Give your team easy access to AI and they will build their own tools, fast. Dustin Burns, who has run technology at McCownGordon Construction since 2005, calls the coming pile-up “AI sprawl.” His fix is not to lock it down. It is to govern the on-ramp: let people experiment, watch what they build, and step in before a weekend agent turns into leaked client data or a runaway bill. Here is his playbook, and why it matters whether you run a contractor or sell into one.
Something shifted this year on job sites and in back offices. A superintendent can talk to an app instead of typing. An estimator can spin up a custom agent over the weekend. A project manager can drop drawings into a model and get a number back. The tools got cheap, fast, and easy, and that is exactly the problem.
Dustin Burns, senior vice president of technology and business solutions at McCownGordon Construction, has watched construction tech go from “unnecessary” to unavoidable across two decades at the same firm. His warning on a recent Bricks & Bytes episode is not that AI is overhyped. It is that most companies are about to make a mess with it, and the ones who plan for the cleanup now will come out ahead.
The risk with AI right now is not that your people use it. It is that hundreds of small, unmanaged experiments scatter your data, blow your budget, and leave you with tools nobody can vouch for. Governance is what turns that chaos into capability.
AI sprawl is the next mess construction has to clean up
A pattern this tech leader has watched before
Hand your people free access to AI and something predictable happens. Someone builds a little tool over the weekend to solve a real headache. It works. They share it. Ten more sprout. Before long your data is scattered across a dozen silos and nobody quite knows what is running where. Burns has a name for it: AI sprawl.
He has seen the shape of this before. Around 2010 the iPad landed, and teams held monthly meetings just to compare which new apps everyone was using. That energy was good, but it splintered company data into disconnected pockets. The industry eventually pulled all of it back into common data environments, the single sources of truth most large contractors run today.
Burns expects agents to run the same loop: a burst of building, then a moment where leadership looks up and finds tools everywhere. His read stays calm. The explosion is a healthy phase to move through. The job is making sure it happens on rails.
Blocking AI backfires harder than you think
Govern the on-ramp, do not slam the door
The instinct at a risk-averse firm is to block anything unsanctioned. Burns argues that is the surest way to stifle innovation. Cut off the tools people are curious about and you do not stop them, you push the activity into the shadows. Worse, if your own approved AI becomes the thing nobody is allowed to touch, you have made it less useful than a banned app.
So he does the opposite. He does not block the other tools, he governs access to them. His team watches the citizen developers build custom agents inside Copilot, monitors the interactions on the back end, and lets experimentation run. The guardrail kicks in at one moment: when someone starts sharing an agent. That is when the technology group checks the structure and instructions, tightens them, and only then pushes the tool out more widely.
Underneath all of it sits a human. Burns describes catching a case where a staffer leaned on raw ChatGPT figures that could have carried real weight on a decision. A person in the loop caught it before anything left the four walls. That posture, where AI handles the task but a human still owns the outcome, is the same shape other operators are landing on. We have written about enterprise buyers standing up AI subcommittees and audit-then-assist-then-autonomy on-ramps to ease conservative teams into trusting an agent.
Data governance beats buying another AI tool
The quiet way AI budgets get wasted
Ask Burns for the single biggest mistake contractors make with AI and he does not hesitate: overinvesting in AI while underinvesting in data governance. The unglamorous work of quality, control, verification, and structure is what makes any of the flashy stuff function. Data only becomes useful once it is refined, and it is far too easy to bolt AI onto a broken process and call it progress.
There is a hard-dollar angle too. On tools that bill by tokens, one poorly formed request can burn through a thousand dollars in a single shot. Part of why Burns leans on predictable-billing tools while people learn is that a curious employee can question them all day without triggering a nasty overrun. A new model will not rescue messy inputs, so the firms getting real value treat their data as the asset and the AI as the layer on top, not the other way around.
The operators actually shipping AI read this first
One sharp briefing a week on how construction leaders are governing, buying, and deploying AI that works.
Join 3000+ ReadersWhat it looks like when the guardrails work
The payoff: capability without the chaos
Governance sounds like a tax until you see what it buys. By Burns’s own account, 72% of McCownGordon uses AI weekly, up sharply over a few months. A recent four-week stretch ran close to a thousand AI-assisted work hours through Copilot alone, before counting the other tools in play. In preconstruction, AI-based takeoff has cut 60 to 70% of the clicking, which he ties to a few hundred thousand dollars in quantified time savings.
