AI for invoice processing in healthcare: supplier AP, contract pricing and theatre receipts

A rep opens a consignment tray at eleven at night, three implants are used, two sizes are put back. Somebody writes the lot numbers on a sticker and puts it in the theatre record. Six weeks later an invoice arrives for four items at a price nobody in finance recognises, against a purchase order raised retrospectively by an assistant who was told the sizes over the phone.

That invoice is not a data entry problem. It is a question about what happened in a room accounts payable has never been in, and it is why healthcare supplier AP behaves differently from AP in manufacturing or retail, where goods arrive on a loading bay and someone scans them.

This is the accounts payable ledger, not patient billing. Different systems, different failure modes, and considerably less attention.

The invoice population is not one population

Health systems tend to run one AP function over six or seven trades that share almost nothing except a supplier master and a payment run. Treating them as one queue produces a backlog that never clears.

Medical and surgical consumables are high volume, low value per line, and increasingly invoiced from usage rather than delivery. Implants, orthopaedic trays, cardiac rhythm devices and interventional stock arrive on consignment, sit on a shelf in a theatre or cath lab still owned by the supplier, and are sold to you only at the moment they enter a patient.

Pharmacy runs through wholesalers and direct manufacturer accounts, at very high line counts, with contract prices that change on a cycle you do not control and rebates that settle months behind the invoice. Agency and locum staffing is invoiced weekly against timesheets, at rates set by a framework, with a margin on top that is the actual thing worth checking. Capital equipment and its maintenance contracts are low volume and high value, with instalment schedules, acceptance milestones and service cover that may or may not have started. Facilities and estates covers everything from a planned maintenance contract to an emergency call out at a weekend, with wide price variation and weak purchase order discipline.

Each has a different truth source for whether the invoice is right. An agent built to read invoices will do reasonably well on all of them and resolve none of them, because the difficulty is never the reading.

Three way matching when the receipt happened in theatre

Ordinary three way matching assumes a purchase order, a goods receipt and an invoice, created in that order by three separate parties. In a hospital that assumption breaks predictably.

The purchase order for a consignment implant is frequently raised after use, built from what the theatre system or the rep's usage sheet says was consumed. The goods receipt is not a receipt at all. It is a record of usage, entered by a scrub nurse or a theatre coordinator, in a clinical system, in clinical vocabulary, against a catalogue that does not use the supplier's part numbers.

So the match that matters is not purchase order to receipt to invoice. It is invoice line to recorded clinical usage, a reconciliation across two catalogues and two vocabularies.

This is genuinely good work for an agent. It can hold a mapping between supplier part numbers, your item master, the manufacturer catalogue number and whatever free text the theatre record actually contains, and maintain it as suppliers reissue part numbers. It can propose a match with a confidence level and the evidence attached: this invoice line, this procedure, this patient episode, this lot number, this implant log entry.

Where there is no usage record at all, it can say so and route the line to the theatre manager rather than to a buyer who has no way of knowing. That routing decision is most of the value. A large share of consignment disputes are not price disputes. They are invoices for something nobody can evidence was used, and they sit unresolved because the only person who could answer was never asked.

Explant, wastage and the item that was opened but not implanted

A tray is opened, a size is trialled, it is contaminated and discarded. An implant is removed in a revision and returned. A device is used in a case that was abandoned. Depending on the contract, each is either chargeable or not.

An agent can flag every line where the usage record and the invoice disagree on quantity, and classify the likely reason from the theatre note. It should not decide whether wastage is chargeable. That is a contract question and it belongs to the clinical procurement lead.

Contract price, framework price and the price on the invoice

The line price on a healthcare supplier invoice is supposed to be governed by something: a group purchasing organisation agreement, a national framework, a local contract variation, a tender award, a pharmacy contract price, or a manufacturer agreement with tiered volume bands.

In practice that price lives in a contract document, a category manager's spreadsheet, a framework schedule and a price file in the purchase to pay system, and those four do not agree. Variance goes unchallenged because proving the correct price takes longer than the variance is worth on any single invoice.

