Everyone Records It Their Own Way

Reading four hundred delivery records to work out which supplier keeps missing its slot is real work that nobody on a live project has time for and software gets through it in seconds. It'll also spot that a trade has been waiting on the same predecessor for six weeks, or that the same reason code keeps landing on the same zone. Construction has needed that for years and what AI can do with a decent set of records is genuinely impressive.

The question is what we're giving it to work with.

What's actually in the record

On most projects a single day ends up spread across five or six places. The diary gets written from memory after the site closes by the person who spent that day dealing with the events in it. The dockets are in a folder in the site office and the photos are on somebody's phone with no labels on them. A fair bit of it is in a group chat somewhere between the parking and the coffee run. Whatever is left sits in what three or four people remember about a Tuesday that has already started to blur by Friday.

No AI on the market will know what time the truck arrived, since nobody wrote down what time the truck arrived. What comes back reads well and sounds certain, then in a few places it's wrong with nothing to mark out which places.

A person doing the same job hedges, so ask a site manager to reconstruct a delayed pour from three weeks ago and they'll tell you which bits they're sure of and which bits they're filling in. AI won't do that unless you ask it to and the answer looks the same whether the record underneath it was complete or nearly empty.

That matters more the further the answer travels, since somebody reading a summary in November can't tell which bits came off a real timestamp and which bits were worked out later. Both sorts end up in the same paragraph looking exactly alike.

Most sites already record plenty

The usual assumption is that construction doesn't write enough down and that recording more of it would finally give these tools something to work with.

Most projects already record a great deal and they just don't record it the same way, so very little of it can be put side by side.

Here's what that looks like on the page, where one site manager writes that the concrete didn't turn up and the pour moved to Thursday. On the next job the same event goes down as a Zone B pour not commenced, awaiting materials. A third site files it under weather, since it had also started raining by the time anyone gave up on it, so all three entries are honest accounts of the same afternoon and none of them match.

Now try to count how many pours were lost to late concrete last quarter. Two of those three entries never mention concrete and one of them blames the sky. Everyone involved is competent and everyone has their own way of working, which is entirely normal and makes the number impossible to arrive at.

The cost of that isn't the missing number, it's that the supplier never comes up. Nobody questions the concrete because on the one record that names a cause it was the weather, so the conversation that should happen at the next performance review never happens at all. Inconsistent recording ends up protecting whatever is actually going wrong, quietly, for years.

The same thing happens further up with reports, where four people write up the same problem and all four get to roughly the same conclusion. One does it in a spreadsheet, another in a slide deck, a third in an email with the actual finding buried in the third paragraph and the fourth in a two page memo. None of the information was missing and none of it lined up, so working out that the four accounts actually agreed took months when it should have taken days.

Put those four reports through AI and you get a tidier version of the same disagreement.

Manufacturing sorted this decades ago

In a factory the data gets captured all the way along the chain without stopping or leaving gaps. A stoppage gets logged against a code by the operator at the machine before the line starts again, so nobody writes it up at the end of the shift and nobody types a sentence about it. That code means the same thing in March as it does in November and the same thing again on the other line.

AI works well in a factory for that reason, since decades of fairly dull work went into getting the recording consistent and the clever software turned up afterwards to find the job already done.

The productivity gap between the two industries is well documented and it has plenty of causes. This is one of them and it doesn't get discussed much, since you can't buy your way out of it.

What it takes to close the gap

None of it is complicated, which is probably why it keeps getting skipped.

The event gets recorded when it happens by the person who was there, rather than at the end of the day from memory. The reason comes off a set list so a missed pour is Delivery Not Made or Access Not Available or Trade Not Attended rather than whatever phrasing came to mind at the time. That same list gets used on the next job so this month can be set against last month and this project against the one before it. The day gets locked at the end of it so what was recorded stays recorded.

If that sounds like a programme of work, it doesn't have to be. Pick the one thing that costs you most, agree six or seven reasons for it and nothing else, then use those same six or seven on the next job. One category recorded consistently is worth more than a whole diary written freely.

What you then do with it is a subject of its own and there's a longer argument about measurement elsewhere. Nobody has ever been promoted for writing things down properly and it decides whether everything built on top of it works.

The ceiling

None of this is an argument against AI. It's about what we're handing it. Construction is going to get some very good tools over the next few years and most of them will work as advertised, on records written from memory by four different people using four different words for the same afternoon. They'll keep getting better regardless of what the industry does about the records underneath them. How much those tools end up being worth comes down to what gets written down on site, every day, in the same words.

The ceiling isn't the technology, it's what we hand it.