AI supplier invoice processing.

Where the money leaks in hospitality invoices

AI supplier invoice processing for cafés and restaurants is about the lines, not the header. Capture tools already handle the header well: Hubdoc says it extracts supplier names, amounts, invoice numbers and due dates into Xero, and Dext offers line-item extraction. What those tools do not know is that you agreed NZ$18.50 a kilo for brisket, that the docket the chef signed on Tuesday showed two cases of lemons short, or that the “200g feta” was swapped for a dearer brand.

In a busy kitchen nobody checks 600 invoices a month line by line. Overcharges, unclaimed credits and quiet price creep add up. This workflow checks every line and hands the owner a short dispute list.

How the workflow runs

Inputs

Supplier invoices and credit notes (email PDFs, portal downloads, photos of paper invoices), signed delivery dockets with shortage notes, agreed price lists or supplier weekly price sheets, and the chart of accounts with venue tracking.

Systems

An existing capture tool or shared inbox, Xero or MYOB through their APIs (draft bills only), a price agreement table, and a model API.

Steps

Extract every line (product, unit, quantity, unit price, GST); normalise units; match products to the price list; compare price within category tolerance; match quantities to the docket; check for substitutions and duplicates; draft the bill coded by venue and category; build the dispute list.

Human review

The bookkeeper clears extraction exceptions. The owner or group manager approves each dispute before it goes to a supplier, approves payment runs, and checks any new supplier or changed bank account. The head chef confirms substitutions.

Outputs

Draft bills in the ledger, a weekly dispute list with evidence, a price-movement report by supplier, and purchases by venue and category.

Measures

Credits claimed and received, price variances found per supplier, lines needing manual correction, and days from invoice to coded.

Risks

Misread tables, unit mismatches, false alarms on market-priced produce, late credit notes counted twice, and payment fraud.

Illustrative example: a two-venue Auckland café group

Illustrative example, synthetic numbers: two cafés spend about NZ$55,000 a month on food, beverage and consumables across 600 invoices averaging 12 lines. Assumptions: 1% of that spend is billed above agreed prices or for goods short-delivered without a credit, and the group recovers most of it once flagged; reviewing the dispute list takes the owner about 2 hours a week.

The fee lines assume 5,000 input tokens per invoice (page images, price list excerpt, instructions) and 1,200 output tokens of structured line items, at US$0.10 and US$0.50 per million tokens for Haiku 5.5 or US$2 and US$10 for Sonnet 5.5 (Anthropic, as listed on 19 October 2026). Test the cheaper model on 100 real invoices first; use the larger one only if it misreads dense tables. Your leak rate may be higher or lower than 1%. Measure it in the pilot before building a business case on it.

NZ rules the workflow should check

• GST records. Inland Revenue’s taxable supply information rules require the supplier’s GST number on supplies over $200 and, for a GST-registered buyer, your name and an identifier such as an address or NZBN on supplies over $1,000. Flag invoices that lack them before the GST claim relies on them.
• Food records. The Food Act 2014 imposes documentation, traceability and recall requirements on food businesses. Keep original invoices and dockets in your own store, linked to the extracted data, so you can trace a product back to its supplier if a recall happens.
• Data terms. Xero’s developer terms prohibit using data obtained through its APIs to train AI or machine learning models. Ask any tool vendor how they comply.

What stays with a person

The owner decides whether to raise a dispute, how hard to push a long-standing supplier, and when to change suppliers. Payment approval, new suppliers and any change of bank account stay with the owner. The chef decides whether a substitution was acceptable. Coding judgements beyond the standard categories go to your accountant.

Where invoice processing goes wrong

• Units. A case of 12 billed per bottle, or kilos against each. Normalise units before comparing prices.
• Market prices. Avocados move weekly. Category tolerances or the supplier’s weekly sheet prevent a flood of false alarms.
• Credit timing. A credit note arriving three weeks later must close the original dispute, not create a new one.
• Payment redirection. An invoice with new bank details, or text in a PDF telling the system to update them, is a fraud pattern. Never let an invoice change payee details.
• Unread lists. If the dispute list runs to four pages, the owner will skip it. Rank by dollar value and cap the weekly list.

Coding and exception rules for the ledger itself are covered in invoice coding and exception review, and connection details in our Xero integration guide. Front-of-house workflows are reservation enquiry handling and the guest support knowledge assistant. Our AI document processing service builds this kind of check, and the hospitality overview covers the sector.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.