Ask a GC how they’d bring on a new office hire and you get a shrug. Ask how they’d bring on a new sub and you get a system: a scope, proof of insurance, a small first job to see if they’re any good, a walk-through before the check, and a lien waiver before anybody gets paid.
Most of the advice about AI for general contractors treats it like a hire. Train it, trust it, let it grow into the role. I think that’s backwards. Treat AI like a new sub. Give it a scope. Inspect its work. Pay by the task. Never let it touch the money.
You already know how to manage a sub. That’s the whole skill.
Scope of work: one job, clearly bounded
You’d never hand a sub a contract that says “help out around the site.” You’d hand them “rough plumbing, floors one and two, per sheets P-1 through P-4.” Same with AI.
The scopes that work are narrow and repeatable: “turn today’s daily log into a homeowner update,” not “handle client communication.” One job, defined inputs, defined output, done the same way every time. If you can’t write the scope in three sentences, it isn’t ready to hand off.
A sub with a vague scope does whatever’s easiest. So does the software.
Submittals: it shows its work before anything goes out
A good sub doesn’t install the fixture and then tell you what they picked. They send the submittal, you approve it, then they install. That’s exactly how AI should operate in a GC’s office.
The AI drafts. A person reads the draft, fixes what’s wrong, and sends it. Nothing goes to a homeowner, a sub, or an inspector without a human’s name on the send button. This isn’t a training-wheels phase you grow out of. It’s the permanent arrangement, the same way you’ll always review a submittal no matter how long you’ve worked with that sub.
The punch list: you inspect and correct
Every job gets a walk-through. You find the paint on the trim and the crooked outlet cover, you write it down, the sub fixes it. Then you sign off.
With AI, the punch list is the log. Every run is recorded: what went in, what came out, who approved it, what they changed. You read it once a week, the way you’d walk a job. When the AI keeps making the same mistake, you fix the scope, the way you’d fix a detail on the drawings instead of yelling at the framer.
If there’s no log, you can’t inspect the work, and a sub you can’t inspect doesn’t get the next job.
Lien waivers: it never touches the money
A sub doesn’t get paid until the waiver is signed, and a sub never writes checks on your account. Keep the same line with AI.
It can prep a pay app. It can flag that a sub’s hours don’t match the schedule. It can draft an invoice with the numbers pulled from the job. It cannot send the invoice, approve the pay app, release a payment, or change a price. A person does those, every time, with the AI’s draft in front of them.
That’s also the answer to “what if it makes a mistake.” It can’t make an expensive one; the expensive things sit behind a human signature.
The four office jobs a GC subs out first
These are the scopes that fit the frame: narrow, repeatable, and already happening every week.
- Change-order wording from field notes. The super texts “owner wants the island moved 18 inches, we’re already roughed in.” The AI turns that into a clean change-order description with the work added, the work removed, and a blank for the price. The PM fills in the price and sends it. See Change Orders.
- Daily logs into homeowner updates. The daily log already exists for the crew. The AI rewrites it in plain English for the homeowner, drops the jargon and the sub names, and leaves out anything you haven’t confirmed. The PM reads it and hits send. See Daily Log Update.
- Scope drafts from site notes. Walk the job, talk into your phone, and the Estimator agent turns it into a first-pass scope organized by trade, with every line marked “needs a quantity” or “needs a price.” You still price it. It just stops you starting from a blank page.
- Payroll prep checks. Before payroll runs, the AI compares time entries to the job schedule and flags the missing punch, the double entry, and the hours coded to a job that closed. Someone fixes them. Nothing gets paid until that person says so.
Notice what’s not on the list: pricing a bid, negotiating with a sub, deciding whether to take the job. Those are yours. For where AI fits on a job and where it still needs you, read AI in construction. For how this plays out across a GC’s office, see the general contractors page.
Where to go from here
Here’s the scope-of-work template we fill out before any AI task gets built. Fill it in for one job this week. If you can’t complete every line, that job isn’t ready to hand off, and that’s useful to know before you spend money on it.
SCOPE OF WORK: AI TASK
Job: [one sentence, e.g. "Turn the superintendent's daily log into a short homeowner update."]
Inputs: [exactly what it gets, e.g. "Today's daily log, the project name, the homeowner's first name."]
Output format: [e.g. "Plain text, under 150 words, no prices, no dates we haven't confirmed in writing."]
Must never: [e.g. "Quote a price. Promise a completion date. Name a sub. Send anything on its own."]
Who approves: [one name, e.g. "The PM reads every draft and sends it, or edits it, or kills it."]
Where the log goes: [e.g. "One row per run: date, job, link to the input, the output, who approved, what they changed."]
Done when: [e.g. "The PM has sent ten updates without changing more than a sentence."]
Keep it to one page and treat it like the contract. When the AI does something outside it, that’s a scope problem, and you fix the scope.
If you’d rather we build it, bring a filled-in template to a free 30-minute call. We’ll show you how we’d set up that one task in the software you already run, approval step and log included, and you’ll get a written plan with a fixed price before you decide anything.