Construction's AI Is Mostly in the Back Office
AGC of America and Sage publish an annual outlook survey. This year's, titled "Dampened Expectations," was completed by 951 respondents across 49 states and the District of Columbia, fielded November 4 through December 15, 2025.
The line that gets quoted is that 61 percent of respondents use AI or plan to increase investments in it, up from 44 percent a year earlier. Hold that sentence loosely; I will come back to it, because AGC also publishes the question it came from and the two do not match.
The breakdown underneath is the part worth reading. Asked whether their firm had initiated or increased its use of AI in the past year, in a question with 857 responses, surveyed contractors answered:
Office applications, 45 percent
Estimating, 23 percent
Design or preconstruction, 20 percent
Recruitment, training and other HR, 16 percent
Onsite construction activity, monitoring and documentation, 13 percent
Scheduling, 12 percent
Procurement, 7 percent
None of the above, 39 percent
So the most common application, by roughly two to one over estimating, is the back office.
The obvious reading is wrong
The obvious reading is that construction is conservative, that estimators are set in their ways, that the industry is slow. I have heard all three said out loud at conferences by people selling something.
I think the ordering is rational, and you can derive it from one property of the two kinds of work.
An administrative error is usually recoverable and it usually announces itself. A misfiled document gets found. A badly worded email gets a confused reply and you send another one. A wrong calendar invite produces somebody standing in the wrong room, immediately.
I want to be careful with that, because "office applications" is a broad category and the expensive end of it is genuinely expensive. A conditional lien waiver issued with the wrong through-date is an administrative act, it is silent, and it is not recoverable in an hour. Certified payroll filed wrong on a prevailing wage job is administrative and it is a compliance event. So the asymmetry is not absolute, and anybody who has run a back office will have a counterexample ready.
The point survives in the aggregate. Most administrative work has a short feedback loop and a bounded cost when it goes wrong.
An estimating error has neither. It gets priced into a number, the number goes into a bid, the bid becomes a contract, and the error is now a term of a signed agreement. Nobody calls to tell you. You find out during construction, when the material runs short or the labor hours run past, and by then the recovery mechanism is a change order you probably will not win or a margin you simply eat.
That is not caution. It is a correct read of where mistakes are cheap and where they are expensive.
Why the estimate is genuinely harder, and it is not the arithmetic
There is a second reason, and it is the one I would want anyone building tools for this to understand, because it gets missed constantly.
The hard part of a takeoff is not the arithmetic. It has never been the arithmetic. Multiplying a length by a height is not what takes the time.
The hard part is that the answer is not in one document.
Take something as ordinary as how much wall tile goes in a bathroom. Four documents have a say, and they have different jobs.
The finish schedule tells you whether that room gets tile at all, and which one. The floor plan tells you the shape of the room, and you need the traced interior face, not the dimension string, because the dimension string is often to structure or to the stud line and the tile does not stop there. The interior elevation tells you how high it goes, and it is frequently the only place the height appears anywhere in the set. The reflected ceiling plan tells you what is happening above, which decides whether you are running to a ceiling, to a soffit, or to a hard lid at a different elevation.
Miss any one of those and you still get a number. It just is not the right number, and nothing in the arithmetic will tell you.
And then the rule experienced estimators internalize and almost never write down: when the schedule and the drawing disagree, the schedule wins. The drawing is a picture of intent. The schedule is the specification of what actually goes there. A drawing showing tile in a room the schedule says gets paint is a drawing that is wrong, and an estimator who follows the picture has bought material for a room that is not getting it.
That precedence rule is not in the drawing. It is not in the schedule either. It is in the estimator's head, and it got there by being wrong about it once.
So when the survey reports 23 percent on estimating, it is describing the harder problem being attempted by fewer people, not the easier problem being avoided by cowards.
A prediction, and the survey already falsifies it
Here is where I would have left this, and I want to show the check rather than skip it.
If the recoverability asymmetry is the real driver, then the next thing to move should not be estimating. It should be whatever sits between administration and the estimate: work with the document-handling character of the back office and the consequence profile of a mistake that is still cheap to fix. Scheduling and procurement are exactly that middle ground.
AGC publishes both, and both come in below estimating. Scheduling at 12 percent, procurement at 7 percent, against estimating at 23.
So the prediction fails. Whatever is ordering these applications, it is not a clean gradient of how recoverable the mistake is, or the two middle categories would sit above the expensive one rather than well below it.
And there is a worse problem for my argument in the same list, which I should state rather than wait to be told. Estimating is not last. It ranks second of the seven application options, behind office work and ahead of design and preconstruction, HR, onsite activity, scheduling and procurement. The whole premise of this piece is that estimating trails because estimating mistakes are expensive and silent. It outranks five other categories, including two that are cheaper to get wrong.
I could rescue the hypothesis by arguing that scheduling and procurement are smaller surfaces with less repetitive document handling than an office. Maybe. But that is a rescue. The honest version is that the recoverability asymmetry is consistent with office work leading, and it does not explain the rest of the ordering. It is one comparison that holds, not a theory of the whole list, and I am not going to inflate it into one.
What is wrong with these numbers
AGC's own summary sentence does not match AGC's own question, and the summary is what everyone quotes. The report's prose says firms "use AI or plan to increase investments in it." The published instrument asks something narrower: "Has your firm initiated or increased its use of artificial intelligence (AI) in the past year for any of the following?" That is a past-year action question. There is no forward-looking investment option in it. The 61 percent is simply 100 minus the 39 percent who selected "None of the above."
So the headline figure is real, and the sentence carrying it around is a loose paraphrase written by whoever drafted the summary, not a description of what was asked. Anyone repeating "61 percent of contractors use AI or plan to invest more" is repeating the paraphrase.
One loose end I cannot close, and it is the place someone could push back on this. The report says "61 percent of respondents," and 951 respondents completed the survey, so a reader naturally takes it as 61 percent of 951. But question 15 drew 857 responses, and if the 61 percent is the complement of that question's "none of the above" then it is 61 percent of 857. Those are different bases. Question 15 is the only AI question in the instrument, so the complement is the only derivation I can see, but the report never reconciles the two numbers and neither can I.
The instrument is published, in a fact sheet linked from the same materials index as the report itself. If the argument is that everyone else is misreading a survey, the survey is the thing to open.
The four categories everyone quotes are four of eight. The full option set, with a shared base of 857 responses, is listed above. The seven application options sum to 136 percent, which is well over 100, so a firm using AI in more than one place is counted in each. They are not slices of a pie and they never were. That makes the overlap point stronger than the four-category version of it suggests, not weaker.
It is called the 2026 Outlook and it was fielded in 2025. November 4 through December 15, 2025. Every figure in it is a late-2025 report of the past year, not a measurement of anything that has happened in 2026.
This is AGC's membership, and it skews toward general contractors. Asked what best describes the firm, 60 percent said general contractor or construction manager and 31 percent said specialty or subcontractor. So on a piece aimed at estimators, the respondent base is majority GC. It is not a random sample of American construction firms.