Operations
The Count You Still Check
By John J. BakerNovember 21, 20268 min read
An AI takeoff tool can now do the first count on a bid set. On clean, repeated symbols it is close to exact. On areas, odd shapes, linear runs and anything that takes reading the drawing like a person, it drifts. So the estimator's job shifts from counting everything to checking the specific places the tool is known to miss.
Here is how a bid week usually goes. An electrical sub is bidding a new two-story school. The estimator opens the lighting plans and starts clicking: type A fixtures, type B, the emergency fixtures, the exit signs. Then the power plans: duplex receptacles, quads, floor boxes, data drops. Then an addendum lands on Wednesday with three revised sheets, and part of the count starts over.
On the mechanical side it is the same week with different symbols: diffusers, grilles, VAV boxes, thermostats, and every run of duct and pipe.
By Thursday night, the work that actually wins or loses the job is still waiting. Reading the scope. Checking the vendor quotes. Setting labor units. Writing the qualifications. That is the part that needs a person with years behind them, and it gets whatever time is left.
Where does an estimator's bid week actually go?
Most of it goes to counting and measuring, which is the part of the job that needs the least judgment. A research team at Utica University put it plainly in a 2026 paper: estimators "often spend substantial time on measurement rather than higher-value tasks such as pricing analysis, constructability review, and risk assessment."
The bench is not getting bigger. The U.S. Bureau of Labor Statistics counted about 226,400 cost estimator jobs in 2025 and projects the occupation to decline 3 percent from 2025 to 2035. A third of all estimators work for specialty trade contractors, more than in any other industry. For most electrical and mechanical subs, estimator hours are what limit how many jobs you can bid.
What does an AI takeoff tool do, and how is it different from asking a chatbot?
An AI takeoff tool is software built to read construction drawings. It sets the scale on each sheet, finds the symbols that match your legend, counts and measures them, and puts a mark on every item it found so the estimator can see and click each one. Every number traces back to something on the sheet.
A general chatbot does not do that, and the research is clear on this. A benchmark called AECV-Bench, posted in January 2026 as a preprint that has not been peer reviewed, tested general purpose AI models on architectural drawings. They read text on drawings well, up to 0.95 accuracy. But counting doors and windows often landed between 0.40 and 0.55 accuracy, with large errors, and the authors wrote that "symbol-centric drawing understanding" remains "unsolved."
Pasting a plan sheet into a chatbot and asking for a fixture count is not a takeoff. The tools worth testing are the ones built for drawings, and they show their work.
What does the research say AI counts well?
AI counts well when the item is a repeated, clearly drawn symbol. The best independent test I have found is the Utica University study by Sadeh, Geloso and Todorov, published in June 2026 in the proceedings of the Associated Schools of Construction conference.
The team ran an AI takeoff tool, Togal AI, on a real 16,000 square foot commercial building on their own campus and compared it with the general contractor's bid takeoff, item by item, across 29 line items. Of the 13 count items, 11 matched exactly. All ten window types came out identical.
The largest data set comes from a vendor. Trimble reported in August 2026 that across more than 4,000 U.S. contractors using its AI takeoff features in the first half of 2026, estimators counted an average of 460,000 objects a month, with 60 percent time savings on counting even after a short review step. It also reported that 97 percent of drawing scales set automatically needed no manual change. That is vendor data with no accuracy figure published, so read it as a picture of use, not proof of accuracy. Trimble's own rule is worth repeating: "Nothing gets into a bid without a qualified person putting eyes on it."
Where the AI count drifted from the contractor’s takeoff
One 16,000 square foot building, 29 line items. Each bar shows the largest miss in its group.
Windows10 types, counted. All 10 matched exactly
0%
Doors3 types, counted. 1 exact, 1 at 9% low, 1 at 67% high
67% high
Floor finishes5 items, area. All low, 1% to 8%
8% low
Ceiling finishes5 items, area. All low, 3% to 11%
11% low
Exterior finishes5 items, area. 2% low to 13% high (stone)
13% high
Rubber base1 item, linear feet. About 2,700 ft against 12,978 ft
79% low
Repeated symbols were close to exact. Areas, odd shapes and linear runs are where the estimator’s eyes earn their keep.
Where does AI takeoff drift, and why?
AI drifts on areas, irregular shapes, linear runs, and symbols that look alike. In the Utica study, 12 of the 15 area items came in lower than the contractor's numbers. Floor finishes ran 1 to 8 percent low. Ceiling finishes ran 3 to 11 percent low. Stone on the exterior ran 13 percent high.
