Governing AI Agents on Capital Projects
Autonomy is not the goal. A governance model for construction agents — permissions, approval gates, evidence requirements and the audit trail that makes them defensible.
The question executives ask about construction agents is never "can it do the task." It is "what happens when it is wrong, and who is answerable."
That question deserves a better answer than a confidence score.
Autonomy is the wrong axis
Vendor demos tend to present agent capability as a slider from assists to acts independently, with independence as the aspiration. On a capital project this framing is close to useless. The relevant question is not how much the agent does on its own. It is which class of decision it is touching.
We separate three:
Reversible and low-consequence. Retrieving evidence, drafting a summary, classifying a document, flagging an anomaly. Wrong output costs a few minutes. Agents operate freely here.
Reversible but consequential. Proposing a resequence, drafting an RFI response, assembling a delay narrative. Wrong output costs credibility and rework. Agents produce; humans review before anything leaves the system.
Irreversible or contractually significant. Anything that enters the contractual record, commits resources, or forms the basis of a claim. Agents never execute these. They prepare them, with evidence attached, for a named human to approve.
The slider is not a slider. It is a set of gates.
Four requirements
1. Scoped permissions. An agent operates within a project, a package, a role and a tool set. It cannot read what its principal cannot read. This sounds obvious and is routinely violated by systems that index everything into one vector store.
2. Evidence before conclusion. No claim without a traceable source. If the agent cannot cite the record that supports a statement, the statement does not get made. This is a hard constraint in the architecture, not a prompt instruction — prompt instructions are suggestions, and this cannot be a suggestion.
3. Explicit approval, explicitly recorded. Not a checkbox at the end of a workflow. A named person, a timestamp, the exact artefact approved, and the evidence set that was visible at the moment of approval. Approval that cannot be reconstructed later is not governance.
4. The complete trail. What the agent observed, how it reasoned, what it recommended, who approved, what was executed, and what actually happened. Including the recommendations that were rejected — those are often the most valuable record you have when a dispute arrives.
Why this makes agents more useful, not less
There is a persistent assumption that governance is a tax on capability. On capital projects the opposite holds.
An ungoverned agent produces output nobody will stake a decision on. It gets used for drafting and abandoned for anything that matters. A governed agent produces output with its justification attached — which is precisely the form in which a project director can act on it immediately.
Constraint is what makes the output usable. The audit trail is not the compliance overhead around the product. On a project where every significant decision may eventually be examined by a claims consultant, it is the product.