When an AI Agent Is Wrong, Who Signs for It?

Autonomous AI is moving into construction workflows. The accountability has not moved with it.

Industry Trends
Renata Pinheiro
Renata Pinheiro
June 25, 2026 · 7 min read
A construction executive pausing over a contract with a pen, a project dashboard on a laptop beside them, warm window light

An AI agent can now reorder your submittals, push an update to your Primavera schedule, and place a material order with a supplier, all without a person approving any single step. The companies selling these agents say it plainly. One construction procurement agent advertises that it "auto-orders materials before you run dry, and handles every supplier call. 24/7. No humans required."

Read that last sentence again. "No humans required" is not a feature. It is a transfer of accountability, and the direction of that transfer should worry every contractor about to sign up.

The agents are already on your jobsite, or about to be

This is not a forecast. Autonomous AI agents moved from demo to deployment across construction in the first half of 2026. The shift is loudest in the back office, not the field. As one digital construction leader put it in June 2026, the real change "isn't happening on the jobsite. It's in the back office. Agentic ERP, PDF-to-Procore pipelines, voice agents fielding intake." That is where the hours leak, so that is where the agents are going first.

The pitches are specific and the numbers are real. Procurement agents claim contractors lose 8 to 14 percent on materials because ordering runs on phone calls and gut instinct, then offer to run the ordering themselves. Project controls tools now generate monthly cost forecasts and surface at-risk projects before the CFO asks. These are not toys. They touch money, schedule, and contractual obligations.

The question is no longer whether agents arrive. It is what you allow them to decide.

"No humans required" is a liability transfer

Here is what the vendor slide does not show. When an agent acts and the action is wrong, the contractor who deployed it carries the consequence, not the company that built it.

Several jurisdictions in 2026 are converging on a "reasonable oversight" standard. The organization deploying the agent is liable unless it can prove it had real monitoring, auditing, and controls in place. In the United States, the FTC has been direct that a business cannot outsource its compliance. Singapore went further and published a Model AI Governance Framework for Agentic AI in January 2026, stating that organizations remain legally accountable for their agents' behavior regardless of any voluntary vendor assurances.

Now read the contracts. As Clifford Chance flagged in early 2026, most technology agreements place full responsibility for legal and regulatory compliance on the customer. The supplier controls whether the agent behaves, yet the customer absorbs the consequence when it does not. If an agent misprices a change order, authorizes the wrong supplier payment, or pushes a schedule update that misses a contractual milestone, the vendor disclaimer often holds and the liability lands on you.

"

Autonomy without accountability is not innovation. It is unpriced risk.

"The AI did it" does not survive a courtroom, a claim, or an owner dispute. Someone on your side of the table signs for the agent's actions. The only choice you control is whether you decided what the agent was allowed to do before it acted, or after.

The mid-size firm is the most exposed

The contractors most at risk here are not the giants. They are firms your size.

AI adoption in construction is splitting into two tiers. By one June 2026 account, roughly 67 percent of construction AI adoption sits with firms doing more than $1 billion in revenue, while only about 12 percent of contractors under $50 million have deployed any AI tool at all. Among firms over $50 million, adoption climbed from 8 percent in 2020 to 39 percent by 2023, and per AGC, 61 percent of construction firms now use AI or plan to increase investment, up from 44 percent in 2024. The momentum is real and it is moving down-market fast.

That speed is the exposure. The $1 billion firms have IT departments, in-house counsel, and innovation teams that build oversight scaffolding around an agent before it touches a live project. A mid-size general contractor moves faster, which is an advantage everywhere except here. Deploying an autonomous agent at speed, without the monitoring and decision boundaries the large firms put in place, means adopting the liability without the controls that make it survivable.

You do not need a million-dollar innovation lab to close that gap. You need to decide the boundaries before the agent runs.

Decide the gates before the agent acts

The fix is not to avoid agents. It is to govern them. Before any autonomous tool touches a live project, every action it can take should be sorted into one of three zones.

In a Workflow Map™, those boundaries are explicit and visible to everyone who touches the process:

The procurement agent that wants to auto-order materials "with no humans required" does not belong in the "AI handles it" tier. Ordering against a live budget is an "AI recommends, you decide" action at minimum. Naming that out loud, before deployment, is the difference between an agent that saves your team hours and an agent that signs your firm up for unpriced risk.

The order matters. We start with the questions about which decisions actually carry risk for your firm. We agree on governance boundaries next. The Workflow Map is the artifact that captures those decisions, not the place the conversation begins. It is the same principle behind what your firm should still own as AI moves in: humans lead, AI assists.

The takeaway

Autonomous agents are coming into construction workflows whether or not your firm has a position on them. The contractors who do well will not be the ones who adopt the fastest or the ones who refuse the longest. They will be the ones who decided, in writing, what their agents are allowed to do before the first one acted.

If your team is evaluating an AI agent for procurement, scheduling, or document work, our Discovery engagement maps every action it would take to a clear decision zone, so you know exactly where your people stay in control. That is the whole point. Humanity for what matters. AI for everything else.

Common questions
What is an AI agent in a construction workflow?
An AI agent is software that takes actions on its own, not just recommendations. In construction that can mean reordering submittals, updating the schedule, flagging risk, or placing material orders without a person approving each step. That autonomy is what separates an agent from a standard tool.
If we use a vendor's AI agent and it makes a costly mistake, is the vendor liable?
Usually not. Most vendor agreements push compliance and legal responsibility onto the deploying customer, and the 2026 reasonable oversight standard holds the deployer accountable unless they can show real monitoring and controls. The contractor running the agent generally owns the outcome.
Should mid-size contractors avoid AI agents because of this risk?
No. The answer is not avoidance, it is governance. Decide which actions an agent can take on its own, which it recommends for human approval, and which a human must own before you deploy it. That boundary is what protects you.
How do we decide what an AI agent is allowed to do?
Start with the questions about which decisions carry real risk, agree on governance boundaries, then map every action to one of three levels of authority: AI handles it, AI recommends and you decide, or you lead and AI supports. A Workflow Map captures those boundaries so everyone knows where authority sits.
What is a Workflow Map in construction?
A Workflow Map is a visual blueprint that shows, for a given process, where AI acts on its own, where AI recommends and a person approves, and where a human makes the decision with AI support. It makes the boundary between human judgment and AI execution explicit before any tool is deployed.
Which AI agent actions are safe to fully automate on a construction project?
Generally the high-volume, low-judgment work where a mistake is small and recoverable: pulling data across documents, formatting reports, drafting routine intake responses. Anything that commits money, changes the schedule against a contractual milestone, or affects safety or a client relationship should require human approval, not full automation.
Does putting governance around AI agents slow down adoption?
No. Defining the decision boundaries up front is what lets a firm adopt agents faster, because the team trusts where the limits are. The delay comes from cleaning up after an agent acted outside its authority, not from agreeing on its authority beforehand.