We built software because the paperwork was winning

SafeDesk did not begin with an idea for artificial intelligence.

It began with a folder.

Actually, with many folders.

Anyone who has spent enough time around construction sites knows what I mean. Every company arrives with its own documents. Safety plans, registrations, certificates, training records, medical fitness documents, equipment documentation, insurance, appointments, expiry dates. Then workers change, subcontractors arrive, certificates expire, new documents replace old ones, and the same information appears again in another PDF, another email, another folder.

None of this is secondary. Construction is an environment where mistakes can have real consequences, and documentation exists for a reason. Behind a training certificate there is a worker who is supposed to understand a risk. Behind a medical fitness document there is a person who has been considered suitable for a particular activity. Behind a safety plan there are choices that should affect what happens on site.

The problem is not that these documents are useless. The problem is what we have gradually asked people to do with them.

In Italy, a significant part of this responsibility often ends up in the hands of engineers, architects, safety coordinators, RSPPs and technical offices. People who have studied for years to understand structures, construction processes, regulations and risk are asked to spend an extraordinary amount of time performing work that is essentially administrative.

Open the file. Find the date. Find the company. Check the worker. Look for the training certificate. Check whether it has expired. Go back to the spreadsheet. Update the row. Open another PDF. Repeat.

I have done this work myself.

That is probably the most important fact behind SafeDesk. I did not discover this problem by interviewing construction professionals and asking them what software they would like to buy. I was one of the professionals opening those files.

At some point I remember looking at the work in front of me and thinking that something was fundamentally wrong with the way we had organised it.

Not because the control was unnecessary. The control was essential. What felt wrong was that so much professional attention was being consumed by the mechanics of finding information rather than by understanding what that information meant.

There is an important difference between the two.

Finding an expiry date is not professional judgement. Understanding the consequences of an expired qualification on the activities taking place tomorrow morning is.

Recognising that a document is missing is not professional judgement. Understanding whether that absence represents a formal irregularity, an operational problem or a real safety risk is.

Reading the name of a worker from a certificate is not professional judgement. Knowing whether that worker should be doing a particular task, under those conditions, with that equipment, is.

Yet we had created processes in which highly qualified professionals were spending enormous amounts of time on the first category of work and, consequently, had less time available for the second.

That contradiction stayed with me.

I have always had a complicated relationship with repetitive work. Long before SafeDesk, and long before I thought of myself as someone who might build a technology company, I was interested in programming. When I encountered a repetitive calculation or a process that could be formalised, my instinct was often to build a small tool for it. Sometimes it was a spreadsheet. Sometimes a piece of software. Sometimes simply a better way of organising the information.

Engineering encourages this kind of thinking. You learn very quickly that repeating a calculation one hundred times does not make you a better engineer. If the calculation can be expressed correctly as a process, the intelligent thing to do is to automate the repetition and spend your time understanding the result.

For years, however, I thought about this mainly in relation to calculations. What changed was realising that the same principle applies to information.

A large part of technical work is not difficult because each individual operation is intellectually complex. It is difficult because thousands of small operations have to be performed correctly, continuously and without losing context.

Construction documentation is a perfect example.

A person can easily read a certificate. A person can easily find an expiry date. A person can easily identify the name of a company. The problem begins when that person has to do it hundreds of times, across dozens of companies, while documents are constantly arriving and changing.

Human beings are very good at judgement. We are much less impressive at repeating the same mechanical operation indefinitely without making mistakes. Computers have almost exactly the opposite characteristics.

Once you look at the problem from that perspective, the question changes.

It is no longer, “Can artificial intelligence replace the safety coordinator?” That was never the question I was interested in.

The useful question is much simpler:

Which part of this work genuinely requires the safety coordinator?

That question became the foundation of SafeDesk.

If a machine can open a document, recognise what kind of document it is, identify the company, read the relevant information, extract a date, associate it with the correct subject and compare it with what is required, then there is very little value in asking a professional to perform those same operations manually.

The professional should receive the result and decide what it means.

That distinction is fundamental to the way I think about technology in engineering.

There is a tendency, particularly when new technologies appear, to frame everything in terms of replacement.

Will artificial intelligence replace engineers? Will software replace architects? Will automation eliminate this profession or that profession?

I find most of these questions less interesting than they appear.

Engineering has always evolved by moving certain tasks from human effort into tools. Nobody believes that using finite element software makes someone less of an engineer because they are no longer solving thousands of equations by hand. We do not ask a structural engineer to calculate every matrix manually in order to prove that the final judgement belongs to them.

