AI Agents for Construction Coordination

A note for the people who own the P&L on construction work, not the people who swing the hammers. Field productivity is a real problem, but it is not your problem to solve this decade. Your problem is that you run a coordination business that still prices coordination as overhead, and the cost structure of coordination is about to change underneath you.

What a general contractor actually sells

Strip away the branding and a GC sells exactly one thing: the promise that several dozen independent companies, none of which report to you and several of which actively dislike each other, will show up in the right order, with the right materials, and produce a building that matches a set of documents produced by yet another independent company. The self-perform work, where it exists, is a rounding error on this promise. The promise is sequencing, information flow, and accountability. The promise is coordination.

Owners understand this, which is why fee percentages have been compressed to the low single digits for decades. You are not paid for laying block. You are paid a thin sliver for making sure the block, the blocklayers, the rebar, the inspection, and the drawings all intersect at the same point in spacetime. When they do, you keep your two or three percent. When they don't, the slippage comes out of that same two or three percent, which is why a project can be ninety-five percent well-run and still lose money.

Now ask an uncomfortable question. If coordination is the product, how much of your organization is actually built to do it well, and how much is built to absorb its failures?

Walk through your own org chart with fresh eyes. Project engineers who spend their days formatting RFIs and chasing submittal returns. Project managers whose calendars are wall-to-wall status meetings that exist because status does not flow on its own. Superintendents who spend, by their own estimates, a third of the day relaying information between the field and the office. Assistant PMs hired substantially to send follow-up emails to subs about insurance certificates, closeout documents, and unreturned buyout paperwork. Precon teams re-typing scope clarifications across bid leveling sheets.

None of these people were hired to do those things. All of them do those things most of the day. The industry's answer to coordination load has always been the same: add another coordinator. It is the only industry I can think of where the standard fix for too much overhead is more overhead, and it works, in the sense that projects get delivered, and it fails, in the sense that the margin stays pinned to the floor.

Why software bounced off

Construction tech was supposed to fix this, and executives who lived through the last fifteen years of it have earned their skepticism. The pattern repeated with remarkable consistency. A point solution arrives: document management, then punch lists, then daily reports, then submittal logs, then drawing viewers, then reality capture. Each one digitizes a record that used to be paper. Each one gets adopted, sort of. And each one adds another place where information lives without adding any mechanism for information to move.

The result is the modern jobsite information stack: the truth about any given project is distributed across a project management platform, an accounting system, a scheduling tool, two or three sub-specific portals, a shared drive, a group text, and the personal inboxes of about fifteen people. Every piece of data is stored better than it was in 1995. Moving a question from the person who has it to the person who can answer it still requires a human being to notice, format, route, follow up, and relay back. The systems remember. They do not act.

The systems remember. They do not act.

This is the specific gap that matters, because a new kind of participant has started to close it, and the early adopters are not startups building demo houses. They are ordinary GCs and developers running ordinary jobs who have added AI agents to their project teams the way they would add a sharp coordinator, except the coordinator reads every log, never sleeps, and works inside the channels where the team already talks.

What mixed crews look like in the office trailer

Concreteness matters here, because "AI on the jobsite" has been vaporware for so long that abstraction reads as fiction. So here is what the pattern actually looks like on teams running it today, described as the industry trend it is becoming rather than any one company's setup.

The submittal chain is the cleanest example. A mechanical sub uploads a fan schedule submittal. An agent does the first pass: checks it against the spec section, flags that two units reference a model number discontinued in the latest catalog, and drafts the transmittal. A second agent, or the same one wearing a different hat, checks the flagged items against the approved equipment list from design development. Only then does it land on the project engineer's desk, pre-reviewed, discrepancies highlighted, ready for the human judgment call about whether the substitution is acceptable. The engineer makes the call in ten minutes instead of finding the discrepancy on his own in forty, or worse, not finding it. Then the package moves to the architect with the review history attached. Human sign-off never left the chain. The clerical layers around the sign-off did.

Insurance and compliance chasing, the assistant PM's purgatory, has quietly become agent work almost everywhere it has been tried. The agent watches expiration dates across every sub on the job, requests updated certificates directly from the sub's office, parses what comes back, and confirms coverage matches contract requirements. When a sub goes silent for a week, the agent escalates to the human PM with the history attached: requested twice, no response, coverage lapses Friday, this sub is on the critical path Monday. The PM makes one phone call with full context instead of discovering the lapse during an incident.

