Can You Trust AI On A Fit-Out Project?

Not on its own. Use it to get a first draft quickly, check that draft the way you would check one from a new starter, and leave the signature with the person who is responsible for it. That is the short answer.

A lot of people in construction are sceptical about AI, and I think they are right to be. A tool can sound certain and be wrong. It can fill a gap with something that looks plausible. And when it goes wrong on a live project, the tool is not the one who answers for it. I am sceptical too, and I have built 140 tools for UK fit-out without writing a line of code by hand.

This article is what that taught me. It is not a pitch. It is what I check, what I will not hand over, and the questions I would ask of any tool before it goes near a live project, mine included.

Why the scepticism is reasonable

Three things worry people, and all three are real.

• •

  It can be wrong and sound right. A confident wrong answer is more dangerous than a blank page, because nobody stops to question it.

• •

  It fills gaps. Leave a detail out of what you give it, such as the lifting arrangements for a heavy item, and a careless tool will write something anyway. It will read well. It may not be true.

• •

  Nobody signs for it. Responsibility sits where it always did, with the person whose name is on the document.

None of that is a reason to ignore the technology. It is a description of what you have to check.

What building 140 tools taught me

The hard part was never the building. It was finding and fixing what was wrong.

Writing the tools was quick compared with checking them. I checked the results against what I know from site, and when a result did not match, that was the cue to go back and fix it, not a sign the tool had finished.

That changes where the work goes. There is less time typing the first version and more time reading it critically. If you do not have the experience to read a draft critically, a fast first draft is not a gain. It is a quicker way to be wrong.

What a tool should do when it does not know

The most dangerous failure is a gap filled quietly. A programme with an invented duration. A RAMS with an invented control measure. A record that looks complete and is not.

So the tools I built are set to write NOT PROVIDED BY AUTHOR where a fact is missing, rather than invent one. It looks untidy on a draft. I would rather the software told me what it does not know than handed me something that looks finished.

If a tool you are testing never says that it does not have something, ask why.

Where it earns its place, and where it does not

The work AI is good at is repetitive, document-heavy and checkable: a first programme, a first RAMS, a set of records, a first draft of a letter. You can put the output next to what you know and see quickly whether it holds.

The work it is poor at depends on being on site. Knowing which sub will turn up short, what the client will actually accept, whether the lift is really free on Thursday. That is judgement, and it comes from having been there.

I use it freely for the first kind of work and keep my own hands on the second.

Who signs it off

If your name is on a RAMS, you answer for what it says. That does not change because a tool drafted it.

The RAMS Reviewer is a good example. Upload a RAMS and it gives a suggested review against a 60-point checklist drawn from CDM 2015 and the Work at Height, COSHH and LOLER Regulations. It is a suggestion. The decision to accept it stays with you or your competent person.

That is how I wanted all 140 tools to work. Everything the platform produces is a draft. You review it and you stay the competent person.

It can still get things wrong, which is why every draft needs checking by someone with the experience to judge it.

What checking a draft actually means

Checking is not skimming. Take a RAMS draft. These are the questions I put to it:

• •

  Does it describe the job as it will actually be done on this site, or a generic version of it?

• •

  Are the hazards the ones this site has: the occupied floor, the out-of-hours access, the neighbouring tenant?

• •

  Are the controls ones the crew can carry out with the kit and the numbers they will have?

• •

  Are the sequence, the permits and the interfaces with other trades right?

• •

  Are the names, dates and references correct?

• •

  Is there anything in it that nobody on the team can explain?

The last question matters most. If nobody can explain a line, it should not stay in the document.

How I test a tool before I trust it

The same way I would test a new planner or a new coordinator. I run it on a realistic scope, compare it with what I would have produced, and fix what it misses.

1. 1.

   Give it a realistic scope, one you know well, with the awkward parts left in.

2. 2.

   Compare what comes back with what you would have produced yourself.

3. 3.

   Mark what it missed, what it got wrong, and anything it did that you had not thought of.

4. 4.

   Ask the real question: would I have produced this, and could I use it on site?

Take a programme. Built the old-fashioned way, from the brief and drawings, it takes me one to two days. A first draft from the tool usually takes a few minutes, and then I check it. The saving is real, and it is in the first draft. It is not in the checking.

Give the same scope to ten planners and you get ten different programmes. So a tool gives you a draft to work from, not the one right answer, and a programme that has not been checked by the person running the job is only a first draft, however quickly it appeared.

Seven questions to ask before you trust any AI tool on a project

These work on any tool, mine included.

1. 1.

   What does it do when a fact is missing: stop and say so, or fill the gap?

2. 2.

   Who checks the draft, and do they have the experience to judge it?

3. 3.

   Can you see what it used to reach the result: which documents, which inputs?

4. 4.

   Does it tell you what it is not sure about?

5. 5.

   Does it put anything in the document that you did not give it?

6. 6.

   Can you take the result into the formats and tools your team already uses, and change it?

7. 7.

   If it is wrong, who finds out, when, and what does it cost?

What I would never let AI sign off

My own line is anything that carries a name and a consequence. A RAMS. A payment or pay less notice. An extension of time submission. A certificate.

AI can draft any of those, and a draft saves real time. But the person responsible reads it as though they had written it, because as far as anyone else is concerned, they did.

The short version

Treat AI as a fast first draft. It is not a second opinion and it is not a signature. Test it on a scope you know. Prefer tools that tell you what they do not know. Keep the person who is responsible exactly where they are.

Scepticism about AI is healthy. The answer is not to refuse it. The answer is to check it, and fast is only useful if it is also checked.

The tools mentioned in this article are on the FitOut Insider platform.

Related reading