Three things I keep noticing, and one I have not solved.
Dear reader,
The most important evidence about a building is often the evidence nobody has collected yet.
It might be behind a wall. Inside a concrete core nobody has cut. Or waiting on the next hard rain.
That is a strange problem to have right now, because AI can read through years of records in the time it takes to make coffee, and not one of them contains evidence nobody went and collected.
I see my son’s school most mornings, the buildings and the structures across the campus. Part of it is being renovated and expanded right now. Sooner or later somebody decides: fix it, reuse it, add to it, or knock it down.
Twenty years of looking at buildings, structures, and the infrastructure around us has taught me they tell you things. A patch on a wall. A stain under a window. A seam where one pour of concrete met another that had already started to set. Those marks say something happened here. They do not say what. Three different problems can leave the same mark, and each one arrives with a different bill.
Ever had a report tell you everything except the one thing you needed?
Some of that history reaches the drawings, reports, photographs, and asset database. Some never does. The file holds a version of the building. Not the building.
Now AI reads the file and hands back an answer before you have finished reading the question.
Three things I keep noticing. One I have not solved.
There is a pitch going around. Connect everything. Every drawing, every work order, every sensor, every inspection photo. The fragmentation disappears and the answers appear.
Some of this is real. Platforms already pull condition, cost, maintenance, and risk data together to help owners prioritize maintenance and capital investment, and several now embed AI into that analysis.
Start with something simpler than a cracked wall. Does the record even know what is out there? Take a large campus. How many buildings. How many structures. How many light poles, how many sets of bleachers, and where each one sits. Much of the time that inventory gets built properly. Sometimes it does not, nobody notices, and then we point AI at it and ask which one to replace first.
Then there is the part the demo skips. If nobody opened the wall, there is nothing behind the wall in the database. Connecting ten incomplete records gives you one large incomplete record. It looks better on a dashboard.
AI can tell you what often happens to buildings like yours. But when the evidence was never collected, it cannot establish what happened to yours.
The answer is not always sitting in another system. Sometimes it is sitting in the concrete.
The question to ask before acting on any recommendation is not a clever one. What do we not know, and who is going to go find out?
That gap is what (AI + MI) × HI™ is for. Artificial Intelligence brings speed. Material Intelligence, the earned habit of knowing how buildings behave, names what the file left out. Human Intelligence decides whether those two add up to a good decision or a fast bad one.
AI works remarkably well at a great many things. It also fails, and the dangerous failure is when it fails well.
A bad answer gets questioned. A good-looking answer gets forwarded.
Ever sat in a meeting where a document was so well organized nobody argued with it? That is the failure mode. Formatting is not evidence. Confidence is not evidence. A tidy paragraph explaining why a wall is cracking is worth nothing until somebody checks whether that explanation is physically possible.
The argument I used to hear was about better prompting. What I hear now is connection: link every workflow, every system, every record, and the problem takes care of itself. That is the part to watch, because none of it tests whether the explanation could have happened.
Forensic practice, the work of figuring out why something cracked, leaked, or fell down, has used a less glamorous method for decades. List every explanation that could produce what you are seeing. Work out what each one predicts you would find somewhere else. Then go look.
Back to the stain under the window. It could be failed flashing, the metal that is supposed to catch water and send it back outside. It could be a blocked weep, the small opening that lets trapped water drain. It could be a detail that was wrong on the drawing and wrong in the wall. All three can explain the stain, and more than one can be at work at once. Each predicts a different pattern around it and a different response after three days of rain. The evidence, not the prose, separates them.
An explanation that reads beautifully and cannot physically happen is still wrong.
AI made the thinking faster. Nobody told the concrete.
Cutting a core out of a slab takes what it takes. Getting access. Taking the sample. Preparing it, putting it under a microscope, and working out what it is telling you about what that concrete has been through. None of that compresses to model speed. AI can answer in seconds. Next to seconds, weeks look like an excuse.
The same goes for opening up an assembly, or waiting out a wet season to see what a wall does when it rains for three days straight. The analysis compressed. The evidence did not.
Which creates a problem nobody writes down. Ever been the person asking for six more weeks in a room where a finished-looking answer arrived on Tuesday? You are not the careful one in that room. You are the holdup.
In my experience, nobody announces that the evidence is being skipped. It never reaches the agenda. The schedule tightens by a week, then another, and the investigation that would have changed the answer becomes the thing there was no time for. A year later it is in the capital plan and the wall has still never been opened.
Here is the one I have not worked out.
Say you did everything right. Real evidence, competing explanations tested, a proper investigation. The report is correct.
It is correct on the day you sign it.
Then the building keeps going. The use changes. Somebody makes a repair and tells nobody. A wet season shows up and reveals what three dry ones hid. The report did not become wrong. The building moved.
We are not oblivious to this. Inspection cycles exist. Capital plans get refreshed. But I have never seen anyone put an expiration date on a conclusion. Or a trigger. Something that says: if this happens, open this back up.
We are good at paying for answers. We are less good at paying to revisit them. AI can make that worse, because its output looks finished, and finished things get quoted for years by people who have no idea when it was written or what it was written from.
I have asked people who ought to know. Nobody has a good answer yet.
PostscriptThis argument is the basis of my SXSW 2027 PanelPicker proposal, Concrete Remembers: Rules for AI in the Physical World. Community voting is open through August 23, and voting requires a free SXSW account. Vote here.
Sources: IBM, Asset Investment Planning with IBM Maximo Application Suite. Siemens, Origin: AI-driven asset lifecycle and capital planning software.
(AI + MI) × HI™ is an original framework developed by Jigar B. Desai.