AI for Patient Intake in Insurance

The fax server releases its overnight batch at seven and the utilisation management queue gains ninety one prior authorisation requests. One is a scan with a handwritten procedure code that could be a three or an eight. Three are duplicates of portal submissions from yesterday, under a different spelling of the member's surname. The reviewer opening them works to a turnaround standard measured in hours for urgent cases, and that clock started when the request arrived, not when anybody read it.

Nobody here calls that patient intake. It is utilisation management, or pre certification, and the people doing it have never met the member whose knee is in the file.

No waiting room, and the same work anyway

Strip the job titles away and the function is the one a clinic performs at its front desk: particulars arrive, get checked against a rulebook, and a decision follows. A payer runs that loop through several doors on different systems. Prior authorisation takes in a clinical request from a provider's office. Enrolment takes in an application and, for late joiners, a health declaration. Case management takes in a member identified from data rather than one who asked. The nurse line takes a symptom description by telephone; the claims floor takes medical records attached to invoices. What differs is that the payer never meets the person, the rulebook is a plan document rather than a clinic protocol, and somebody else started the clock.

A clinical request that arrives three ways at once

The same request reaches you as a fax, a portal submission, a phone call for an urgent case, and in some markets a structured electronic transaction. The portal gives clean member and provider fields and then one free text box holding everything that matters. The fax gives a form the provider designed, clinical detail buried in attachments that may be out of order. The phone gives a request transcribed under pressure. An agent earns its place here by turning those shapes into one: extracting member, providers, procedure, diagnosis and dates, pairing each value with the page it came from, and checking against the open queue, because the duplicate sent through a second channel is otherwise found only by a reviewer who recognises the name.

The policy document behind the decision is longer than anyone reads

Behind every authorisation sits a medical policy, a criteria set licensed or written in house, a plan document with its own exclusions, and whatever mandates apply locally. The criteria are conditional and nested: conservative therapy for a defined duration, imaging inside a window, a documented failure rather than a stated one. An agent can hold them as checkable questions, each carrying the policy version and its effective date, then report which the documents answer, which they do not, and where each answer was found. That output is a gap list with pointers, not a recommendation. Assessing necessity was never the slow part. Discovering that the office never sent the six weeks of physiotherapy notes is, and today that happens after the file has waited a day in a queue.

Turnaround clocks that started before anyone opened the fax

Authorisation turnaround is regulated almost everywhere payers operate, with separate standards for urgent and standard requests, rules on pausing the clock for an information request, and reporting obligations. The costly failure is rarely a wrong determination. It is a late one, or an extension taken because nobody looked at the file in time to see that one document was missing. Completeness triage on arrival changes that arithmetic more than anything else here: an agent that reads each request within minutes, checks it against the policy and raises the missing item the same morning turns a three day round trip into a same day one. Pair it with an ageing view that knows which clock applies and how much has run.

Why the same facts get typed out again downstream

Follow one authorisation through a payer and count the times a person enters the same handful of facts. Member identifier, diagnosis, procedure code, provider number and dates of service go into the utilisation management system. If the case is referred they go into the care management platform. A note goes onto the claim so the adjudicator knows what was approved, and a row goes into the spreadsheet feeding the regulatory turnaround report. Every keystroke is a chance for those records to disagree, and they do, usually discovered when an appeal arrives and the authorisation note does not match the claim. An agent that captures the facts once, carries the provenance of each value and writes downstream through system interfaces removes the divergence rather than reconciling it later.

Joining the scheme, and what the member declares about their health

Enrolment is intake wearing a commercial hat. An application arrives with dependants to verify, effective dates to establish, and in many products a health declaration, because late joiners and voluntary cover carry waiting periods, pre existing condition rules or underwriting questions. Medical schemes add chronic condition registration, where a member applies for a condition to be recognised for benefit purposes and supplies evidence. The checks themselves are objective. Is the declaration complete and signed. Does the attached evidence support the condition declared. Does the application contradict itself. An agent is good at all of it and should stop there, because treating a declaration as inaccurate has consequences for the member and belongs to a person who can be named.