That asymmetry is what an agent removes. It can check every line against the contracted price rather than a sample, hold the clause that establishes that price, detect uplifts applied earlier than the contract permits, catch indexation applied to the wrong base, and spot a product substituted outside the awarded range at a higher price.

Rebates are their own trade. Volume, growth and compliance rebates accrue against invoiced activity and settle later, and most health systems cannot say with confidence what they are owed. An agent can build the entitlement position from invoice history and the rebate terms, and compare it to what actually arrived as a credit. It will not tell you whether the supplier agrees.

Timesheets, coding, credit notes and short shipments

Agency and locum invoices are a matching problem wearing a pricing problem. The invoice must reconcile to an authorised timesheet for a named worker, at a rate matching the framework band for that grade and shift type, with the correct unsocial hours and on call treatment, and a margin within the capped rate. Duplicate submission of a shift across two agencies, shifts billed against a ward that did not book them, and grade drift are common and detectable. An agent can check all of it per line. A person still confirms the shift was worked, because the ward knows and the ledger does not.

Cost centre and department coding is where most AP rework is created. A theatre consumable coded to the wrong speciality distorts the procedure cost, the service line position and next year's budget, and nobody notices until a clinical director queries their numbers. Coding should come from the evidence chain, the requisitioner, the location of use, the procedure, the item category, rather than a supplier default. When the agent is unsure it should say so, because a confident wrong code is worse than a held invoice.

Credit notes and short shipments are the standing mess of healthcare AP. Stock is returned as out of date, a cold chain line is rejected on delivery, a case is cancelled and consignment restocked, and the credit is promised on a phone call. An agent can hold an open expectation for every promised credit, match arriving credit notes to the invoice line that caused them, and chase what has not appeared. Unapplied credits are real money, usually found only at audit.

The month end accrual scramble

Every month the same exercise runs. Goods received not invoiced, invoices received not approved, theatre usage in the period with no supplier invoice yet, locum shifts worked in the last week, maintenance cover part consumed. Finance builds an accrual from whatever the ledger will produce, adds judgement, and moves on.

An agent can build that position continuously rather than in the last three days. Usage recorded in theatre with no matching invoice is a known liability the moment it is recorded, as are shifts authorised and unbilled, and instalments due and not invoiced. The output is a standing accrual with a line level explanation, which turns month end from estimation into review and shrinks the following month's true up.

Where AI fails here

It fails when the underlying record does not exist. If theatre usage was not recorded, or was recorded on paper and scanned to a folder, no agent can match to it. It will generate confident sounding proposals from partial evidence, and if you let those auto approve you will pay for implants that were never used. The honest output there is an exception, and exception volume will be high in the first year.

It fails on catalogue drift. Suppliers change part numbers, repackage and rebrand. A mapping correct in March silently mismatches in September, and the failure is quiet: lines match to the wrong item at a plausible price. Mappings need ownership and review, not just a model.

It fails on contract interpretation. Whether wastage is chargeable, whether an uplift clause was validly triggered, whether a substitution falls inside an awarded range, these are readings of a document with money and a relationship attached. An agent can assemble the argument and cite the clause. It cannot decide.

It fails on rate cards it has never been given. Most price variance work stalls because the governing price was never loaded anywhere machine readable, and building that reference is a procurement project, not an AI deployment. Anything promising price checking without asking where your contract prices live is checking against the last price paid, which launders the error.

It fails on urgency. A theatre list runs tomorrow and the supplier has put the account on stop. No matching logic resolves that.

And it fails where the fraud is. Duplicate invoices with altered reference numbers, bank detail change requests, and invoices from a supplier whose relationship exists only in the ledger are pattern problems an agent flags well. Acting on the flag stays human, and a bank detail change is verified on a known number.

What this leaves the AP team

Less keying, less chasing, and a much better argument. The team stops reconciling and starts adjudicating exceptions that arrive with the evidence assembled. The clinical procurement lead gets a per line view of whether what was invoiced matches what was used and what was agreed.

Keep approval authority, contract interpretation, supplier relationships and payment release with people. Give the agent the matching, the price checking, the credit tracking and the accrual build. Adopt AI Partners can be reached at hello@adoptaipartners.com.