An always-on AI agent can now run most of claims intake for a managing general agent (MGA), a third-party administrator (TPA), a loss adjuster or a broker that handles claims. It watches the claims inbox and portal, reads the notification and its attachments, matches the claimant to the policy, spots signs of vulnerability, requests and chases missing documents, sets the claim up in your claims system, routes it to the right handler and tells the claimant what happens next. It never declines, settles or values a claim. A named handler makes every decision that matters. That split is what makes it work. The agent takes the reading, matching and chasing, and your experienced people keep the judgement that the regulator, the claimant and your capacity provider expect from a person. The technology is ready. What decides whether it reaches production is the state of your records, rules that are written down, and a rollout that earns trust one slice of volume at a time.
Why claims intake suits AI agents now
Claims teams are being squeezed from two sides at once. Talent is now the top challenge for UK insurers, up from seventh a year earlier: 72% find it harder to hire qualified people, 48% say claims management and adjusting face the most acute shortages, and six in ten say labour market conditions are directly affecting their ability to manage claims efficiently (Gallagher Bassett, 2026 Claims Insights). Experience is thinning too. In the Lloyd's market, the share of claims staff with more than 15 years' experience fell from 37% in 2023 to 31% in 2025 (LMA Claims Talent Survey, via Insurance Edge, July 2026).
Claimants, meanwhile, notice the wait. Buildings insurance complaints to the Financial Ombudsman reached a ten-year high, with claim delay the second most common reason at 23%, and the ombudsman found that communication tends to break down when third parties such as loss adjusters are involved (Financial Ombudsman Service).
Intake is where those two pressures meet. Before anyone exercises judgement on a claim, someone has to read it, find the policy, work out what is missing and ask for it. That is exactly the work an agent does well.
What the agent does, step by step
Every process I put on agents follows the same pattern. For claims intake it looks like this.
| Step | What the agent does | What your handler gets |
|---|---|---|
| Watch | Monitors the claims inbox, portal, web forms and call notes, around the clock | Nothing sits unread over a weekend or a storm surge |
| Read | Extracts the facts from the notification, photos, invoices, reports and emails | A structured claim, not a pile of attachments |
| Match | Ties the claimant to the right policy, checks basic cover facts and finds previous claims | The right file, first time |
| Triage | Applies your written rules for claim type, complexity and urgency, with a confidence score | A plain-English reason for every routing |
| Act | Sets up the claim, acknowledges the claimant, requests and chases missing documents | A complete file instead of a half-finished one |
| Escalate | Sends vulnerability signals, low-confidence cases and anything unusual to a named person, with everything it found | Context, not a cold handover |
| Prove | Logs every step and reports volume handled, time to first contact, exceptions and cost | Evidence for your board and your capacity provider |
The table hides the hardest step. Matching only works if your policy and claims records agree with each other, and in most companies I have worked with, they do not quite. That is why the records come before the agent, not after it. I explain why in your AI does not need a bigger model, it needs to know your business.
What the agent never does
The agent does not decline a claim, settle it, or put a value on it. Not at 10% of volume, not at 90%.
There are three reasons, and any one of them would be enough. A wrong decision lands on someone who has just had a flood or a break-in, and it is hard to undo. UK data protection law only allows significant decisions made solely by automated means with safeguards, including the right to ask for a person to step in, and restricts them further where health data is involved, as it often is in injury and travel claims (ICO guidance on automated decision-making). And claims decisions sit squarely inside the FCA's Consumer Duty.
Allianz, which runs one of the best-documented claims agents in production, draws the same line: "By design, payout decisions are never automated" (Allianz, 2025).
The line between reading and deciding is what lets a cautious compliance officer say yes.
What the FCA expects
The FCA is not writing an AI rulebook for claims. It relies on the rules it already has, the Consumer Duty and the Senior Managers regime, and its 2026 insurance priorities say it plainly: "We support firms' use of AI, but they must monitor outcomes for consumers closely" (FCA Regulatory Priorities: Insurance, February 2026).
The same document puts claims near the top of the list. Under "handle claims fairly", firms are expected to "respond promptly, fairly and transparently to consumers and deliver good outcomes when they claim", and the FCA's review of home and travel claims found areas for improvement, "particularly where firms outsource processes". From the second quarter of 2026 it is reviewing how firms oversee outsourcers and delegated authority claims models, aiming to report its findings in early 2027.
For an MGA, a TPA or an adjusting firm, that review is the one to plan for. If an agent handles part of your intake, your insurer partners and capacity providers will want to see what it did and why, claim by claim, and under the Senior Managers regime someone in your firm answers for it. Design for that question from the first week and the findings become something to welcome rather than dread.
Compliance designed in, not bolted on
Many claims products leave these controls for the client to configure. I would rather build them in from the start, because retrofitting an audit trail onto a live agent is far harder than designing one. These are the controls I put into every build, and for claims intake they look like this:
- Vulnerability checks at first contact, using the four drivers in the FCA's guidance on vulnerable customers (health, life events, resilience and capability), with a warm handover to a person, logged.
