Measuring the value
The store's photo in the group chat fills in the delivery note: 2 photos matched, but only 3 of 46 stores linked
The store receives the goods after the driver has left, checks them and sends a photo of the corrected note, "10 − 1 = 9", to the Zalo group; the machine finds the delivery note, records delivered, support and returns, and queues it for invoicing. Each photo is worth a few thousand dong — the real money depends on how many stores have their group linked.
Stores send the note's photo to the chat group, the numbers land on the delivery note — how many stores make it worth it?
All it needs is delivery notes to the chain per month, the share from stores whose chat group is linked, the share the store photographs and the share matched to the right note by itself. Every money figure follows from those. Every one is editable.
Derived from the four numbers above
Assumptions — click to see and edit each one
Accountant hours
Cash in earlier
Cost
This is an estimate, not a promise. Set the share of notes with a linked group to 0: every line is zero, cost included. Each "needed" tick holds the other three levers where they are — a tick past the end of the scale means that lever cannot do it alone.
At a busy chain store the driver does not wait for a signature. The goods are left, the truck moves on. Some time later a store employee checks them: one bag short, writes "10 − 1 = 9" on the delivery note, signs, photographs it and sends it to the Zalo chat group shared with the supplier. On the supplier's side the photo drifts down the group until an accountant has time to dig it out, find the right note, fix the numbers, and only then make the invoice.
The change in this piece: the moment the photo lands in the group, the machine picks it up, finds the delivery note, writes the numbers from the paper, and queues the note for invoicing — nobody digs. We measured the first live days at a fresh-food distributor delivering to a supermarket chain. The result fits in one sentence: both photos that arrived landed on the right note by themselves — and that is every photo there was, because only 3 of 46 stores have a chat group linked to the system so far. This piece works out what one photo is worth, and how many stores it takes for the whole thing to be worth it.
1. What changes
- The store's chat group is linked to the customer record. A company Zalo account is already in the group; its messages are saved to the system's inbox against the right store. A message with a photo, from the group of a store in the chain, is the machine's job.
- The AI reads each photo once: is it a delivery note, which one, what each line says. The only candidates are the delivery notes to that same store in the last seven days.
- Matching. If the AI is certain, or the note number printed on the paper matches exactly one note, the photo goes into that note's handover record. If not, it waits in an "unmatched" frame on the invoice review screen, with suggestions, for the accountant to pick.
- The numbers are written from the paper — on draft notes only — and the note is marked delivered, just as when the driver presses Complete on the phone. A posted note is not changed; the gap goes to the review screen.
- Queued for invoicing. A note that qualifies is invoiced at exactly what the store received; a note with anything odd waits for the accountant. No "gap" message is sent back into the group — the store has just sent that very paper.
The photo goes through the same reader as the signed delivery note the driver photographs; the difference is that the sender is the store, so the numbers on the paper count as its confirmation, even when the signature lands in the wrong box.
2. The writing rule: "10 − 1 = 9" is not "10 corrected to 7"
Two handwritten forms that look almost the same are two different transactions, and the rule is written down so the machine does not have to guess:
| The store writes | The system records |
|---|---|
| A subtraction: "10 − 1 = 9" | Delivered 9, support 1 |
| Only a new number: "10" → "7" | Delivered 7, returned 3 |
| "return", "exchange", "taken back", "refused", or a separate return number | That part is a return |
| The whole line struck out | The whole line returned |
| Received more than shipped | Nothing written, waits for a person |
| An item added by hand that is not on the note | Look it up in what the chain has received; if certain, add a line, if not, wait for the accountant to confirm |
An item on two lines of the note is not split by the machine; a photo sent twice is handled once.
3. How much has run
| Item | Measured |
|---|---|
| Stores in the chain | 46 |
| Delivery notes to the chain in 30 days | 406 (33 stores) |
| Stores with a linked chat group | 3 |
| Share of the 30 days' notes from those 3 stores | 45 / 406 = 11% |
| Notes to those 3 stores since launch / photos the stores sent | 5 / 2 |
| Photos landing on the right note by themselves / placed by hand | 2 / 0 |
First photo: the paper matched the note, no numbers to change; the note was queued after 6 minutes, qualified, and is waiting for the accountant to approve — the photo arrived before automatic invoicing was switched on. Second photo: an item shipped as 10, the paper says 7 with no subtraction, so the system recorded delivered 7, returned 3; the store had also written in 5 kg of a fruit not on the note, and the machine found that item among what the chain has received and added a line. The photo arrived; 18 seconds later the numbers were on the note and the note was queued; 16 minutes later there was an invoice. That first invoice had to be cancelled and redone the same afternoon because the buyer details, item names and prices on the invoice were not yet configured — a first-day setup mistake, but worth telling.
