DORA mean for an AI project at an insurer?
The AI supplier becomes an ICT third-party service provider. It goes into the register of information, and the contract has to state where the service is provided and where data is processed, plus audit and exit rights for critical functions. DORA does not require EU-only data, but it does require clarity.
Does AI work with an old policy system that has no API?
Yes, usually through database extracts, scheduled file drops or, as a last resort, automation over the existing interface. You do not need to replace the policy system to start. We begin from the data that can be pulled out today and who approves it.
Want to discuss a project?
Book a free discovery call with the Sapio team.