Case management intake begins with someone who did not ask to be called

Case and disease management invert the direction. Nobody submits anything. The member is identified from claims, pharmacy and authorisation data and telephoned by a care manager who has to earn the conversation in thirty seconds. Intake is the assessment that follows: history, medications, barriers, consent, and a care plan in a structured template. Two tasks suit an agent and neither is the conversation. One is preparation, assembling what is known about this member into a brief so the care manager is not reading three screens while a stranger is talking. The other is drafting the assessment record and care plan fields from the call notes for approval. A quieter third is reporting which identified members were never reached.

The triage line is intake with a clinician on the other end

A nurse advice line is symptom intake performed live against a licensed triage algorithm, producing a disposition that may send someone to an emergency department. The algorithm is the regulated part and should not be touched. Around it: identifying the caller and pulling recent claims, authorisations and medications before the nurse speaks, checking for an open case elsewhere, and handling the after call documentation. The nurse asks the questions and makes the call. The agent removes the typing.

Clinical intake on the claims side of the house

Plenty of clinical intake arrives attached to money: inpatient claims with discharge summaries, out of network claims pulled for review, itemised bills and operative reports, retrospective requests where the service happened without authorisation and the justification arrives afterwards. A clinician or coder reads the bundle against what was billed. An agent can index it so it is navigable, saying what each document is, which dates of service it covers, which billed line it supports, and what is conspicuously absent, such as an operative note for a procedure that has been charged. It can also compare the claim against any authorisation on file and surface mismatches in units, dates or site of service.

Coordination of benefits surfaces at the worst moment

Other coverage is the question nobody asks at the right time. A member has a spouse's plan, a retained employer scheme, a government programme that pays primary or secondary, or an injury another insurer may be liable for. The signals show up early: an accident related diagnosis, a workplace injury indicator, a status that changes the order of payment. The value here is timing rather than cleverness. An agent reading intake for those signals raises the question while the authorisation or enrolment is open, when a member or a provider's office will answer willingly, rather than leaving it to a recovery team after payment. Recovery then is slow and damaging. The same question at intake is routine.

Status chasing is a symptom, not a service problem

A meaningful share of every payer's call volume is a provider's office asking where an authorisation has got to, often about a file waiting in a queue for a document that office could have supplied days ago. An agent answering accurately from the case record takes load off the phones, but treat that as a symptom. The structural fix is telling the requesting office precisely what is missing within an hour of arrival, naming the document rather than issuing a generic notice. Keep the same record complete enough to survive an appeal: which criteria version applied, which documents the reviewer saw, who decided, and when the clock stopped.

Where AI fails here

Medical necessity determination is not available for automation. In most jurisdictions an adverse determination must be made by an appropriately licensed clinician, and for many service categories by a physician in the relevant specialty. An agent may show that every criterion has been addressed and where the evidence sits; the determination belongs to a named reviewer who can be held to it.

Deciding that a request is not urgent carries the same weight and is more dangerous because it feels administrative. Treat a claimed urgency as urgent and let a clinician downgrade it.

Extraction from poor scans and handwriting fails quietly. A misread digit in a procedure code produces a plausible wrong answer, not an error message, so anything driving a decision needs a confidence threshold and a person reading the source image below it.

The agent cannot reconcile documents that contradict each other. Where the referral letter and the operative note disagree on which side was operated on, it presents both and chooses neither. It cannot pursue a document the office has not sent, decide that a health declaration was inaccurate, give a member clinical advice, or rescue a policy whose criteria cannot be answered from what providers routinely submit. The regulated clock also remains yours, and an agent unavailable for a morning is not a defence, so the manual path has to stay usable.

Where to start

Take one service line, one channel, one task. Prior authorisation requests arriving by fax for a single high volume procedure group, read on arrival, checked for completeness against the criteria that apply, with the missing information request drafted for a coordinator to send. Leave the determination alone. Leave the urgency classification alone. Leave the denial letters alone. Within a few weeks you will know what share of requests arrive incomplete and what that costs in days, and that figure settles more internal argument than any projection.

The second scope is almost always the rekeying, because once the facts are captured cleanly at the door, writing them downstream is mechanical and the divergence stops accumulating.

Adopt AI Partners builds these agents inside the utilisation management, enrolment and care management systems payers already run, so a completeness check lands on the case rather than in another inbox. If it would be useful to talk it through, describe how clinical requests reach your team to hello@adoptaipartners.com.