When the Agent Acts First and Asks Never

An RFI lands on a Tuesday morning. By the time the engineer who owns it opens the project platform, a reply is already drafted and waiting, citing the specification sections and the drawing revision it drew from. Nobody approved it. Nobody was asked. The agent that wrote it triggered when the RFI arrived, searched the project record, and produced its answer with a log entry where a signature should be.

That is not a pilot or a preview. PlanRadar put AI agents into general availability across every plan it sells, and the agents act once triggered rather than waiting for a human to approve each step [1]. Every action is logged and attributed to the agent, every agent runs against its creator's own access rights each time it executes, and the whole thing sits inside the platform's existing ISO-certified data-protection framework [2]. The flagship agent reads an incoming RFI, searches the project's documents, and drafts a reply citing its exact sources [3].

Context matters here. In April, Procore and Autodesk launched agents that kept humans in the approval loop by default. PlanRadar's release is the most aggressive autonomy position a major construction platform has shipped: no approval step, on every plan, for everyone [2]. Document Crunch's Project Assist, landing inside Trimble, auto-generates redlines, submittals, notices and RFIs [4][5]. Whatever your view, the action-owning wave is no longer hypothetical, and one vendor has chosen the boldest default.

There are two honest ways to read that choice. They lead to opposite advice, and both deserve a full hearing.

The case for switching it on

Start with the governance, because it is more coherent than the headline suggests. Permission inheritance means an agent can never touch more than the person who built it could. Attribution means every drafted line traces back to a specific agent, on a specific run, with a log entry. These are real controls, designed by people who understood that "who did this" is the first question any dispute will ask.

Then there is the work itself. RFI cycles are where project weeks quietly disappear. A cited first draft waiting in the queue converts authoring time into verification time, and verifying an answer against named sources is a far cheaper task than composing one from scratch. The reviewer's job changes shape, not value.

The deeper argument is about defaults. Default settings become industry practice: if no-approval agents prove safe on live projects, every competing platform will face pressure to match the autonomy level, and QA workflows will start assuming machines initiate actions rather than suggest them. A contractor who prefers to move deliberately can do exactly that today, with their own gates configured on top. Fluency now beats inheritance later.

So the readiness checklist before switching autonomy on is concrete. Test citation quality against your own project documents, not a demo set. Confirm the permission model matches your actual org chart, including the consultants. Decide who countersigns machine-initiated correspondence before the first machine-initiated correspondence exists. Do that, and the agent is a fast junior with a complete audit trail.

The case that it's a land grab

"Customers don't want another login" is doing a lot of work in that pitch. Translate it, and the sentence says something else: the platform does not want to hand you a suggestion you could act on somewhere else. It wants to own the action. And owning the action means owning the record of the work.

Follow that to the evidence trail. When the agent drafts the response inside the vendor's platform, the log of what happened is written by the vendor's system, stored in the vendor's environment. Your only proof of what was said, when, and on what basis is an audit trail you do not control, describing actions no human approved.

The liability is not symmetric either. Your contract with the employer makes you answer for what issues from your side of the job. The vendor's terms cap its exposure at the price of a subscription. When an unapproved, machine-drafted response turns out to be wrong, the party standing in the room is the contractor who switched the agent on, not the company that shipped it.

Even the log, sold as the safeguard, can invert outside the platform's walls. Read by an insurer or opposing counsel, thousands of unreviewed machine actions may look less like mitigation and more like aggregation: a complete, timestamped record of how systematically nobody looked.

And whatever the first real incident proves becomes practice overnight. If the log saves the contractor, every platform ships the same default within a year. If the log damns one, autonomy gets quietly walked back and the whole category wears the scar. Either way, the answer gets written by whoever moves first, not by the industry's judgment.

Where both sides already agree

For all the disagreement, the two readings share their conclusions more than their advocates admit. The contracts are not ready: standard forms have no clause for machine-initiated correspondence, and legal commentary has flagged the gap all year. Nobody has published a measured error rate for agent-drafted RFIs on live projects. Nobody has published how an insurer treats an agent-enabled project. Logging, attribution and permission inheritance are necessary controls and completely unproven ones.

That is why the first documented dispute involving an autonomous construction agent matters more than any launch. It will set the tone for the entire product category, and it will do so using whatever contract language, or absence of language, happens to be in the file.

The factors that will decide which reading ages well are already visible: citation quality under field scrutiny, whether permission inheritance survives a real audit, who countersigns, what the first dispute proves about the log, and how the first insurer prices the exposure.

Two postures are on the table: switch autonomy on deliberately, with your own gates on top, or let the shipped default write industry practice while you watch. Which do you think will prove better on live projects? Tell us what you would need to see before trusting a draft no human approved.

Sources

[1] AEC Magazine. "PlanRadar adds AI Agents to automate routine tasks." August 20, 2026. https://aecmag.com/project-management/planradar-adds-ai-agents-to-automate-routine-tasks

[2] PlanRadar (release via Markets Insider). "PlanRadar AI Agents Close the Gap Between What Happens on a Project and What Gets Done." August 19, 2026. https://markets.businessinsider.com/news/stocks/planradar-ai-agents-close-the-gap-between-what-happens-on-a-project-and-what-gets-done-1036473446

[3] IT Brief Asia. "PlanRadar launches AI agents for project workflows." August 20, 2026. https://itbrief.asia/story/planradar-launches-ai-agents-for-project-workflows

[4] Trimble Newsroom. "Document Crunch Launches Construction's First Project-Level AI Risk Intelligence Platform." June 9, 2026. https://news.trimble.com/2026-06-09-Document-Crunch-Launches-Constructions-First-Project-Level-AI-Risk-Intelligence-Platform

[5] ENR (Jeff Yoders). "Trimble Acquires Document Crunch, Plans to Integrate Agentic AI Contract Review." April 3, 2026. https://www.enr.com/articles/62770-trimble-acquires-document-crunch-plans-to-integrate-agentic-ai-contract-review