Walk any trade show floor this year and every booth has three letters on the banner: A-I. The pitch is always the same: smarter, faster, fewer headaches. But out on an actual project, most of that promise is still vaporware. Here's the honest map of where AI earns its keep on a jobsite today, and where it's still a demo looking for a problem.
Where it actually saves hours
The wins are unglamorous and specific. AI is good right now at the paperwork layer that sits between the field and the office, the stuff that eats a project manager's evenings.
1. First-draft estimating
Feeding a scope description to a model and getting back a structured takeoff (line items, quantities, ballpark unit costs) is genuinely useful. It won't replace a real estimator, and you'll correct it. But going from a blank sheet to an 80%-there draft in thirty seconds changes how fast you can turn around a budget. The key word is draft: it's a starting point a human sharpens, not a number you send to an owner.
2. Submittals and RFIs
Generating a properly formatted RFI from a rough note, or checking a submittal package for the documents it's missing against the spec section, is a real time sink that AI handles well. This is pattern-matching against known formats, exactly what these tools are built for.
3. Punch lists from photos
Snap a photo of a defect, get a suggested punch item back with a title, trade, priority, and location. It's not magic, but on a closeout walk where you're logging forty items, shaving fifteen seconds off each one adds up to a coffee break.
The best AI on a jobsite is the AI you don't notice. It just makes the boring part shorter.
4. Voice-first field capture
This is the sleeper. A superintendent standing in mud, gloves on, is never going to tap through a form. But they'll press one button and talk. Turning "we lost the morning to rain, poured footings on grid C after lunch" into a structured daily log, a weather delay, and a time entry, all automatically, is the difference between a log that gets filled out and one that doesn't.
Where it's still a demo
Be skeptical when the pitch gets bigger than the paperwork layer.
- "Predictive schedule risk." Models can flag phases drifting behind, and that's fine as a prompt to go look. But a confident "this project will finish 18 days late" from a tool that's never met your subs is a number to distrust.
- Autonomous anything. Nobody should be letting AI approve a change order, release a payment, or sign a waiver. The liability math doesn't work, and it shouldn't.
- "AI project manager." The job is 70% relationships and judgment. A model can draft the email. It can't read the room on why the owner is actually upset.
The honest takeaway
The right question isn't "does it have AI?" It's "does the AI shorten a task I actually do every day?" Estimating drafts, submittal checks, photo-to-punch, voice logs: those clear that bar today. Everything else is worth watching but not worth betting a project on.
That's the line we hold building AccuDone: AI should make the tedious 20 minutes into 2, and then get out of the way. The person still runs the job.