- An audit trail for every claim that cannot be edited after the fact: the inputs, the model and prompt version, every extracted field with its confidence, the routing reason and any human override.
- Plain-English reasons for every triage and routing outcome, so Consumer Duty outcome monitoring can go to the board.
- An outcomes dashboard: time to first contact, how often files bounce back for missing information, and outcomes for vulnerable claimants compared with everyone else.
- Data kept in the UK or EU, a published list of sub-processors, no training on your data, and the data protection impact assessment drafted for you.
- An outsourcing pack your oversight team and capacity provider can use: exit plan, audit rights and incident notification terms.
- Clear disclosure to claimants that AI is involved, and a route to a person, always.
Why most claims AI never leaves the pilot
None of this is held back by the technology. Today's models read a loss adjuster's report or a photo of a water-damaged kitchen well enough to do this work. Yet Bain's 2025 assessment of 81 property and casualty insurers found 78% using generative AI in some capacity and only 4% scaling it across the organisation (Bain, via Insurance Business).
The gap is the work around the model. A claims pilot proves that a model can classify or prioritise claims on a curated dataset. Production needs reliable data lineage, integration with the claims platform, clear exception thresholds, a named decision owner, human review, monitoring and an audit trail. As a July 2026 piece drawing on Future Processing's research put it: "a pilot tests whether the model can produce an answer. Production tests whether the insurer can rely on that answer" (Insurance Edge, July 2026).
Every item on that list is a gap in skill, data or organisational readiness, and every one is fixable. I describe what fixing them looks like in turning AI theatre into AI that moves the numbers.
Where my method comes from
The method I use for claims intake comes from production work, and I want to be precise about what that work was. At one of Europe's largest insurance brokerages, I led the design and build of an agentic system that resolved 67% of customer service cases without a human. Customer service cases, not claims.
The process was different. The shape is the same: work arriving by email and portal, documents to read, records to match, routine rules, and judgement calls that must stay with a person. The discipline carries straight across. We sorted the data out before building any model. The escalation logic was designed with the people who knew the work. Rollout started at 10% of case volume and grew only as the numbers justified it. Every automated step left an audit trail.
That is the method I bring to claims intake. A claims result will stand on its own measured baseline, not borrow this one.
Start narrow. Allianz shows how
The best public example of claims agents in production is deliberately small. Allianz's Project Nemo, live in Australia since July 2025, handles food spoilage claims under AUD 500 after severe weather. Seven agents check cover, confirm the weather event, look for fraud, propose a payout and write an audit summary, and a claims professional makes the payment decision. Allianz reports an 80% cut in processing and settlement time, with the agents' work done in under five minutes before human review, and the system went live in under 100 days (Allianz, 2025, company report).
The scope is the lesson. One line of business, one kind of claim, one decision held by a person. For a mid-market intake build I would stop earlier still: the agent prepares the file and the handler values it. You do not need a core-system replacement or a team of data scientists to start. You need one well-chosen slice of intake, an agent running on the inbox and claims system you already have, shadow mode first, then a small share of live volume.
The appetite is there. In Gallagher Bassett's survey of MGAs and programme administrators, 61.3% already use AI but only 35.5% are actively budgeting for it (Gallagher Bassett MGA Market Pulse, November 2025). The use is ad hoc. The budget follows a measured result.
What it gives back
Here is the arithmetic I would run for a typical MGA, with every number an assumption until a baseline replaces it with yours. Say 1,500 claims arrive a month, and intake admin (reading, keying, matching, chasing) takes 20 minutes each. That is 500 handler hours a month, roughly three full-time handlers doing work that needs no judgement. If the agent takes 60% of that work end to end, it returns about 300 hours a month, close to two handlers' capacity, and every claimant hears back sooner.
That is what growing without hiring ahead looks like in claims: the next scheme, the next binder or the next storm surge absorbed by the team you already have. Your experienced handlers spend their time on the claims that need them, and that is also how you keep them. Nobody stays in claims for fifteen years to re-key PDFs.
Where to start
If you run claims at an MGA, a TPA, an adjusting firm or a broker, this is the order I would do it in:
- Pick one line of business and one intake channel, ideally high volume and low complexity.
- Measure the baseline: volume, minutes per claim at intake, time to first contact, and how often files bounce back incomplete.
- Check the records the agent will rely on, policy data above all.
- Write down the triage and routing rules your best handlers carry in their heads.
- Run the agent in shadow mode alongside your team, then on 10% of live volume, then scale only on the evidence.
- Keep every decision about cover, value and settlement with a named handler.
A Process Audit covers the first four in two weeks and ends with a fixed-price build spec. A Process Build takes it into production. The same pattern runs well beyond insurance, from credit control to contract renewals, and what AI agents can run sets out the evidence for each. For why the agent itself has become the easy part, see always-on AI agents are here.
If claims intake is where your team's week disappears, I would be glad to hear how it works today. Let's talk.
Related: Always-on AI agents are here. The agent is now the easy part · Your AI does not need a bigger model. It needs to know your business · Turning AI theatre into AI that moves the numbers · What to do when your AI pilot fails
