What AI actually does for a 20-person construction company

If you run a 20-person construction company and you've been Googling "AI for construction" lately, here's what you've probably seen: jobsite robots laying brick, drones doing site surveys, computer vision spotting safety violations, and a hundred startups telling you that the future of building is BIM-plus-AI.

Most of that is interesting. None of it is going to change your week.

I spent twelve years in construction — field, estimating, project management, operations. The companies I worked at lost more profit to lost change orders, slow estimating, and late vendor invoices than they ever lost to bad layout or slow trades. The work that quietly burns margin in a small GC is back-office work. And that's exactly where AI is finally cheap and reliable enough to matter.

This is what AI actually does for a 20-person construction company in 2026. Not theoretical. Not a demo. Real wins, with real tools, scoped for a real budget.

The unsexy truth about where your time goes

Pull a week's calendar from your project manager, your estimator, and your office admin. Tally the hours. The pattern is consistent across every small GC I've worked with:

• 30–45% of admin time is intake. Reading RFPs, parsing scope, classifying RFIs, summarizing emails, pulling key dates out of long PDFs.
• 20–30% is data entry. Pulling line items from supplier quotes into your estimating template. Re-typing PO numbers. Updating spreadsheets that already exist somewhere else.
• 15–20% is status communication. Writing daily reports. Updating the owner. Fielding "where are we at" questions from subs.
• The rest is the actual work.

If you can shave 30% off the first three buckets, you've effectively given a 20-person company two extra people without hiring. That's the case for AI in your shop. Not robots. Time.

Five things AI actually does well right now

These are the engagements I see deliver real, measurable returns inside 30 days. Cheap to start, easy to extend, and they all run inside Microsoft 365 or Google Workspace — no rebuild required.

1. RFP intake and triage

You get a 40-page RFP at 4 p.m. with a 9 a.m. response deadline. AI doesn't write the bid for you, but it absolutely:

• Pulls the project address, owner, due date, scope summary, and required forms into a single intake sheet.
• Flags missing info you need to chase (insurance limits? bonding requirements? union scope?).
• Drafts the cover letter and the standard sections you say the same way every time.

What used to be a two-hour intake by a senior estimator becomes 15 minutes of review. Tooling: Claude or Gemini connected to your shared drive, plus a saved prompt template.

2. Estimating helpers, not estimators

AI is bad at deciding the number. It's surprisingly good at the prep that surrounds the number:

• Extracting line items from supplier PDF quotes into your spreadsheet.
• Cross-checking your takeoff against the spec section it should match.
• Generating a clean cost narrative from your raw markup notes.

Your senior estimator still owns the bid. They just spend their time on judgment, not data entry.

3. Daily reports that get actually read

Every PM hates writing daily reports. Every owner hates reading the lazy ones. AI fixes both:

• Voice memo from the field gets transcribed, structured against your daily-report template, and sent to the owner by 5:30 p.m. with photos attached.
• The same data feeds your weekly schedule narrative, your monthly owner report, and your project-close documentation. Write once, reuse four times.

This is the highest-immediate-trust win we ship. Owners feel the change in week one.

4. Change orders before they become arguments

Late or sloppy change orders cost you money in two ways: the ones you forget to bill, and the ones you bill late. AI helps both ends:

• A scope-change flag in your RFI process: "this conversation contains a probable scope change — should it be a CO?"
• Auto-drafting the CO narrative from the RFI thread, the field photos, and the contract language.
• A weekly "open CO" digest pushed to the PM and the owner.

The friction drops to near zero. The CO either ships or gets explicitly killed — both better than dying in someone's inbox.

5. Vendor and trade invoice triage

Your AP person opens 40 invoices a week. Half are coded right. The rest need chasing — wrong PO, wrong job number, wrong total against the contract. AI matches each invoice against the open commitments and surfaces only the ones that need a human.

You go from "person scanning everything" to "person resolving exceptions." Same headcount, double the throughput.

The Five Question Filter

When clients ask which AI use case to start with, I don't lead with technology. I run the Five Question Filter. If a candidate use case scores 4 or 5 out of 5, it's a good first build. Anything below 3 — wait.

#	Question	Why it matters
1	Frequency — does this happen weekly or more?	Less than weekly and you'll never recoup the setup cost.
2	Repeatable shape — is the input mostly the same kind of thing?	AI eats consistent inputs; one-off bespoke work is for humans.
3	Tolerance — is "90% right, reviewed by a human" acceptable?	If a single error costs a job, the use case isn't ready.
4	Owner — is there one person whose week gets visibly better?	No internal champion = no adoption, no matter how clean the build.
5	Measurable — can we name a number that should move?	Cycle time, throughput, error rate. If you can't measure it, you can't justify it.

For a typical 20-person GC, the use cases that score 4–5 out of 5 are usually: daily reports, RFP intake, vendor invoice matching, and change-order drafting — in roughly that order of leverage.

What it actually costs

The conversation usually ends with: "OK, but what does it cost?"

Realistic numbers for a 20-person GC, 2026:

• AI tooling — $50–$150 per seat per month (Claude, Gemini, OpenAI workspace, plus one or two integrations like Make or n8n). Call it $300–$800 a month total.
• Setup — a focused Bottleneck Audit to map where your time goes and what to fix first, plus the Automation Build that follows, lands at $2,500–$20,000 depending on scope.
• Internal time — about 4–6 hours per week from one champion during the build, then dropping to under an hour after handoff.

Compare that to one extra hire ($75,000+ all-in for an admin or junior estimator) and the math is rarely close. The AI build pays back inside two months on the right use case. The hire takes a year.

What it doesn't do

I'd be a poor consultant if I let you read this and walk away thinking AI is magic. The things AI is still bad at, in 2026:

• Final-mile decisions on bids. Don't let it set your number. It'll be confidently wrong.
• Anything safety-critical without a human in the loop. Period.
• Replacing tribal knowledge. Your senior super knows things that aren't written anywhere. AI can't extract what was never captured.
• Working with bad data. If your job costs are a mess in the GL, AI just gives you a faster mess.

Where to start

If you're running a 20-person construction company and wondering what to do with all of this, here's the honest answer: don't buy software. Don't hire a "Chief AI Officer." Don't tell your team to "experiment with ChatGPT."

Pick one of the five use cases above. Score it on the Five Question Filter. If it's a 4 or 5, scope a two-week sprint, ship it, and measure the result. If it works, the next build is obvious. If it doesn't, you learned something for the cost of a single sub-trade lunch budget.

That's how this gets real. Small, measured, owned by someone on your team. Then the second one. Then the third.

If you'd rather have someone help you pick the first one, that's literally what we do — the Bottleneck Audit is designed for exactly this conversation. One week, $2,500, you walk away with a prioritized fix list and a clear next build.

Either way: start small, start with what already burns you, and don't fall for the demo.