How to reconcile delivery notes and invoices for physical deliveries automatically with AI

A truck arrives with lemons. The driver photographs the signed delivery note and sends it to the WhatsApp group where deliveries are coordinated. A few days later the carrier's invoice for the freight arrives, and at some point the supplier's invoice for the fruit.

Three pieces of paper for the same delivery, issued by three different companies, arriving through three different channels at three different times. The only way to know whether what you are paying for is what you received is to put them side by side.

How it is done today

Somebody downloads the photo from the group when they can. Both invoices come in by email and go into the pending payments folder. And the comparison, which is the only moment a discrepancy gets caught, happens at month end, if it happens.

Step by step: find the photo among the group's messages, open it and read the note number and quantities by hand, find the purchase order in the ERP, compare by eye, find the supplier invoice in the inbox, compare again, and only then post the goods receipt. For every delivery. In an operation with ten deliveries a day that is a full person doing copying work.

By the time month end arrives, both invoices have been approved. Claiming a two-crate discrepancy on an invoice already paid costs more than the two crates, so nobody claims it. The loss does not show up as a loss: it shows up as a margin slightly worse than last month's.

What the technology does, concretely

The agent works on all three documents as they arrive, not at month end.

It reads the photo of the delivery note. It interprets the image the driver sent over WhatsApp and pulls out the number, date, quantities and whatever is handwritten at the bottom, which is usually the most important part because that is where the person receiving wrote that two crates were missing or one arrived damaged. A photo taken at an angle, in shadow or with the paper creased is processed just the same.

It queries your ERP and checks against the purchase order. What was ordered, what the note says arrived, and what the receiver noted. If all three match, the delivery is clean.

It posts the goods receipt. It writes the movement into the management system against the corresponding purchase order, with the note attached and the batch if there is one. That entry is typed by a person today, two days later.

It checks the supplier invoice against the delivery note. Not against the purchase order, which is the usual mistake: what is invoiced is compared against what actually arrived. An invoice for a hundred crates against a note for ninety-eight is a discrepancy that flags itself, the day the invoice comes in.

It checks the freight invoice against the trip. Transport is billed separately and almost never cross-checked against anything. The agent links it to the delivery that generated it and compares what was charged against what was quoted for that trip.

It opens a case when something does not add up. With all three documents inside, the assigned owner and the difference calculated, before the invoice gets approved.

What changes

The discrepancy is caught the day the paper arrives, not thirty days later. That turns an impossible claim into a two-minute call, and turns copying work into exception handling: one person looks at the deliveries that did not reconcile instead of keying in all of them.

Where it does not go alone

On the decision. The agent cross-checks, flags and assembles the case, but it does not approve a payment, does not issue a credit note and does not decide whether a discrepancy gets claimed or let go. A person does that, and that decision is recorded with a date and an author.

It does not guess either. A handwritten note with illegible writing is not interpreted: it is passed to a person with the photo alongside. An invented figure in a reconciliation is worse than a missing one, because a missing one is visible and an invented one is believed.

Why the channel matters

Because the driver is not going to log into a portal. He is going to keep taking the photo and sending it over WhatsApp, which is what he has in his hand as he climbs down from the cab. What changes is not the driver's habit: it is where that photo lands. Today it lands in a group where addresses, schedule changes and jokes live together. When it lands attached to a delivery, with its timestamp and its author, it is a document.

Where to start

With one route or one supplier, not the whole operation. Connect the channel the photos already arrive through, load that supplier's delivery note format once, and watch for two weeks how many discrepancies show up that were not visible before. That number is what decides whether it is worth extending to the rest.