Manpower Ledger Book a demo

Can AI Read a Handwritten Timesheet? What We Measured

On our sample sheets, yes. Well enough to stop typing, and not well enough to stop checking. Here are our own figures, the mistakes it makes, and the three checks that catch them.

In short
  • In our tests on sample sheets, AI read 99 boxes in 100 correctly. The samples were neat, and a real supervisor's writing will be harder.
  • One wrong box in a hundred is still several wrong boxes on a monthly sheet, so the reading must never go straight into a salary.
  • Reading every sheet twice and comparing the two readings cut the mistakes that slipped through from about one a sheet to one in five sheets.
  • A clear photo matters more than anything else. Flat, whole, in good light, written in pen.

The short answer

A site supervisor fills in a timesheet by hand. Somebody in the office then types 30 workers times 31 days into a spreadsheet. That is close to a thousand boxes for one site, typed by a person who is reading someone else's handwriting.

AI can now do that typing. The new models that read pictures do not look at one character at a time, as the old OCR programs did. They read the sheet as a table, the way a person does: this row is Imran, this column is the fourth, the box says 8.

They are also wrong sometimes. The whole question is how often, and what you do about it.

What we measured

We do not trust a figure that comes without its test, so here is ours.

The sheets Three sample timesheets in the layout manpower offices use, with 10, 13 and 40 workers and a full month of hours. We made them on a computer, in a handwriting-style script
How they were given Each as a clean scan, and as a picture turned and blurred a little, as a phone photo would be
Readings 18 in all
Boxes read correctly 99.3 in every 100
The same sheets with an older AI model 74 in every 100
Measured November 2026

Two things are worth seeing in that table. The first is that 99 in 100 is good. The second is the last row: on the same sheets, another model got one box in four wrong. "AI" is not one thing. A tool that reads handwriting well this year may be built on a model that could not do it last year, and a seller's "99.9% accurate" means nothing until you see the sheets it was measured on.

Our limits, plainly. The sheets were made by us, with made-up workers, and the writing on them is a neat handwriting-style script, not a real person's hand. No sheet from a real supervisor has been measured yet. A supervisor writing on a clipboard in August is harder, so expect a real sheet to do worse than our figure. The checks below exist because of that, and we will publish the real figure when we have it.

What it gets wrong

A monthly sheet for 40 workers has about 1,200 boxes. At 99 in 100, that is around ten wrong boxes. They are not spread at random. They gather in a few places.

Name1234567Total
Suresh B.998898951
Anil T.8A889841
The marked box is the hard one: a 9 was crossed out and an 8 written beside it. A person hesitates there too.
  • A figure written over another, or crossed out. The most common cause by far.
  • Figures that look alike. 1 and 7, 4 and 9, 3 and 8, and a quick A that looks like a 4.
  • An empty box against a dash. Is it a rest day, or did the supervisor mean zero?
  • Boxes that slip along a row. This was the reader's rare mistake in our tests: a few figures of one worker moved one day to the right. It is the most dangerous kind, because the row still adds up to the right total.
  • A bad photo. A shadow across the sheet, a folded corner, the last column cut off.

The three checks that catch it

A reading that is right 99 times in 100 is useful only if you know which box is the hundredth. Nobody can check 1,200 boxes by eye every month: that is the typing job again. The checks have to point at the few boxes worth looking at.

1. Read it twice and compare. The same sheet is read two separate times. Where both readings give the same figure, the box is very probably right. Where they differ, the box is marked. In our tests a single reading let about 1.2 wrong boxes per sheet through the other checks. Reading twice brought that down to about 0.2, one wrong box in every five sheets.

2. Add up the row. The supervisor wrote a total at the end of each row. If the boxes that were read do not add up to it, something in the row is wrong: a box, or the supervisor's own sum. Either way a person should look. This check cannot see figures that slipped along a row, because the sum stays the same. That is what the second reading is for: two readings almost never slip in the same place.

3. A person approves. The reading sits beside the picture of the sheet. One of your staff looks at the marked boxes, corrects them and approves. Nothing counts for a salary before that.

How to take a photo that reads well

The cheapest way to better readings is a better picture. Give your supervisors these six lines.

  1. Lay the sheet flat on a table. Do not hold it in your hand.
  2. Stand over it, so the camera looks straight down.
  3. Get the whole sheet into the picture, with all four corners.
  4. Use daylight or a bright room, with no shadow of your hand or phone on the paper.
  5. Write in blue or black pen, never in pencil.
  6. Cross a mistake out and write the right figure beside it. Do not write over a figure.

And one for the form itself: keep the column of row totals. It costs the supervisor a minute and it is the best check there is.

Does it replace the person who types?

It replaces the typing. Somebody in the office still looks at the marked boxes and approves the sheet, and that person now spends minutes on a sheet where it was an hour. One honest detail: a reading is not instant. A full monthly sheet takes a minute or two to be read twice. The difference is that nobody has to sit there while it happens.

Questions people ask

Can AI read handwriting on a timesheet?

Current AI models read figures in a table well. On our sample sheets, which were in a neat handwriting-style script, 99 boxes in 100 were right. Real handwriting will do worse. Either way the rest have to be caught by checks and by a person, which is why a reading must be checked before it is used.

Is this the same as OCR?

Not quite. Classic OCR recognises printed characters one at a time and does badly with handwriting and with tables. The newer AI models read the whole picture, including which row and column a figure belongs to. In daily talk people call both "OCR".

What happens when the AI reads a number wrong?

If the two readings differ, the box is marked. If the row no longer adds up to the supervisor's total, the row is marked. Your staff correct it against the picture of the sheet. A mistake that passes all of that is rare, and it can still be found later, because the original picture stays with the timesheet.

Do our supervisors need a new form?

No. The reading works from the sheet they use today: a row for each worker and a column for each day. If you are designing a new one anyway, this format reads well.

Can it read a timesheet sent on WhatsApp?

It reads the picture, wherever the picture came from. Save the photo and upload it. Messaging apps often shrink pictures, so ask for the photo to be sent as a file, or by email, when the writing is small.

See it with your own timesheet

Send us one sheet from one of your sites. In a short call we show you your own hours, read and worked out into salaries.

Book a demo