Ninety percent of claims arrive in reasonable shape: a recognisable form, a readable policy number, the supporting documents attached. Automating those is well understood. The programme's reputation, and most of the handling cost, is decided by the other ten percent. The handwritten discharge summary. The nine-page PDF where the bill is on page six. The policy number with a transposed digit. The design for claims intake has to start from the messy cases, because that is where the people will be spending their time.
What arrives, and why it is hard.
- Multiple channels: portal, email, hospital or garage desk, third-party administrator feeds, post that is scanned.
- Compound documents: one PDF containing the claim form, bills, prescriptions, reports and identity proof, in no fixed order.
- Mixed quality: printed, handwritten, photographed on a phone, faxed and scanned again.
- Reference data that does not resolve: policy lapsed, member not on the policy, provider not empanelled.
- Sensitive content throughout: diagnosis, treatment, identity, in jurisdictions with strict health-data rules.
A pipeline built for the hard cases.
Split before you classify.
The first job on a compound document is page-level splitting: decide where one document ends and the next begins. Then classify each resulting document. Pipelines that classify the whole file as a claim and try to extract from page one are the ones that fail silently on page six.
Classify with an escape hatch.
Claim form, itemised bill, discharge summary, prescription, diagnostic report, identity proof, and other. The other class is the escape hatch. Anything the model is not confident about goes there and is routed to a person with the page image. Over time the other class tells you which document types deserve their own model.
Extract with field-level thresholds.
Set thresholds by field, not by document. A low-confidence hospital name is a routing nuisance. A low-confidence claimed amount is a financial risk. Route the first for a quick confirmation and the second for full review. Handwritten fields get their own, lower, expectations and a review step by default.
Validate against the policy, not just the form.
Once the fields are out, the real checks begin: policy active on the date of treatment, member covered, provider in network, treatment within the product's coverage, waiting periods, sub-limits, prior claims against the same event. These are rules that change with every product version, so they belong in a rules engine or a configuration, versioned, and never in the bot's code.
The reviewer's screen is part of the design.
The ten percent will be handled by people. Their tooling decides whether the automation helps them or merely moves the work. The review screen should show the page image and the extracted fields side by side, highlight the fields that need a decision, pre-fill everything else, and let the reviewer correct a value with one click. Every correction is logged and fed back as training data. A reviewer who has to retype the form has been given a slower process, not a faster one.
Health data, handled properly.
- Minimise what the bot stores. Extract what adjudication needs and leave the rest in the source document store with access controls.
- Mask identity and health fields in logs and in any data sent to a model or service outside the controlled environment.
- Apply the applicable rules: HIPAA for US health data, GDPR for European subjects, the DPDP Act 2023 for Indian operations, plus sector regulators' outsourcing and data-location requirements.
- Keep the audit trail of who viewed and changed what on each claim, including bot accounts.
- Define retention and deletion for intake documents, and automate it.
What good looks like.
In a well-run intake automation, clean claims move to adjudication within hours of arrival rather than days. The reviewer queue holds only the cases that need judgement, each with a reason code and the evidence laid out. The straight-through rate rises month on month as the top reason codes are fixed. And the people who used to key claims now review them, which is a better job and a better control.
Design for the claim that arrives as a phone photograph of a handwritten bill. The clean ones will take care of themselves.