Status has stopped being a meeting on these teams. A PM types a question into the job channel at 9 pm, is the level four inspection scheduled and does anything block it, and gets an answer assembled from the inspection log, the lookahead, and this morning's daily report. Owners' reps have started doing the same thing, which has an interesting second-order effect: the monthly OAC meeting gets shorter and more substantive, because nobody needs the first forty minutes of readouts.

And then there is the pattern that sounds most like science fiction and is actually the most mundane: agents talking to each other across company lines. The GC's agent notices the revised structural drawings shifted an embed layout. It queries the steel fabricator's agent about whether the change affects pieces already in fabrication. The answer comes back in minutes: two pieces affected, not yet cut, revised shop drawings to follow. Two humans then approve what their agents worked out. The same conversation, run through traditional channels, is an RFI, a two-week cycle, and a decent chance the pieces get cut wrong in the meantime because nobody connected the drawing revision to the fabrication queue in time.

Notice the shape of every one of these scenes. The agent gathers, checks, drafts, watches, nudges, relays. The human decides. When the stakes rise, the agent's job is to escalate early with the facts assembled, not to improvise. Teams that get this division of labor right describe the effect less as automation and more as finally having enough coordinators, except the coordinators cost a rounding error and the humans they support get promoted into the judgment work they were hired for.

The math that should keep you up

Run the numbers on your own overhead. A typical commercial GC carries project management staff at somewhere between four and eight percent of contract value, depending on project type. Sit with your ops leaders and honestly estimate what fraction of those hours goes to the fetch-format-route-remind-relay layer. Most executives who do this exercise land somewhere between a third and a half, and the ones who land lower usually change their answer after a week of actually watching.

Now assume, conservatively, that half of that layer becomes agent work over the next few years, because at other firms it will. You are looking at one of two futures. In the first, you convert that capacity into margin and into the coordination-heavy work you currently ration: real lookahead planning, real buyout diligence, real risk reviews, precon depth that wins better work. Your fee stays at three percent and your cost to earn it drops. In the second future, your competitors do this and you don't, and since construction is a bid business, their new cost structure becomes the market price, and your three percent quietly becomes one and a half.

And the effect will not stop at your own overhead line, because your subs are running the same math on theirs. A mechanical contractor whose office runs this way returns submittals faster, answers RFIs faster, and staffs its PM desk leaner, which shows up in its numbers, which shows up in its bids. Within a few cycles the coordination speed of your trade partners becomes a real selection criterion, the way safety scores became one, and the GCs who learned to work in mixed human-and-agent project teams first will be the ones the good subs prefer to work with, for the simple reason that fast counterparties make everyone's job cheaper.

There is no third future where nothing changes. The technology is not waiting on a breakthrough. It is waiting on adoption, and adoption in construction follows a known curve: slow, skeptical, then sudden, the moment the first regional competitor starts winning bids with it.

What to actually do, and what to refuse to do

The executives getting this right are conspicuously not launching innovation initiatives. The failed playbook of the last decade, pilot team, lighthouse project, eighteen-month rollout, dies here too. What works is duller. Pick one live project with a PM who is respected and slightly impatient. Put agents on the three chores everyone already agrees are chores: submittal first-pass, compliance chasing, status in the channel. Keep human sign-off exactly where it is today, and write down, explicitly, where agents must escalate to a person, because your subs and your owner will ask, and "we haven't decided" is the answer that kills trust. Then let the PM tell the other PMs, because in this industry the only technology endorsement that has ever mattered is one super telling another it isn't garbage.

The refusal matters as much as the action. Refuse to frame this as headcount reduction, because the firms that do will teach their best people to sabotage it, and because it is inaccurate. Every GC in the country is short of good project people. The point was never fewer coordinators. The point is that your project engineers stop being routers and start being engineers, your supers get their walk back, and the coordination your fee actually pays for gets done by something with perfect recall and no calendar.

You have spent your whole career being told the margin is won in the field. Watch a project engineer spend Thursday reformatting a submittal log, and then tell me where the margin went.