The calculation is delegated to the machine. Responsibility is not.

The same distinction should exist elsewhere.

Artificial intelligence can read documents at a speed that no human being can reasonably match. It can classify information, compare records and identify inconsistencies. That is useful precisely because none of those activities is where the real professional value should be concentrated.

The real value begins afterwards.

Why is this document missing? Does it matter for the activity being performed? Can the company enter the site? Is the worker properly trained for tomorrow’s operation? Is there a pattern of non compliance that suggests a larger problem? Does what appears correct on paper correspond to what is actually happening on site?

These are different questions. They require context, experience and responsibility. They require a professional.

The more we developed SafeDesk, the more I realised that the interesting part of artificial intelligence in construction was not making the engineer disappear.

It was giving the engineer back the time to be an engineer.

That may sound like a small distinction, but I think it is an important one.

For years we have digitised processes without necessarily improving them. We took documents that were once printed and put them into PDFs. We took physical folders and replaced them with digital folders. We took paper registers and recreated them as spreadsheets.

That made information easier to store and transmit, but it did not fundamentally change the work.

A digital bureaucracy is still bureaucracy. If I have to open fifty PDFs instead of fifty paper documents, the storage medium has improved. The process has not.

Real digitalisation begins when we stop reproducing the old workflow on a screen and start asking whether every step of that workflow is still necessary.

This is where artificial intelligence becomes interesting. For the first time, software can work with information that was previously difficult to automate because it was written for humans rather than machines.

A PDF does not need to have perfectly structured fields. A certificate does not need to follow exactly the same layout as the previous certificate. A document can contain text, tables, stamps, signatures and information in different positions. The software can begin to understand the document as a document, rather than simply as a collection of predefined database fields.

For construction, this opens a much larger door than people sometimes realise.

Our industry produces an enormous amount of information. Safety documentation is only one part of it. Technical specifications, inspection reports, contracts, drawings, minutes, certificates, test reports, maintenance records, material documentation, site photographs and correspondence all contain pieces of the project.

The difficulty is rarely that the information does not exist. Very often, the difficulty is that nobody has enough time to read all of it, connect it and keep it continuously updated.

That is a very different problem from the ones software could solve ten years ago.

And it is one of the reasons I returned to programming with such interest after spending so many years primarily as a structural engineer. I suddenly found myself looking at the problems I had encountered throughout my professional life and seeing them differently.

Not every engineering problem requires another engineer. Sometimes it requires better software.

At the same time, not every problem that can be automated should be automated.

This is probably the part that matters most.

Technology is useful when we understand exactly what we want to remove from human work and, equally, what we want to protect.

I want to automate searching for an expiry date. I do not want to automate responsibility.

I want software to identify that something is missing. I do not want software to pretend that understanding the consequences of that absence requires no professional judgement.

I want a system to organise information better than a person can. I still want a person to decide what should happen because of that information.

The objective is not to remove people from construction.

The objective is to stop wasting them.

There are experienced engineers spending afternoons renaming files. There are safety coordinators moving dates from PDFs into spreadsheets. There are technical offices searching through email attachments for documents they know somebody sent three weeks ago. There are people whose knowledge took ten or fifteen years to build who are performing tasks that a machine can now execute in seconds.

That is not efficiency.

It is a poor allocation of intelligence.

SafeDesk came from trying to correct a very small part of that problem. It is still a young product, and there is much more to build. But the reason behind it has become clearer to me as we have worked on it.

The best technology does not begin with technology.

It begins with watching how people actually work.

You sit next to someone. You follow the process. You notice where they stop. You notice what they repeat. You notice the spreadsheet that everybody complains about but nobody removes. You notice the folder whose organisation only one person understands. You notice the task that takes three hours every Friday simply because it has always taken three hours every Friday.

Then you ask whether the problem is really inevitable.

Sometimes it is. Often it is not.

I think this is particularly important now, because artificial intelligence has made it very easy to begin with the solution. Every week there is another tool looking for a problem to solve.

For me, the more interesting direction is the opposite.

Start with the work. Start with the person doing it. Start with the annoying, repetitive, unglamorous part that nobody puts in a presentation. Understand why it exists. Then decide whether technology has anything useful to say.

SafeDesk did not begin because we wanted to build an artificial intelligence product.

It began because the paperwork was winning.

And after spending enough time on construction sites, I had become convinced that there were better things for engineers to lose their time on.