Measuring the value
One photo of the day's delivery list: 462 printed cells, 13 handwritten numbers, the machine sure of 1
A driver photographs one printed list for the whole day; the AI looks for hand marks on the grid and the machine compares handwritten numbers with delivered quantities. Two live days: 10 lists, 13 handwritten numbers, the machine certain of 1 — the value lies in putting the right cells in front of a person, and this is how to price it.
One photo of the day's delivery list — how much can the machine conclude, and what is it worth?
All it needs is drivers per day, notes per driver per day, the share of days the list is photographed and the share of cells with a handwritten number. Every money figure follows from those. Every one is editable.
Derived from the four numbers above
Assumptions — click to see and edit each one
The list
Reading
Money
This is an estimate, not a promise. Try two corners: set the AI's certain share to 0 — the benefit barely moves; set the share a person checks to 0 — it falls to almost nothing. Set the share of days photographed to 0: every line is zero, cost included. An honest model has to be able to do that.
Every morning a driver gets a printout from the system: their own delivery list for the day — a row per store, a column per item, the quantity in each cell. By the end of the round the sheet is full of pen: a cell crossed out because the store did not take it, a 12 corrected to 10, a few circles as reminders. In the afternoon the sheet is in a pocket, and the system still holds the morning's printed numbers.
The change in this piece is small: photograph that sheet once for the whole day, instead of not at all. The AI looks for hand marks on the printed grid, and the machine compares a handwritten number with the delivered quantity already entered. We measured the first two live days at a fresh-food distributor: 10 lists, 462 printed cells, 13 handwritten numbers. The machine read exactly 1 of those 13 with certainty; the other 12 were put in front of a person to check on the photo. That result decides where the value is — and it is not where people usually expect it.
1. What changes
- One list, one driver, one day. The delivery screen on the phone has a "Delivery list" card; the driver takes one or a few photos (six at most). The system already knows what is PRINTED on the list — that driver's own delivery notes for the day — so it sends that grid along to the AI. A photo of the list can also sit in a batch of the whole day's photos; the machine recognises it as the list and files it there.
- The AI does not re-read the sheet, it only looks for hand marks: struck out, corrected, added, circled, annotated — and says, for each mark, whether it is sure.
- The machine compares, by a written rule. Struck out means the paper says 0; a corrected or added number means the paper says that number. Only a mark the AI is certain of that differs from the delivered quantity becomes a "gap", reported to the driver on the phone. An uncertain mark is shown under the item with "check the photo", not reported. Circles and notes are only shown, not compared.
- A clean printed cell concludes nothing. The AI can miss a faint mark, so a cell with no hand mark is not counted as a match.
- When the driver corrects the delivered quantity on the phone, the machine compares again against the stored reading at once, without calling the AI.
2. What the first two days measured
| Item | Measured |
|---|---|
| Lists with a photo | 10 lists, 17 photos |
| Driver-days with a list photographed / driver-days with delivery notes | 10 / 29 |
| Delivery notes on those lists | 93 (9.3 per list) |
| Printed cells on the lists | 462 |
| Hand marks the AI found | 67 |
| — of which circled or annotated, no number | 54 |
| — of which handwritten numbers (struck out, corrected) | 13 |
| Handwritten numbers the AI read with certainty | 1 |
| Gap alerts raised | 0 |
Of the 13 handwritten numbers: 5 already matched the delivered quantity in the system — the driver had corrected it on the phone; 7 differed from it, but the AI marked all of them "unsure", so they were only shown for a person; 1 was written over and could not be read. Results came back within fourteen minutes of the first photo, usually in one or two.
It also has to be said plainly: in those two days nobody stopped photographing each note. 87 of the 93 notes on the lists still had their own photo of the signed delivery note, and that reader flagged a gap on 24 of them. The list is the driver's paper and carries no store signature; it is a cheap extra check, not yet something that replaces the signed sheet.
3. First stream: gaps caught before invoicing
A gap is caught one of two ways: a number the AI read with certainty that differs from the delivered quantity (the machine reports it), plus an uncertain number that differs and that a person opens the photo to check.
How to measure it. Take the last two month-end statements: how many mismatched lines trace back to something the driver struck out or corrected on the list, and how long each took to chase down. The "after" figure reads straight off the photographed lists: marks, marks a person checked, marks corrected.
