You think you're sending a meeting recap. You're actually setting a baseline for a delay claim.
That's the whole problem with construction communication. Most of it looks the same on the page. A 200-word email recapping the OAC reads identical to a 200-word email confirming a price. One is a paper trail. The other is a contract.
When AI starts writing your emails, that line stops being academic.
What People Think Is Happening vs. What Actually Happens
People think: AI helps me write faster.
What actually happens: a PE on their second job lets a model draft a response to a GC on a coordination question. The model uses the word "confirm." The GC takes "confirm" as a commitment. The PE never reread the third paragraph. Three weeks later there's a change order request that points to that email.
This is not a hypothetical. It's a Tuesday.
The Stuff That Doesn't Hurt You
Some communication is volume. There's a lot of it, it follows patterns, and it doesn't bind anything to money or scope. Push it to AI without losing sleep.
- RFIs that are pure clarification. "Detail 4/A502 and structural notes conflict—which governs?" That sentence writes itself. The model can format it, log it, and route it. The reviewing engineer is the one carrying the weight.
- ITB intake. Pulling the GC name, bid date, walk-through time, and link to drawings out of an invitation email. Twenty minutes a pop, and BD does this thirty times a week.
- OAC minutes and due-out lists. The minutes describe what was said. They don't commit anyone to anything new. AI is good at this—just don't let it summarize away the part where the owner pushed back on a substitution.
- Toolbox talks, daily log narrative cleanup, training reminders, status nudges. The boring backbone of office work.
If the document moves information without moving money, it's fine.
The Stuff That Will End You
Now the other column. The one where the wrong sentence costs you the job's margin.
**Clarifications and qualifications on a proposal - ** One word swing. "Includes" vs. "excludes" temp power on a 14-month industrial build is a six-figure swing. AI can lay out the structure. A human reads every line out loud before it leaves the building.
**Anything tied to price - ** Change order narratives. VE memos. Pay-app cover letters. Schedule impact letters. AI can format. Estimating and the PM own the number, the tone, and whether the letter says "reservation of rights" or doesn't.
**Contract redlines and scope interpretations - ** AI is decent at flagging weird indemnity language. It is terrible at deciding whether to push back. That decision involves the GC's history with you, your pipeline, and how much risk your bonding company will tolerate this quarter. None of which the model knows.
**Incident reports, claims correspondence, anything an attorney might read - ** AI does not author this. Period. It can organize facts. A human—often with legal—writes the words.
**Anything that creates a binding commitment - ** If the email could appear in a deposition or a mediation binder, a human signs off.
A Quick Test Before You Hit Send
Before you let a model send anything externally, ask three questions:
- Does this define or shift scope?
- Does this reference price, schedule, or assumptions?
- Could a fair reading of this be used against us in a dispute?
If yes to any of them, it's a draft and only a draft. The PM, the estimator, or the project exec owns the words.
The Line, Drawn
Think about the laborers running conduit. Fast, accurate, repetitive—and they don't decide where the panel ties in. AI works the same way. Let it run the cable. The operator decides where the power lands.
The teams that figure this out first communicate faster, coordinate cleaner, and stop accidentally writing checks they didn't mean to write. The teams that don't will find out in mediation.