AI for Invoice Processing in Hospitality: Food Invoices, Price Lists and the Site P&L Deadline

Hospitality accounts payable is not a harder version of manufacturing accounts payable. It is a different problem with the same name. The invoice count is enormous, the values are small, the goods arrive daily, and the person receiving them is a chef at six in the morning with a service to get through and no interest in your document flow.

That combination produces a specific failure. The money leaks two and five pounds at a time, thousands of times a month, in price creep, short deliveries and credits that never arrive. No single invoice is worth stopping for, and the aggregate is worth a point or two of gross profit.

Then the timetable sits on top. The business expects a site P&L within a few days of period end, with cost of sales broken down well enough that an operations director can challenge a head chef on food GP. Accounts payable is the bottleneck in that promise, and everybody knows it.

What makes the hospitality invoice file distinctive

Start with the volume shape. A thirty site group can take tens of thousands of purchase invoices a year, most worth a few hundred pounds or less, from three or four national distributors plus a long tail of butchers, bakers, fishmongers, local breweries and the man who brings the eggs.

The national suppliers are the easy half. They send structured data, hold an agreed price file, and will send a statement you can reconcile. The tail is where the cost sits, because those invoices arrive as paper left in a drawer at the site, as a phone photograph, or on a delivery note doubling as the invoice.

Two other things separate this from most accounts payable work. Very little of it has a purchase order in the conventional sense: ordering happens against a catalogue and an agreed price list, often by phone, and the commitment is a standing arrangement rather than a document. And the goods are consumed before the invoice is approved, so there is no holding delivery pending a query. The food was served on Saturday night.

Capture starts at the back door, not in the finance inbox

The most important design decision here is where capture happens. If the delivery note is signed at the back door and then travels to head office in an envelope, you have already lost. That signature is the only moment anybody compares what was ordered to what turned up, and it happens in a cold yard under time pressure.

So push capture to that moment. The practical version is a phone in the hands of the person signing: photograph the delivery note, confirm or amend quantities received, flag anything rejected or returned. An AI agent reads a creased handwritten docket far better than template based scanning did, and does not need a layout you have seen before.

What the agent can reliably extract

Supplier identity, document number, date, site, line descriptions, quantities, units, unit prices, line totals, VAT treatment, and whether the document is an invoice, a credit note, a statement or a delivery note. That last classification matters more than people expect, because credit notes posted as invoices are a common source of unexplained variances.

Units are the hard part. A line reading "beef mince 6 x 2.5kg" has to become a comparable unit price before any price check means anything, and suppliers change pack sizes without telling anybody. An agent that normalises units catches the price rise disguised as a pack change, which is the most profitable thing it will do for a food buyer.

Matching against the price list, the order and the delivery note

In this industry the three way match has a different third leg. You match invoice to delivery note to agreed price list rather than to a purchase order, and the two questions are separate.

Did we receive what we are billed for? That is quantity, answered by the delivery note as amended at the door. Lines billed but not delivered, delivered short, rejected on quality, and substitutions the driver made on the day all surface here.

Are we being charged the price we agreed? That is the price file question, pure arithmetic once units are normalised. The agent compares the invoiced unit price to the contracted price for that site, supplier and effective date, and flags the variance with the contracted figure beside it. Promotional prices, seasonal market price lines on fish and produce, and volume bandings need loading as rules, not treated as exceptions, or the queue fills with noise.

Credit notes, short deliveries and returns

Credit note chasing is the quiet drain on an accounts payable team here, and where an agent earns its place. A short delivery is noted at the door, the driver promises a credit, the invoice is paid in full because month end is tomorrow, and the credit either never arrives or turns up six weeks later against a different invoice number, in the wrong period.

An agent can hold that thread. It records the expected credit when the shortage is flagged, keeps it open against the site and supplier, recognises the credit note when it appears even under a different reference, and matches it back. Where nothing arrives it chases and drafts the query for a person to send. The open credit position becomes a number the controller can see rather than a list in somebody's notebook.

