Building Easier AI

The AI Punch List: Rank by Payback, Not by Hype

Search for how long AI takes to pay for itself in a construction business and you'll have an answer inside a minute. One search gave me four: a payback window in months, two survey shares of early adopters already banking gains, a five-figure saving per firm.

Now try to find who measured any of it. Every one traces back to a company selling the thing it measures — an AI bidding tool, an app-development shop, two consultancies. Not one carries a sample size, a methodology, or a named author. I went looking for an independent, published payback period for AI in construction to anchor this article. There isn't one, so no number from that pile appears below.

The absence is the finding, and it isn't a gap you can wait out. The research that does exist points somewhere the sales decks don't: in the AGC and Sage 2026 outlook, 45% of firms deploy AI on office and administrative work, against 23% on estimating. The hype is aimed at the takeoff; the usage is in the file room.

So the ranking has to come off your own P&L rather than a vendor's. Here is how to build that list, and the three findings that should change the order you put things in.

The 61% everyone quotes is two numbers glued together

The 61% figure in circulation this year comes from the AGC/Sage 2026 Construction Hiring and Business Outlook, worded like this: "61 percent of respondents say their firms use AI or plan to increase investments in it, up from 44 percent in last year's survey." Read that twice. It blends firms already running AI with firms that merely intend to spend more next year. One of those has a payback; the other has a budget line.

The breakdown underneath is worth keeping, because each figure counts a deployment rather than an intention: office and administrative functions 45%, estimating 23%, design or preconstruction 20%, recruitment and training 16%.

Three surveys measured AI adoption. They came back 18%, 41% and 78%

A Federal Reserve FEDS Note published April 3, 2026 by Jeffrey S. Allen lines up three serious surveys of AI adoption and gets three incompatible answers. Census's Business Trends and Outlook Survey, weighted by firm, puts adoption near 18% at the end of 2025. The Real-Time Population Survey, which asks individuals, says about 41% of the workforce uses generative AI at work. The Survey of Business Uncertainty, weighted by employment, puts roughly 78% of the labor force at a firm that has adopted it. Allen attributes the spread first to what each survey counts — firms, individuals, or employees at adopting firms — then to question wording and social desirability bias; people like saying yes to AI questions.

Sixty points between the high and the low estimate, and not one of the three is sloppy work. If the country can't produce a single number for how many businesses are doing this, no ranking built on adoption rates is worth anything — not "you're behind," not "X% of your competitors already have it."

One figure deserves wall space. In Census's Business Trends and Outlook Survey, 37% of firms with at least 250 employees reported using AI, against a national rate of 19.8% as of May 3, 2026 — and adoption among firms with fewer than 20 employees didn't change significantly over the prior six months. If you have eleven people and no AI, you're not behind. You're the middle of the distribution.

Nineteen percent slower, and they were sure they were faster

The sharpest counterweight available is a randomized controlled trial METR published in July 2025, on data collected that February through June: 16 experienced open-source developers, 246 real tasks, randomly assigned to allow or forbid AI tools. They finished 19% slower with AI. Afterwards they estimated AI had made them 20% faster — and going in, they had forecast a 24% speedup.

Two caveats in the same breath: it is a 2025 study, and METR itself now calls the result historical — in February 2026 it wrote that these results "no longer reflect the current impact of AI models," and that developers are likely faster with 2026 tools. Grant both and something still stands: a 39-point gap between measured and felt productivity, in a small group of experts paying close attention. Better models don't close a perception gap, which is why you rank by payback rather than by how the pilot felt — and why "the team loves it" is not evidence.

Rank it the way you'd rank a punch list

A punch list gets ordered by what's keeping the keys from changing hands, not by what's interesting to fix. Rank AI candidates the same way, in three columns you can fill in an afternoon.

Hours. Count them, don't estimate them. If your office manager codes and enters supplier invoices, count last month's — say 260 — and time yourself on ten. At three minutes apiece that's 13 hours a month, 156 a year.

The loaded rate of the seat doing it. At $32 an hour, 156 hours is about $5,000 a year. Discount the recovery, because nothing goes to zero and a human still approves every invoice: take out a third and you have roughly $3,300 a year of recoverable cost.

What it costs to get. Software, setup, and the hours your own people burn making it work. Say $1,800 a year in subscription and $2,500 of setup: $4,300 against $3,300. That candidate pays back in about sixteen months, which in a summer when commercial backlog just dropped 0.8 months is a no.

Change one input and it flips. Run the same arithmetic on a task in a $38-an-hour seat that eats five hours a week and the recoverable number is near $6,600 — the same $4,300 is back in your pocket by month eight. The tool didn't get better. The seat got more expensive.

Most of your list dies on that page. The two or three still standing are the ones you could defend to a banker, and the order they come out in won't match the order anybody sold them to you in.

What I can't tell you either

So what's our payback number? We don't have one I'm willing to print.

PO Builder cuts a vendor-priced purchase order by voice from the truck. Permit Tracker watches county portals so nobody has to. Robyn, Jobs to Pay and thirteen production skills run inside our own companies every week, doing estimating, scope drafting and plan revisions. They work — we wouldn't keep paying to run them otherwise. But the audited before-and-after that would let me print a percentage doesn't exist yet, and the minute I print one anyway I become the source I spent the opening of this article telling you to ignore.

The most flattering AI figure in circulation — 38% of commercial contractors reporting measurable business impact from AI, up from 17% a year earlier — comes from a 2026 report by ServiceTitan, which sells software to contractors. It may be perfectly accurate. It is also exactly the number a vendor would want true, published by the vendor. The same test applies to us.

Which leaves the conclusion the evidence keeps pointing at from three directions: nobody outside your business can rank this for you. Not the Fed, not the AGC, not us. Walk the seats, count the hours, run the columns. Then build the cheapest thing at the top of the list, and don't tell anyone it worked until you can point at the line item that shrank.