On the IT side the firm runs full agent orchestration on its data pipelines: one agent plans the work, others do it and verify it. That maturity did not come from buying the best model. It came from years of getting the data house in order first, then layering automation on workflows that were already clean.
The last mile is culture. Getting one team’s win to travel across the whole company is mostly change management, the same wall large contractors keep hitting. It is worth studying how giants like Vinci and Kajima actually scale a pilot: prove it on one job, embed it in the standard, and let peers sell peers.
The bill nobody has priced yet: AI governance insurance
When the agent makes the call, who carries the risk?
Burns points to a question the industry has not answered. Once you hand agentic AI the keys to make decisions on your behalf, liability gets blurry, and construction is about as litigious as industries come. If a company rolls out an AI solution and a worker says the tool told them to do it, where does responsibility actually land?
His half-serious prediction is that AI governance insurance could become a line item on contracts, the way subcontractor default insurance (SDI) did as a way to cover the risk of a partner failing mid-project. Whether or not that specific product shows up, the underlying point is sharp. Automated decisions create new exposure, and somebody has to own it.
The honest answer today is that nobody has this fully worked out. That is not a reason to wait. The governance you build now, the monitoring, the review step, the human who owns the outcome, is what protects you when the rules catch up.
Three ways AI shows up on your team
Match the tool to the risk, not the hype
Not every AI tool carries the same exposure. The way Burns talks about it, the smart move is to know which lane you are in before you scale anything.
| Approach | Cost exposure | Data and governance risk | Best used for |
|---|---|---|---|
| Managed copilot (e.g. Microsoft Copilot) | Predictable, closer to flat | Lower, monitored on the back end | Broad rollout and open experimentation |
| Employee-built custom agents | Low to build | Medium, needs review before it is shared | Targeted use cases and citizen developers |
| Raw LLM API wrappers | Variable, token-metered (a malformed call can cost around $1,000) | Higher, easy to over-permission | Prototypes, not production without scrutiny |
It is the term Dustin Burns uses for the coming pile-up of ungoverned AI tools inside a company. When everyone has easy access, people build small agents and apps for single use cases, and the result is scattered data and tools nobody has vetted. Burns compares it to the app explosion after the iPad arrived around 2010, which the industry later had to consolidate back into common data environments. He expects AI and agents to repeat that loop. (Source)
Burns says no. His view is that blocking access is the fastest way to kill innovation, and that making your own approved AI the thing nobody can touch is worse than banning something like TikTok. Instead he governs access: let people experiment, monitor the activity, and protect corporate and client data without shutting the whole thing down. (Source)
Overinvesting in AI while underinvesting in data governance, according to Burns. He means the quality, control, verification, and structure of your data. Since data has to be refined before it is useful, bolting AI onto a messy process just produces confident nonsense faster. The unglamorous groundwork is what makes the tools pay off. (Source)
Burns describes a citizen-developer approach. Staff build custom agents inside Copilot, and the technology team monitors those interactions on the back end. The review step triggers when someone starts sharing an agent more widely: the team checks the structure and instructions, optimizes them, and only then pushes the tool out to the rest of the organization. A human stays responsible for the outcome throughout. (Source)
Yes, on tools that charge by tokens. A single poorly formed request can burn through around a thousand dollars, which is why Burns favors predictable-billing tools while people are still learning. With a flatter cost model, a curious employee can question the tool repeatedly without triggering an overrun. Cost control is part of governance, not separate from it. (Source)
It is a forward-looking idea Burns raises. As agentic AI starts making decisions on a company’s behalf, liability gets murky in a litigious industry. He wonders whether firms will need an AI governance insurance program written into contracts, similar to how subcontractor default insurance covers the risk of a partner failing. Nobody has this fully solved yet, which is why the monitoring and review you build now matters. (Source)
Related Articles
You Can Only Go As Fast As Your Customers: The Quiet Brake on ConTech AI
The Construction Giants’ Buying Playbook, From Inside Vinci and Kajima