How far to trust it: medium. All seven "differs" were numbers the AI was unsure of; some may be misreadings, not real gaps.
4. Second stream: goods kept while there is still time
How to measure it. At the next stock count, separate the variance of the items most often struck out or corrected on the lists, and compare with the two months before photographing began.
How far to trust it: low — this is the softest line, because the shrinkage share is an assumption.
5. What to subtract: a person checking photos, and the AI reading them
Checking uncertain numbers is real time: 148 numbers × 50% × 2 minutes ≈ VND 0.17 million/month. The AI reads each photo once, 1.7 photos a list, VND 800 a photo: ≈ VND 0.17 million/month. Correcting a number after the reading costs nothing more — the comparison re-runs in code.
6. Adding it up
| Stream | Per month | Per year |
|---|---|---|
| Fewer hours settling gaps | 0.60 M | 7.21 M |
| Goods kept while there is still time | 0.56 M | 6.74 M |
| A person checking photos | −0.17 M | −2.07 M |
| AI reading photos | −0.17 M | −2.08 M |
| Net benefit | 0.82 M | 9.80 M |
A small number, and genuinely small: only about 3% of the cells on a list carry a handwritten number. The treemap in the calculator shows why — almost the whole area is clean printed cells the machine concludes nothing about, and the tile the machine decides on alone is a sliver in the corner.
7. What this ROI stands on
Whether a person opens the photo. This is the most important finding of the two days. Set the AI's certain share to 0 in the calculator: the net benefit only drops from 0.82 to 0.72 million. Set the share a person checks to 0: it falls to almost nothing. The value of the list is that it puts exactly the cells with a pen mark in front of a person, not that the machine rules on them.
Whether "differs" really differs. 54% is the share among the numbers the AI read out, uncertain ones included. If you think half are misreadings, set 25%: the net benefit falls to about 0.19 million.
A clean cell does not mean a correct one. The system deliberately does not count a clean printed cell as a match, because the AI can miss a mark. Do not read "no gap alerts" as "no gaps".
The list carries no store signature. It is the driver's own account. Where the store signs, the signed sheet is still the evidence; the list only stands in for it on notes with no photo of a signed sheet at all — six of 93 in this measurement.
And as with every calculator on this site: set the share of days photographed to 0 and every line is zero, cost included.
8. Measure it in your own company
- Take a baseline. For one week, collect every driver's paper list: count cells with a handwritten number, and how many of those had not been entered in the system when the invoice was made.
- Photograph for four weeks. Track weekly: share of days photographed, hand marks, share the AI read with certainty, share of uncertain numbers a person opened the photo for.
- Compare the month-end statements before and after: the mismatched lines that trace back to a mark on the list are the part this photo has to catch.
9. The thing to remember
One photo for the whole day is cheap enough to be worth taking — but what it buys is a person's attention, not a machine's verdict. Of 462 printed cells, the machine was sure of one number. The rest is a short list of cells with a pen mark, put in the right place, telling a person where to look.
The sheet in the driver's pocket already holds the answer. The photo's job is to bring it back before the invoice.
Measurements are from the first two live days at a food distributor, anonymised; the sample is still small. Goods cost, minutes, shrinkage share and every other assumption in the calculator are illustrative. Replace them with your own.
Read next
-
A G8 file from the work items you already have: how many hours of re-typing it saves, and how many copying typos stop reaching the estimate
Quantities worked out in your own sheet, then every code and every explanation line typed again into the G8 estimating program: a few hours a month, and a few copying typos along with them. When an assistant writes the work items as one file in the import layout of G8 2020's «Công trình» sheet — explanation lines as formulas, group and section total rows — how many hours come back, how many typos go away, and how much AI cost comes off?
-
Never invoice twice: ask the accounting ledger before creating — 199 of 209 invoices were already there
When invoicing moves from the accounting package into the ERP, there is a stretch when both create invoices. If the ERP asks the accounting ledger before it creates one, how many duplicates does a cutover avoid — and what is each duplicate you never clean up worth?
-
Sharing an estimate by one link: what the mailed-Excel loop costs, and "which version is the latest?"
An estimate passes through 2–4 people as Excel files sent by mail, so someone always asks "which version is the latest?". When the whole set — drawings, quantity table, where every number comes from — sits behind one shared project, with people invited by email or sent a view-only link that needs no account, how many hours a month does the review loop get back?