The signed delivery note, checked the same day: the store signed for 9.5 kg, the note says 10

We read 113 photos of signed delivery notes: more than half the notes we could match had a line the store received short, and about half of those were never corrected. The AI reads the paper, the machine compares it with what was entered, a gap is flagged the same day — what knowing that early is worth.

At the back door of a supermarket, the receiving clerk puts a bag of oranges on the scale, writes "9.5" next to the printed 10, signs and stamps. That sheet is the only proof of what the store actually took — and it travels back to the company in the driver's pocket. In the system the line still says 10 kg, the invoice will print 10 kg, and the store will pay for 9.5.

We read 113 photos of such sheets from two delivery days of a fresh-food distributor, and matched each one to its delivery note in the system. Of the 65 notes we could match, 54% had at least one line the store received short — mostly half a kilo, a kilo. And on about half of those short notes, nobody had corrected the number in the system. The photos were sent in by the drivers, not drawn at random, so the true rate across every note may be lower. It is enough to ask this piece's question: what is knowing about a shortfall the same day worth, instead of waiting for the store's month-end statement?

1. What changes

The driver photographs the signed note on the phone delivery screen, one or several pictures per note. Then:

1. The AI only reads the paper: which items, the printed quantity, any handwritten one, whether there is a signature or stamp, the store name and date on the sheet. A supermarket's printed receiving note reads best — it carries received quantity, unit price and amount, so the AI checks quantity × price = amount on every line itself.
2. The machine compares; the AI does not judge. The number on paper (handwritten if there is one, printed otherwise) is compared with what the driver entered. A gap is a subtraction, not the model's impression.
3. A gap is reported at once: a red mark on the note, a message to the delivery manager, and a notice to the driver — who still remembers what happened at the store. Once the number is corrected the system compares again on the spot, without reading the photo a second time.

The number read and the number entered appear in two corners of the item's cell on the dispatch grid, next to the figure the store itself sends back — so the warehouse sees in one cell whether three sources agree.

2. Why month end is too late

Stores pay by their own receiving note, not by your delivery note. A short line nobody corrects goes two ways, and both cost more when found late:

• A wrong invoice, and then a month-end statement from the store that does not match the books. Accounting digs through note after note, calls the driver about a drop three weeks ago, raises a correction. The same line, fixed the same day, takes a minute.
• Short goods that went nowhere anyone knows. With fresh produce, the missing half kilo either came back on the truck or was lost. If it came back and nobody recorded it, the system does not know it exists — it sits there until it spoils. Known the same day, the driver still remembers, corrects the delivered quantity, the goods return to stock and get sold.

The chart in the calculator above is a funnel: every short line in a month → lines with a photo of the signed note → lines read and flagged the same day. Each step loses some, and what is lost waits for month end as before.

3. First stream: less reconciliation time

How to measure it. Take the last two month-end statements from your largest stores: how many lines disagreed, who settled them, how long from spotting a line to the finished correction. The "after" number comes from the handover log itself: when a gap was flagged and when the driver corrected it.

How far to trust it: fair. Disagreeing lines can be counted on the statements; minutes vary a lot between stores that send a tidy statement and stores that send a photo of every sheet.

4. Second stream: keeping the short goods while there is still time

How to measure it. At the next stock count, separate the variance on the items most often received short; compare it with the total kg short on the handover notes for the same period. Two months before and two after are enough to see whether that loss goes down.

How far to trust it: low — this is the softest line. Both loss shares are assumptions. Set them equal and the calculator still shows 3.23 million a month — positive on reconciliation time alone.

5. What has to be subtracted: the AI reading

Reading one note costs one model call. The calculator subtracts 800 VND per photographed note — about 0.58 million a month in the example. Correcting a number after the read costs nothing extra: the reading is kept, and the comparison re-runs in code.

6. Adding it up

With 8 million VND to build it and two months for drivers to get used to photographing, it pays back in about month 1.7.

7. What this ROI rests on

Whether drivers photograph at all. That assumption sits under all three money lines, and it is the easiest field to overlook. Set the photographed share to zero: every line goes to zero, cost included. That is why the photo is not mandatory but is asked for — pressing Complete without one makes the screen ask again — and why the share of notes with a photo is the first number to watch.

Whether the photo can be read. Not every sheet is alike. A supermarket's printed receiving note reads well. A route sheet — one driver, many stores, circled and corrected by hand all over — does not: in the trial most hand marks were circles with no number, and the AI easily read across the wrong column. For that kind trust only one signal: a whole item struck out.

The unit on paper is the store's. Some print "100 gram" as one unit, so "45" means 4.5 kg. The comparison must convert before it subtracts; where it cannot, it reports "different unit" rather than guessing.

A note without a signature is worth nothing — the system counts it as a gap even when the numbers agree. And a short line can be right: a store refusing poor goods is something that should happen. What matters is that the number in the system says what happened.

8. Measuring it in your own company

1. Take a baseline. The last two month-end statements: how many lines disagreed, how many hours they took, how many kg of unexplained loss on those items.
2. Photograph for two months. Each week track: share of notes with a photo, share of unreadable photos, lines flagged, lines the driver corrected.
3. Compare the month-end statements before and after: the lines still left over at month end are the part of the funnel not yet closed.

9. The thing to remember

The number the driver types and the number the store signs are two statements about the same event. Kept apart, both look right. Put side by side on the same day, and a half-kilo shortfall becomes a one-minute task instead of a hunt three weeks later.

A signed delivery note is not worth anything because it sits in a drawer. It is worth something once it has been read.

The share of notes with a short line was measured from real photos of a food distributor, anonymised and rounded; the photos were sent by drivers, so they are not a random sample. Notes per day, kg short, cost per kg, loss shares and every assumption in the calculator are illustrative. Replace them with your own.