Your existing technology renewal packages may look a little different this year. Somewhere in the line items there will be a package with a name like Digital Coworker, and a salesperson will walk you through what agentic AI can do for your team, and you will have to decide what it is worth to a mechanical contractor running eleven jobs with six project managers.
That decision is harder than the demo makes it look, because what is being sold has changed shape twice in eighteen months.
What the vendors shipped
We are already seeing an introduction of agents that take action instead of answering questions. Drafting RFIs. Reviewing submittals. Reading contracts. In every case the intelligence is being layered onto a platform your team is already licensed for and already knows.
The part worth reading closely is the metering. For twenty years construction software was sold by the seat. You bought access for a number of humans, and it cost the vendor nothing extra when those humans worked harder inside the tool. Agents break that arrangement. Every drawing set an agent reads costs real money to read, so the vendor has to meter it and you have to manage a balance. Payments, data and infrastructure software all went through this same transition a decade ago, and it usually caught the finance team off guard.
Where the value is real
Within a narrow band, these tools deliver. Deloitte research commissioned by Autodesk found preconstruction staff spending upward of thirteen hours a week researching and analyzing data. If an agent takes a real bite out of that number, the math gets easy.
Look at how specific those wins are. Counting receptacles. Calculating conduit runs including rises and drops. Pulling what you paid for the same assembly on a job two years ago. The common thread is that the work is mechanically repetitive and the output can be checked in under a minute. Every serious vendor describes these as human-in-the-loop tools, which reflects both engineering reality and a live commercial problem: professional indemnity coverage for AI-generated output remains unsettled, and no estimator is sealing a bid they have not verified themselves.
So the question in front of you is narrower than the marketing suggests. Do you have a workflow with a countable denominator?
Three questions before you sign
Name the number first. Hours per week your estimators spend on symbol takeoff. Days between a submittal landing and a reviewer touching it. How many change order disputes last year turned on a clause nobody found in time. With a baseline you can price an agent against it and know within a quarter whether it earned out. Without one you are buying a capability and waiting for a use case to appear, which rarely happens on its own.
Ask what the agent is reading. Synthesis is only as good as the documents feeding it. A contractor whose subcontracts sit as scanned PDFs in a shared drive will get vague answers back no matter whose logo is on the panel. The document cleanup is the real purchase; the agent is how you access it.
Ask who owns the process knowledge. Switching costs get built there. Decide on purpose whether you want yours built inside someone else's platform.
None of this argues against buying. Estimating desks sit open for two and three months at a stretch in this market, and capacity you cannot hire is capacity worth paying for. It argues for buying the way a CFO would rather than the way an early adopter would: against a measured baseline, on a workflow you can audit, with a clear view of what the invoice looks like when usage triples in the second year.
The contractors who get hurt here will not be the ones who moved early. They will be the ones who bought a tier without first writing down the number it was supposed to move.