Summary
"AI claims processing" is used as one phrase for three jobs. Processing is getting a claim complete and routed. Handling is the human work of review, correction and escalation. Adjudication is the judgement about what the policy says. Most automation on the market does the first job well and stops. The value, and the risk, sit in the second and third.
Straight-through processing is not a feature you switch on. It is a permission you earn, claim type by claim type, by showing that an engine's recommendations agree with your handlers on real cases. In the UK, since February 2026, automated decisions with significant effects are legal with safeguards, but health data narrows the permission and the FCA's Consumer Duty treats claims handling as a place where foreseeable harm must be designed out. The practical conclusion is simple: automate approvals inside thresholds, never automate the decline, and keep the reasons on file.
The last section is a six-point checklist for any vendor conversation. If a product cannot cite the page behind a recommendation or run in shadow mode before it is trusted, it is a document extraction tool with ambitions.
Processing, handling, adjudication: three words, three jobs
Claims processing is the administrative path. A claim arrives as a PDF, an email, a portal submission or scanned post. Someone, or something, reads it, extracts the facts into a record, checks the claimant is covered, checks the amounts against a tariff or a schedule, requests whatever is missing, and moves the claim to the right queue. None of this requires judgement about the policy. All of it consumes time, and in most operations it consumes the time of people who were hired for their judgement.
Claims handling is the human layer. A handler owns the claim, reviews what processing produced, decides what to do with it, talks to the claimant, escalates when the case is outside their authority, and leaves a record of why. Handling is where regulators look when things go wrong, because handling is where the customer meets the insurer.
Claims adjudication is the decision itself: does this policy respond to this loss, in what amount, subject to which conditions. In health claims it is often called assessment; in liability lines it shades into coverage analysis. Adjudication is the job that carries legal and financial consequence, and it is the job where "the AI decided" is not an acceptable sentence.
Keeping the three apart matters for a buying decision because the risk profile of automating each is different. Automating processing is low risk and high volume. Automating handling is a question of tooling and audit trail. Automating adjudication needs citations, thresholds, a measurement regime and a line that is never crossed. A vendor who uses the three words interchangeably has not thought about where that line sits.
What straight-through processing means in insurance claims
Straight-through processing, STP, is a claim that moves from notification to payment with no human touch at any step. The term comes from securities settlement and keeps its original meaning: not faster review, but no review.
Industry figures on STP vary widely and most come from vendors. Consultancy surveys circulated in 2025 and 2026 put the straight-through share of standard personal lines claims at around a third, while vendor case studies claim 70% or more on narrow claim types such as travel, low-value property and simple motor glass. The spread tells you the number is about the design, not the technology. A claim type with a defined tariff, a verifiable claimant and a bounded payout can be automated almost entirely. A claim where documents conflict or the policy has to be interpreted cannot, and should not.
Three conditions make STP safe. First, the automated band is defined in advance: claim types, amount ceilings, and the conditions that pull a case back to a person. Second, it applies to approvals only. An automated yes inside the rules is a service improvement; an automated no is a regulatory event. Third, the engine's agreement with human decisions has been measured on real claims before it is allowed to act, and continues to be measured afterwards through a random sample that handlers re-decide. Our underwriting analysis works through the same logic for applications, with the arithmetic for the threshold at which automation pays.
Claims automation versus agents
"Claims automation" has meant different things in each decade. Rules engines encoded the tariff and the authority limits. Robotic process automation moved data between systems that did not talk to each other. Optical character recognition and later machine-learning extraction pulled fields out of forms. Each of these follows a script, and each stops at the first thing the script did not anticipate: a missing document, a discharge summary that contradicts the invoice, a policy wording with an endorsement nobody keyed in.
Those exceptions are where handler time actually goes. A claims desk does not spend its day on the clean 70%. It spends its day on the 30% that bounced.
An agentic system is different in kind, not degree. It reads the whole file, including the hundred-page policy. It notices what is missing and asks for it. When two documents disagree, it says so rather than picking one silently. It applies the rules in your policy pack and cites them by number. It takes a position, with reasons linked to the page each reason came from. And when a handler corrects it, the correction can be promoted to a rule, so the next case of the same kind is handled the way your senior handler would handle it.
The honest limitation is that agents are probabilistic. They will be wrong some of the time, and modern models are confident when wrong. That is why the design around the agent, not the agent itself, carries the safety: thresholds, shadow mode, citations a reviewer can check in seconds, and a human on every decline.
Where the agent fits: six stages
The pipeline we have run in production for a UK healthcare trust has six stages, and the split between automation and judgement is visible in each. The full account is in the AI claims processing case study; the shape is this.
- Intake and digitisation. Every incoming document, whatever the format, becomes a structured claim record. Fully automated.
- Validation and enrichment. The record is checked against membership, coverage and tariff schedules; missing information is requested automatically. Fully automated, with the requests logged.
- Risk and anomaly scoring. Each claim is scored for tariff variance, anomalies and complexity, and every score is explainable. Automated, but the thresholds are the trust's.
- Adjudication. Claims inside the thresholds settle automatically, with the decision, written rationale and payment instruction drafted by the agent. Anything outside, and every potential decline, is referred.
- Routing and escalation. Referred claims reach the right handler with the evidence bundle, risk profile and a draft assessment attached. The handler decides; the agent has done the reading.
- Audit and learning. Every step is logged as it happens, to information governance standards, and handler overrides feed back into the rules.
Note where the human sits: at stage four for every decline and every edge case, and at stage six for every correction. The agent's job is to make sure no claim reaches a handler cold, and no routine approval waits behind one that needs judgement.
