Digitising handwritten site sheets – without abolishing the sheet

Why “end the paperwork” can be the wrong advice

Almost every construction software advertises abolishing paper: the foreman gets a tablet, the data is born digital, the Zettel disappears. For many trades that is right. In road construction and Tiefbau with public clients it falls short – for two reasons.

First, the signed Zettel is evidence. The crew leader’s signature attests that the work was done this way. A tablet form produces a data record; a signed sheet produces a document you can put in front of an auditor.

Second, nobody operates a tablet on site. Not out of reluctance, but because the conditions argue against it: rain, gloves, dust, one hand on the tool. The ballpoint pen has been winning that comparison for decades.

The more interesting question is therefore not how to get rid of the Zettel, but how to read it without retyping it.

What text recognition can do with handwriting today

With printed text, recognition is a solved problem. With handwriting it is not – and anyone promising otherwise either has very clean test data or is in the middle of selling something.

A realistic picture:

What matters here is not the hit rate but what happens to the uncertain values. A system that guesses and quietly carries on calculating is more dangerous in a world of evidence than one that asks.

The crux: crossed-out lines

On a real take-off sheet, things get corrected. A line is struck through and the right value is written beside it. To a person that is unambiguous. To a recognition method it is the hardest place on the whole sheet, because the crossed-out number is often still perfectly legible.

Anyone who only reads the text takes both values – the struck one and the valid one. The result is a quantity that is too high and an error that surfaces at the audit.

Hence a separation makes sense: one method captures the geometry of the sheet – where the fields are, which lines are struck through, which boxes are ticked. A second reads the writing in the fields that still apply. That division of labour is why a struck item is recognised as struck even when the number under the line is neatly written.

The approval step is not a blemish

At Levora every value that is read is put up for confirmation before it becomes billable – including the ones recognised with confidence. That sounds like a concession, but it is the actual point.

On a public contract your firm is liable for the quantity billed. Software that posts on its own does not shift that responsibility, it only obscures it. When the audit finds a discrepancy, pointing at a model helps nobody.

The practical benefit remains large all the same: instead of typing up eighty Zettel, you confirm eighty prepared entries and correct the few that are flagged.

What you have to prepare

Text recognition on forms works markedly better when the form is known. Three things help more than any model:

1. A fixed template. If your Zettel are always laid out the same way, the system can be calibrated to it. Change the template and it has to be set up once more.
2. Pre-printed fields instead of free text. A box for the quantity is easier to read than a number somewhere on the line.
3. A usable photo. Laid flat, evenly lit, the whole sheet in frame. That is the biggest lever and it costs nothing.

What changes in daily work – and what does not

Nothing changes on site. The Zettel is filled in and signed as before. That is the precondition for the change being accepted at all.

A lot changes in the office. Typing becomes checking. The difference is not only the time but the kind of error: typing produces transposed digits that nobody notices. Checking puts the original next to the value.

Nothing changes about your responsibility. The quantity that goes into the invoice is the one you approved.

Frequently asked

What is the hit rate? A general figure would be an advertising claim. It depends on your template, your handwriting and your photos. It only becomes meaningful on your own Zettel – which is why every demo should happen on those and not on sample data.

What happens to the paper? It stays. The photo hangs on the transaction, the original goes into the folder as before. On a query you see both side by side.

And data protection? A Zettel carries names and addresses. Processing therefore belongs in Germany, with a data processing agreement, and the contents must not be used to train someone else’s models. Ask about it – with every vendor.

Is it worth it if we only have twenty Zettel a month? Honestly: probably not. At twenty Zettel, capture costs a few hours a month, and no changeover recovers that quickly. In our experience the tipping point lies at around fifty to eighty documents – work it out on your own number before someone works it out for you.

Our crew fills in the Zettel in Polish or Turkish. Does that work? Figures and units are language-independent, and they are what is actually read. The same goes for free-text fields: they are a hint whatever language they are written in, not a data source. The one real dependency is item matching: the item number comes from the German bill of quantities, because that is the number that goes on the invoice. Which is why a separate field for the item number on the template beats writing text.

What a good Zettel template looks like

The biggest jump in quality comes not from the software but from the paper. Two templates for the same purpose, on the left the common one, on the right the readable one:

The last point is the most important and costs nothing: overwritten numbers are not reliably legible to a person or to a machine. A struck-through line with a new value beside it is legible to both.

Changing the template is an hour’s work and has more effect than any model improvement. Anyone thinking about digitisation should start here – even if no software follows at all.

The office check, step by step

This is the work that takes the place of typing:

1. Take the photo. Sheet flat, whole sheet in frame, even light.
2. Wait for the automatic capture. Seconds, not minutes.
3. Go through the green ones. Values recognised with confidence are displayed and are as a rule correct – visible all the same, not hidden.
4. Settle the flagged ones. Those are the few where recognition was uncertain. Photo and value stand side by side.
5. Approve. Only then does the quantity become billable.

The decisive difference from typing is not the time but the kind of error. Typing produces a transposed digit nobody notices, because there is no original to compare against. Checking puts the original right beside it.

What a capture error costs

A worked example that sorts the priorities. Suppose 128.50 m² becomes 12.85 m² while being typed:

The second point is the unpleasant one: a transposition downwards is not objected to by the client. It simply never comes to light.

What happens to the paper afterwards

The Zettel has been read in – and then? This question is rarely asked, and it decides whether digitisation actually saves work or merely moves it.

The original remains subject to retention as long as it is a supporting document. A take-off sheet underlying an invoice to a public client is part of the accounting records and falls under the commercial and tax retention periods.

In practice a routine in which both exist works well:

The decisive point: the photo does not automatically replace the paper. Whether purely digital retention is permitted depends on the requirements for proper bookkeeping – that is a question for your tax adviser, not your software vendor.

What digitisation does change immediately: you no longer have to open the folder in order to work. It becomes an archive instead of a working tool.

In summary

On a public contract the Zettel is not a relic but a piece of evidence. Abolishing it does not solve the problem, it relocates it. What actually saves time is the disappearance of the typing – with the routine on site unchanged, and with an approval step that leaves responsibility where it belongs.

What that looks like in practice is on the page about Zettel recognition .