Two photos are not a statistic. They show the mechanism works end to end, and they show where the real bottleneck is: 11% of the notes.
4. First stream: accountant hours
How to measure it. Time the accountant handling ten note photos from the chat groups the old way, then count next month's photos: matched by themselves, placed by hand, "not a note".
How far to trust it: medium per note, low on volume — the 80% matched share is a more cautious assumption than the 2 of 2 measured.
5. Second stream: invoice out earlier, cash in earlier
How far to trust it: low. Before this change the chain's invoices were made in the accounting package and synced back, not linked to delivery notes, so how many days late they were cannot be measured in the system. This line is small, and we leave it small.
6. What to subtract: the AI reading photos
One reading per photo; the calculator subtracts VND 1,000 a photo: about VND 18 thousand/month in the example above.
7. Adding it up
| Stream | Per month | Per year |
|---|---|---|
| Fewer accountant hours | 83 K | 1.00 M |
| Invoice out earlier | 23 K | 0.27 M |
| AI reading photos | −18 K | −0.21 M |
| Net benefit | 88 K | 1.06 M |
Each photo matched by itself is worth about VND 7.4 thousand; after the reading cost, each photo that arrives is worth about 4.9 thousand. The mechanism pays from the very first photo — but 18 photos a month add up to less than a hundred thousand.
The bullet chart in the calculator asks the question the other way round: with a setup cost of VND 8 million and payback wanted within 12 months (about 667 thousand net a month), how far would each lever alone have to go? The share of notes with a linked group would have to reach about 83%. The share the store photographs, the matched share or the minutes saved per note — none of them can do it alone; the "needed" tick runs off the scale.
8. What this ROI stands on
How many groups are linked. This is the only lever big enough, and it is people's work, not the machine's: add the company account to each store's group, link the group to the customer record. Set the share of notes with a linked group to 0: every line is zero, cost included. Set it to 100% with everything else unchanged: about 0.80 million a month; if stores photograph 80% of notes: about 1.6 million.
Whether the store sends a photo. 2 photos for 5 notes is too few to call a rate. A store only sends one when the driver has left before the goods were checked; where the store signs on the spot, the photo is not needed.
The matched share. 2 of 2 is right, but a store that takes two deliveries within a few days will leave the AI unsure, and that photo will wait for the accountant to pick. The calculator uses 80% until there is a first month of numbers.
The writing rule is a business rule. "A subtraction is support, a new number is a return" is this company's convention with this chain. Where the convention differs, the rule has to change with it, not be left for the machine to guess.
9. Measure it in your own company
- Take a baseline. For one week: count the note photos stores send into chat groups, time the accountant handling them, and record the days from delivery to invoice for those notes.
- Link the groups before measuring further. The deciding number is the share of notes going through a linked group; track it weekly as a target.
- Measure again after 7 and 30 days: photos received, matched by themselves, placed by hand, "not a note", hours from photo to invoice — then put the real numbers into the calculator.
10. The thing to remember
A photo in a chat group is already evidence; the system's job is not to let it drift. The machine does that part in minutes, and got both first photos right. The part that decides the money lies outside the machine: how many stores have their group linked.
Automating one photo is easy. The hard part is making every photo come through the same door.
Measurements are from the first live days at a food distributor and a supermarket chain, anonymised; the sample is very small and will be measured again. Minutes, days earlier, cost of capital, AI cost and setup cost are illustrative. Replace them with your own.
Read next
-
Make-up orders from a supermarket chain: one PO in 26, and only the receipt tells
A store raises a PO and receives it the same day to pay back goods already delivered without one. The PO header still says "deliver tomorrow", so a make-up order is easily delivered twice and the goods behind it never invoiced. Measured on 1,258 POs, with a matching threshold chosen from the data.
-
One screen for a delivery crew's day: 43% of notes posted before the truck delivers, and returns nobody can record
The warehouse posts the delivery note as the truck leaves; the driver's screen read it as "done" and locked the returns field. Over 14 days, early-posted notes had a signed slip 1% of the time and a returned quantity 0.4%, against 89% and 15.5%. Reopen the note for the driver, add a vehicle & fuel tab that reads photos with AI — what that saves a month.
-
Exceptions that arrive with a reason: 82 slips to check, but only 68 need someone to think
When invoices to supermarket chains raise themselves, the remaining slips reach the accountant with a reason in words. Read by what has to be done, the reasons fall into three kinds: judgement, wait, one-time fix. Measured on 291 slips, and the price of the minutes spent searching.