Coding, cost centres and cost of sales you can defend

Coding is where multi site structure bites. If each property codes its own spend, the same crate of tomatoes lands in food, in sundries, or in kitchen consumables depending on who touched the invoice, and your cost of sales by category stops being comparable across the estate. GP reporting then becomes an argument about coding rather than a conversation about buying and portion control.

An agent does this well because it is pattern work at volume. It codes to nominal and cost centre from the supplier, the product description and the historical treatment of that line across the group, applying the group's convention rather than the site's habit. It should also flag inconsistency, not just assign a code: where one site has coded a recurring line differently from everyone else for a year, that finding is worth more than the postings.

Fixed cost invoices travel a different path and should be kept separate. Rent, rates, utilities, insurance and service charges are low volume, high value, usually already accrued, and validated against a lease or contract rather than a delivery note. Utilities also need meter reads and estimated versus actual billing checked, which is closer to contract audit. Mixing the streams puts a rates demand in the same queue as a bread invoice.

Approval routing and where the thresholds should sit

Routing by site and threshold is simple in principle: general manager up to a limit, area manager above it, finance director above that. The problem is that approving a consumed food invoice is largely ceremonial. The goods are gone and the money is owed, so a general manager clicking approve on four hundred lines of produce is not a control.

The more useful design routes exceptions, not invoices. If the invoice matches the delivery note and the price file, post it and tell the site it was posted. If it does not, route the specific variance to the person who can answer it, with the contracted price, the delivery note and the history in one view. That turns site approval from a rubber stamp into a short answerable question, which is the only version that survives a busy week.

When to chase a supplier and when to let it go

Statement reconciliation is mechanical and well suited to an agent. It compares the statement line by line to the ledger and produces the differences: invoices on the statement you never received, invoices you hold that they have not billed, expected credits that do not appear, and payments in transit. That is the basis of a sensible conversation rather than a long email thread.

Query handling is partly automatable. Drafting the query, assembling evidence, chasing and logging the response can all be agent work. Deciding whether to pursue it cannot, particularly with a local supplier you depend on.

Where AI fails here

Bad reference data defeats the whole thing. If the price file is out of date, incomplete, or held per site in spreadsheets nobody maintains, the agent validates prices against the wrong figures and produces clean, confident, wrong results. Most of the work in these programmes is getting price lists, product mappings and pack sizes into a current machine readable state, which is worth saying before a deployment rather than during one.

No delivery note means no quantity check. If the site signs a paper docket that never gets captured, or signs a blank one, the agent cannot tell you whether the goods arrived. It can only tell you the price looks right, which is the weaker half of the control.

Handwritten amendments remain genuinely difficult. A crossed out quantity with a figure scrawled beside it and a chef's initials is readable by a person almost every time and by a model noticeably less often. Those lines belong in a human queue, and a system that pretends otherwise posts the wrong quantity.

Judgement about suppliers stays human. Whether to reject a price increase, whether a repeated shortage is a depot problem or a site not counting properly, and whether to escalate during a supply squeeze are commercial decisions. The agent finds and evidences the variance. Someone with a relationship decides what to do.

Stock remains the real control. Invoice accuracy improves cost of sales but does not replace the weekly count, and an agent cannot see waste, theft, over portioning or inter site transfers. If food GP is wrong and the invoices are correct, the answer is in the kitchen.

Where to start

Run it on history first. Take a quarter of posted food invoices from three or four sites, with the price files in force at the time, and have the agent recalculate every line. You risk nothing, the overcharges and missing credits it finds tend to cover the exercise, and you learn which suppliers and sites are the problem.

Then fix capture at the back door, because every downstream control depends on it. If the chef cannot photograph and amend a delivery note in twenty seconds, it will not happen.

If accounts payable at your group means a fortnight of keying followed by a month end nobody enjoys, describe it to Adopt AI Partners at hello@adoptaipartners.com and we will say plainly which parts an agent could take.