Healthcare and medical claims
Health claims are the hardest and the most rewarding case for AI claims processing. Hardest because the documents are clinical, the tariffs are detailed, and the data is special category under UK GDPR, which restricts solely automated decisions more tightly than for a motor claim. Most rewarding because the volume of routine, well-documented claims is enormous and the cost of a handler reading each one is the dominant cost of the operation.
The design response is the same as elsewhere, applied more strictly. Automated settlement stays inside narrow, pre-agreed bands. Anything touching a clinical judgement, a coverage question or a potential decline routes to a person. The evidence bundle the agent assembles, with every clinical and tariff source cited, is what makes the handler's review fast rather than redundant.
What UK rules allow
Two frameworks govern AI claims decisions in the UK, and both changed recently.
The Data (Use and Access) Act 2025 replaced Article 22 of UK GDPR with Articles 22A to 22D, in force since 5 February 2026. Solely automated decisions with legal or similarly significant effects are now permitted in general, provided the individual is told about the decision, can make representations, can obtain human intervention and can contest the outcome. Where special category data is involved, which includes health data, the permission is narrower and the older prohibition largely survives. The ICO's guidance on the new provisions was in consultation in 2026 and firms should check its final form.
The FCA's Consumer Duty requires firms to deliver good outcomes and avoid foreseeable harm, and the FCA has repeatedly singled out claims handling in its general insurance reviews. Its position on AI is that efficiency gains cannot make the claims process slower, more opaque or more burdensome for the customer, and that firms must be able to evidence that automated processes do not produce unfair outcomes.
Put together, the rules do not forbid AI claims processing or even automated settlement. They require that the customer can reach a person, that the firm can explain each decision, and that the firm has checked the system does not harm particular groups. A design that automates approvals inside thresholds, refers every decline to a handler and keeps cited reasons for every recommendation meets all three without a legal argument.
The EU AI Act adds a separate layer for firms with EU reach. Claims handling is not named as high-risk there, as covered in Does the EU AI Act apply to UK insurers?
Judging AI claims management software
Six tests separate a claims engine from a document extraction tool with a chatbot on top. Put each to any vendor, including us.
- Citations to the page. Every recommendation should link each reason to the document and page it came from, so a handler can check it in seconds rather than re-read the file.
- Shadow mode before trust. The engine should run alongside your team on real claims, with agreement measured, before anything is automated. Ask for the agreement figure and how it was computed.
- Overrides become rules. When a handler corrects the engine, the correction should be capturable as a rule, with senior approval, so the system learns your book rather than a generic one.
- Thresholds you control. Claim types, amount ceilings and referral triggers should be yours to set and change, not configuration the vendor owns.
- No automated declines. The engine should be unable to decline a claim without a person, as a property of the design rather than a setting.
- Your cloud, your data. Deployment in your own AWS, Azure or Databricks environment, and a contractual commitment that your claims data is not used to train models.
Our own product, Claim It Quick, is built around exactly these six, and the product film shows each one on anonymised real claims.
Putting this into practice
Start with a fortnight of measurement, not a procurement. Pick two claim types: one high-volume and well-documented, one that regularly bounces. Record, for a sample of each, where handler time goes: reading, chasing missing documents, reconciling sources, deciding, writing up. The split will tell you how much of the work is processing and how much is adjudication, and therefore how much an agent can take without touching a decision.
Then run an engine in shadow mode on the same two claim types for a month. Compare its recommendations to what your handlers decided. Where it agrees, you have a candidate automated band. Where it disagrees, you have either a rule to add or a boundary to draw. Only after that conversation should anyone talk about straight-through processing, and only for approvals.
If you would like to run that exercise with us, book a demo and tell us which claims you handle.
Frequently asked questions
What is AI claims processing?
The use of language models and agents to take a claim from arrival to a validated, structured record: reading PDFs, emails and scanned post, extracting the facts, checking them against cover, tariffs and membership, and requesting anything missing. It is the first stage of a claim's life, and the one most claims automation stops at.
What is the difference between claims processing and claims adjudication?
Processing is administrative: getting the claim complete and routed. Adjudication is judgement: deciding what the policy says about this claim and whether it is paid. Automation of processing is low risk and high volume. Automation of adjudication needs citations, thresholds, shadow-mode evidence and a human on every decline.
What is straight-through processing in insurance claims?
A claim settled with no human touch at any step, from notification to payment. It is only safe inside explicit thresholds, only for approvals, and only after the engine's recommendations have been measured against handler decisions on real claims.
Can AI decline an insurance claim in the UK?
The Data (Use and Access) Act 2025 replaced Article 22 of UK GDPR with Articles 22A to 22D, in force since February 2026. Solely automated decisions with significant effects are allowed only with safeguards: notice, the right to make representations, human intervention and the right to contest. Special category data, which includes health, narrows the permission further. The FCA's Consumer Duty adds a duty to avoid foreseeable harm in claims handling. The practical answer is that declines should stay with a handler, with the engine's reasons on file.
How is agentic AI different from claims automation or RPA?
Rules engines and robotic process automation follow a script and stop at the first exception. An agent reads the file, works out what is missing, requests it, flags conflicts between documents, and explains its reasoning. The gain is not that the happy path gets faster. It is that the exceptions, which are where handler time actually goes, get handled.
What should we look for in AI claims management software?
Citations to the page for every recommendation, a shadow mode that measures agreement with your handlers before anything is automated, logging of overrides so corrections become rules, thresholds you control, deployment in your own cloud, and a contractual commitment that your claims data is not used to train models.