Checking Produce Invoices Against Delivery Notes in a Restaurant With AI
Short answer: A restaurant can have AI read each produce, dairy or meat invoice, read the matching delivery note, and list every line where quantity, unit or price differs. The tool does the tedious comparison; a person still decides whether to dispute, accept or credit. It works best when the delivery note is photographed at the back door the same day, when you hold an agreed price list per supplier, and when anything the AI cannot read confidently goes to a human instead of being guessed. This article walks through the check, the failure patterns you will see, and the limits.
Why kitchens lose money between the back door and the ledger
A busy kitchen receives goods at the worst possible time: during prep, with a driver waiting and a chef holding a clipboard. The delivery note gets a quick signature, the crate goes into the walk-in, and the invoice arrives later by email, often as a PDF attached to a message with three other attachments. Nobody lines the two documents up until the end of the month, if at all.
The gaps that appear are usually small and repeated: a case of lemons invoiced but a half case delivered, a price that crept up since the last order, a unit swapped from kilos to cases. Small and repeated is exactly what a person skimming a stack of paper misses and what a patient machine catches.
What you need before any AI is involved
The comparison is only as good as its inputs. Set these up first.
- A photo of every delivery note, taken at receiving, showing the signature and any handwritten corrections ("2 short", "damaged").
- The invoice as the supplier sent it, saved to one shared folder or a dedicated mailbox.
- An agreed price list per supplier, even if it is just a spreadsheet with item name, unit and expected price.
- A consistent item naming sheet, because "Chantenay carrots 5kg" on one document and "Carrots, baby, case" on another will confuse any matcher.
Without the price list the AI can only compare the invoice with the delivery note, which catches shortages but not overcharging.
The check, step by step
- Capture. Receiving staff photograph the note. A phone with a fixed spot near the door and a plain surface beats a cluttered counter.
- Extract. A document-reading step pulls supplier, date, item descriptions, quantities and units from both documents into rows.
- Align. A matching step pairs each invoice line with a delivery line using your item naming sheet, falling back to fuzzy text match when there is no exact hit.
- Compare. Rules, not the language model, do the arithmetic: delivered quantity versus invoiced quantity, invoiced unit price versus price list, line total versus quantity times price.
- Report. The output is a short exceptions list per invoice, in plain words, with the photo and the PDF page linked beside each flagged line.
- Decide. The chef or manager approves, queries the supplier, or requests a credit note.
Keeping the arithmetic in ordinary code matters. Language models are good at reading messy text and poor at being trusted with sums, so let them read and let a spreadsheet formula or script calculate.
Symptom, likely cause and fix
| What the exceptions list shows | Likely cause | What to do |
|---|---|---|
| Invoiced 10, delivered 8 | Short delivery not corrected on the invoice | Ask for a credit note, attach the photo of the signed note |
| Price per kilo higher than list | Seasonal rise or unannounced change | Check recent supplier messages, then agree or contest in writing |
| Same quantity, different unit | Case versus kilo confusion in the item name | Fix the naming sheet so units are part of the item key |
| Line on invoice with no delivery line | Item added to the invoice by mistake, or the note was incomplete | Hold payment of that line, ask the supplier for proof of delivery |
| Delivery line with no invoice line | Goods received but not yet billed | Keep the note; expect the charge on a later invoice |
| Total does not equal sum of lines | Rounding, delivery charge, or an extraction error | Look at the page image before blaming the supplier |
The last row is the reminder that some flags are the reader's fault. Always show the source image next to the claim.
Where the AI should stop
- It does not approve payment. A flagged invoice goes to a person. An unflagged one can follow your normal approval route, but a sample should still be spot checked.
- It does not message suppliers on its own. Drafting a polite query is fine; sending it without review risks a wrongly worded accusation to someone you depend on.
- It does not guess illegible handwriting. If "2 short" could read "3 short", the line is marked unreadable and a human looks.
- It does not judge quality. Wilted herbs or a bruised delivery are for the person who inspected the goods, not the invoice checker.
A small illustrative scenario
This is an illustration, not a real client. A bistro buys vegetables from one wholesaler three times a week. Over a month the exceptions list shows the same two items arriving short on Mondays and one price drifting upward. None of it was visible in any single invoice. With the photos attached, the owner can raise one clear conversation with the supplier instead of three vague complaints.
Keeping it safe and legal
Delivery notes and invoices usually contain business names and sometimes driver or staff names. If you send them to an outside AI service, check the provider's data handling terms, and confirm whether your documents are used for training or retained. For UK businesses the Information Commissioner's Office publishes guidance on using AI with personal data. As of November 2026, most major providers offer settings or plans that exclude your inputs from training; confirm the position in your own account rather than assuming it.
Measuring whether it is worth it
Track three numbers for a couple of months: how many invoices were checked, how many had at least one real discrepancy, and the value recovered through credits. Do not count flags that turned out to be extraction mistakes as wins; count them as tuning work. If recovered value does not exceed the time and tool cost after a reasonable trial, simplify or stop.
FAQ
Do we need special software, or can a spreadsheet do it?
A spreadsheet plus a document-reading AI step is enough to begin. Many kitchens start with a shared folder, a form that captures the note photo, and a sheet that holds the price list and exception results. Dedicated purchasing software can come later if volume justifies it.
What if suppliers send invoices in different layouts?
That is normal, and modern document readers cope with varied layouts. Test with a dozen real invoices from each supplier before relying on the output, and keep a note of any supplier whose format causes repeated errors.
Can this work with handwritten delivery notes?
Printed notes read well. Handwriting is hit and miss, so treat handwritten corrections as a prompt for human review rather than trusted data. Asking drivers to print quantities clearly helps more than any software setting.
Who should own the exceptions list?
One named person, usually the head chef or a manager, so that flags do not sit unread. A weekly ten minute review is better than an inbox nobody opens.
If you would like help setting up a delivery note and invoice check for your kitchen, we would be glad to talk it through on our enquiry page.