Construction technology has been over-promised on AI for a decade. Computer vision was going to solve safety. Generative design was going to redo the architect's job. Schedule prediction was going to make the GC obsolete. None of that happened, and the firms that bought into it have a graveyard of unused subscriptions.
The good news: somewhere in the rubble, six AI use cases have started producing real, measurable outcomes. We've shipped most of them. Here's what's working, what isn't, and what to do about it.
1. RFP intelligence — works
Reading an inbound RFP and producing a one-page go/no-go brief is the highest-yield AI work on the pursuit side, full stop. We've watched firms shrink RFP analysis from 4 hours to 5 minutes, with explainable scores tied to evidence in the document.
Why it works: the document is structured, the firm's win history is the ground truth, and the output is a brief that humans review. The AI never makes the bid/no-bid call — it makes the call defensible in less time.
2. Proposal drafting — works, with caveats
Federal proposals especially benefit. Your library of past submissions is huge, the requirements are predictable, and 80% of any new proposal is structured content the AI can produce in your voice if it's been trained on enough source material.
The caveat is data hygiene. Firms with messy past-submission archives get bad output. Firms that spend three weeks cleaning their library before training get great output. Plan accordingly.
3. Daily field reports — works
Voice-driven field reports are a quiet, durable win. Super dictates a 90-second voice note on the way back to the truck. PM gets a structured report in their inbox by the time they sit down at their desk. Quality of report is higher than when supers were typing on a phone.
This is the use case that wins skeptics in the field. They don't think they want AI; they want the truck to drive itself home. Daily reports get them there.
4. Submittal and RFI processing — works, mostly
Auto-classifying submittals against specs and routing RFIs to the right reviewer with a draft response is solid work. Cycle time on submittals drops 40-60%. The "mostly" is that you still want human sign-off on the gnarly ones — the AI should auto-handle the 70% that are obvious and surface the 30% that aren't.
5. Schedule risk analysis — works for narrow questions, not broad ones
Asking AI "will this project slip?" gets you a confidently wrong answer. Asking "which of these 40 activities is most likely to slip based on similar past projects?" gets you a useful one. Narrow questions, narrow scope, narrow output. Then a human PM does the integration.
6. Owner reporting — works
Monthly owner reports are 80% structured data (schedule, cost, quality) and 20% narrative judgment. AI handles the 80%, your PM does the 20%, the output looks like a senior PM wrote it.
Time savings are real: 16 hours/month/PM, in the firms we've measured.
What still doesn't work
- Computer vision for safety. The cameras work fine. The models still hallucinate hazards 8% of the time, and the false-positive cost (alert fatigue) outweighs the catch rate. Wait 2026.
- AI estimating from public databases. Public unit prices are too generic. Your historical pricing is your competitive advantage; train on it.
- Generative design. Useful for early massing studies. Not yet useful for permit-ready drawings. Architects are not going to be replaced this decade.
- "AI everywhere" platforms. Anything that promises to do six of the above in one product is going to do six things at 60% quality. Buy one good tool per workflow.
How to choose what to ship first
Pick the one workflow where:
- The output is reviewed by a human before it goes to a client/owner/contractor.
- Your team is the bottleneck — adding AI gives you back hours, not just quality.
- The training data exists (past examples, structured records, etc.).
For most AEC firms, that's pursuit work — RFP intelligence and proposal drafting — because the bottleneck (principal time) is also the most expensive and the past examples are well-organized. If your firm is delivery-heavy, start with field reports and owner reports instead.