Why construction is about to spend millions scaling a problem nobody has measured and what my book warned was coming.

I wrote a book about this. I called it The Utilization Deficit. The argument, in one sentence: every contractor in Canada pays for roughly 100 hours of labour to get roughly 55 hours of installed work, and nobody is measuring the other 45.

Last week, on a construction webinar hosted by On-Site Magazine, an insurance executive named David Bowcott summarized the entire thesis without meaning to. The panel featured Procore, CMiC, EllisDon, and PLATFORM Insurance. The topic was agentic AI digital employees, embedded in workflows, executing work instead of just summarizing it. And in the middle of a polished vendor conversation, Bowcott said this:

           “We don’t have a knowledge problem. We have a capacity and consistency problem.” Then everyone moved on.

They shouldn’t have. Because that sentence is the entire premise of my book and the entire reason most of the AI investments this industry is about to make are going to fail.

The sentence, decoded

“We don’t have a knowledge problem.” Translation: contractors are drowning in data. BIM, schedules, IoT, cost systems, safety reports, daily logs, drawings, RFIs, submittals. More information than any team can process.

“We have a capacity and consistency problem.” Translation: the people paid to do the work don’t have time to do the work. They’re chasing paper, reconciling systems, updating logs, producing reports.

That is the utilization deficit. Named by an insurance executive on a construction webinar, and nodded at by the rest of the panel as if it were wallpaper.

In the book, I call the 45% white space — the administrative, coordination, and consistency work absorbed into every paid hour that does not produce installed work. It is not fraud. It is not laziness. It is the carrying cost of running a project the way projects are run. It sits on every P&L. Nobody measures it. Nobody reports it. Nobody owns it.

And now we’re about to hand it a bigger budget.

What AI agents actually do

Procore has AI agents in open beta. CMiC is building them. EllisDon is already deploying them on Palantir Foundry. The use cases the panel listed — RFIs, submittals, daily logs, QA/QC, materials coordination, schedule reconciliation, compliance  are real. The agents work. The ROI stories are coming in.

Here’s the problem.

Every one of those use cases sits inside the 45%. Every one of them is white space  the administrative carrying cost between units of installed work. The vendors are selling machines that make the non-value-add activity faster, cheaper, and more consistent.

That is a real productivity gain. It is also a dangerous one.

Because if you do not measure the deficit first, you will use AI to industrialize it. That is the warning at the centre of my book, and the industry walked straight into it in real time on that panel.

The math nobody is running

The Utilization Deficit spends a full chapter on this, but here is the compressed version.

Say you run a $500M project with $200M of direct labour. On a 55/45 split, roughly $90M of that labour is carrying cost  paid hours not producing installed work. Some of that is unavoidable. A lot of it is not.

Now deploy AI agents across RFIs, submittals, logs, and reconciliation. You speed up the administrative layer by 30%. That feels like a win. Your dashboards show it. Your CFO is happy.

But what actually happened? You made your waste layer more efficient. The carrying cost is still there. The 45% is still 45%. You have not reclaimed a single hour of installed work. You have automated the symptoms of a problem you never diagnosed.

Worse: the vendors now own the record. Your workflows, your bottlenecks, your exceptions all flowing through their platform, priced on consumption. Your carrying cost is now a line item on someone else’s invoice.

What the leaders are actually doing

Read the EllisDon story carefully. They didn’t buy agents. They spent years building a

unified data ontology on Palantir Foundry. They structured safety, schedule, estimating, cost, logistics, and field data into a single operational view. Then they pointed agents at it.

That sequence matters. The foundation came first. The automation came second. And the foundation is the expensive part  years of structuring work, not a SaaS licence.

Most contractors do not have that foundation. They have systems. They do not have structure. They are about to buy agents and point them at fragmented, inconsistent, untagged data, and the agents will cheerfully automate the mess.

This is the gap. This is the sequence problem. And it is exactly what the book was written to prevent.

The order of operations

Before you buy an agent, you need three numbers. The book devotes its second half to building them.

Your utilization baseline. What percentage of paid hours produce installed work? What percentage is carrying cost? By trade, by crew, by phase, by project. Not an industry benchmark. Yours.

Your white space map. Where is the 45% sitting? Which processes, which handoffs, which meetings, which approvals, which waiting? Not guesses. Measured.

Your leverage points. Which slices of the 45% are the largest, most repeatable, and most automated ? That is where agents earn ROI. Everywhere else, they are theatre.

Without these numbers, an agent is a tool in search of a problem. With them, it is a precision instrument.

The vendors will not give you these numbers. The vendors are not incentivized to give you these numbers. Their business model is consumption. The more of the 45% that flows through their platform, the better their quarter.

The uncomfortable conclusion

AI agents are not the problem. They are a real and valuable technology, and the contractors who deploy them well will outperform the ones who don’t. That part is true.

But deploying them well requires knowing what you’re automating. And the industry is about to skip that step at scale which is why I wrote the book now, and not in five years when the cheques have already cleared.

The utilization deficit is the question every AI investment should answer before the cheque

is signed. How much of the 45% will this recover? How will we measure it? What’s our baseline? If the vendor cannot answer those questions, you are not buying productivity. You are buying velocity on waste.

Measure first. Automate second.

In that order. Always.

The Utilization Deficit is available on Amazon in Kindle, paperback, and hardcover.