Even the counts had two misses. One door type came in at 20 where the contractor had 22. Another came in at 10 where the contractor had 6, which is 67 percent high. And the one linear item, rubber base, came in at about 2,700 feet against the contractor's 12,978, which is 79 percent low.
The authors explain why. For doors and windows, the tool uses symbol recognition and text detection. For areas, it has to draw a boundary around a shape, and offsets, material changes and irregular edges make that boundary harder to find. The misses were not random: they lined up with the kind of item being measured.
This study measured architectural finishes, not electrical or mechanical scope. My read for an MEP sub is that the same pattern points to where you should look: fixtures and devices that repeat across a floor are the easy part, and conduit, duct and pipe runs, plus anything that only appears in a note or a schedule, are the hard part.
What should the estimator still check on every bid?
Check by risk, not by habit. The estimator stops recounting the items the tool is good at and spends that time on the places it is known to miss.
1. The tool is usually right
Spot check
- Repeated symbols that match the legend: light fixtures, receptacles, diffusers, grilles
- Sheet scales set automatically
- Windows and doors drawn the same way across the set
2. The tool drifts here
Check every one
- Areas, and rooms with odd shapes or offsets
- Look-alike symbols, such as a fixture with an emergency battery or a different mounting
- Any count the tool marks as low confidence
3. The tool can miss it entirely
Do by hand or verify fully
- Linear runs: conduit, duct, pipe, base
- Items that live only in notes, schedules or the spec
- Anything on a revised sheet or addendum, and every scope line marked by others or existing to remain
Some calls never move to the tool at all. Is this item in our scope or by others? Is that fixture existing to remain? Which alternates do we price? What labor unit fits this building? Those are judgment calls, and they stay with your estimator. I wrote about where that line sits in Draw the Line: the machine does the repeatable part, and a person owns every decision with your name on it.
Getting the quantities right going in also only solves half of the number. The other half is keeping the prices behind them current, which is the subject of the quote that went stale in the drawer.
How would you test it on one bid before trusting it?
Run one bid both ways and compare the counts line by line. Pick a recent bid your estimator already counted by hand, ideally a job you won, so you also know what was actually installed. Run the same drawings through an AI takeoff tool. Then put the two counts side by side.
The one-bid test, side by side
Same drawings, counted both ways. Every mismatch gets an answer.
- Item
- Type A troffer
- AI count
- 148
- Estimator count
- 148
- Difference
- None
- Who was right, and why
- Matched. Move to spot check.
- Item
- Type A troffer with emergency battery
- AI count
- 6
- Estimator count
- 12
- Difference
- 6 low
- Who was right, and why
- Estimator. Look-alike symbol the tool read as a plain type A.
- Item
- Duplex receptacles
- AI count
- 212
- Estimator count
- 209
- Difference
- 3 high
- Who was right, and why
- Tool. The estimator missed a revised sheet.
- Item
- 3/4 inch EMT
- AI count
- Not counted
- Estimator count
- 2,400 ft
- Difference
- Missing
- Who was right, and why
- Estimator. Linear runs stay on the by-hand list.
For every line that does not match, write down who was right and why. After one bid you will know three things: which items the tool gets right on your kind of drawings, which ones it misses, and how much time the first count took each way. That list becomes your checklist for every bid after it.
This is the same idea as starting it by hand: prove the process on one real job before you change how every job is done.
What do you get back when counting stops eating the week?
You get your estimator's time back, and you get more bids out the door with the same people. When the first count takes hours instead of days, the hours go to the work that wins jobs: reading the scope, checking vendor quotes, setting labor units and writing clean qualifications.
For a sub whose bid volume is capped by one or two estimators, that is a revenue question. More bids reviewed properly means more chances to win the work you want.
The honest limits
The research is early and small. The Utica study is one building, one tool, and architectural finishes rather than MEP, and the researchers still had to review and correct the tool's markups. The Trimble figures are vendor data with no accuracy number. The AECV-Bench results are a preprint.
AI takeoff also does not know what it was not shown. It will miss an item that lives only in the spec, a note, or an addendum it never saw. It can be confidently wrong, and as the door count showed, it can overcount as well as undercount. Scanned or poor quality drawings make every number less reliable. A faster count is only useful if someone still checks the places it tends to miss.
Where ClearOak comes in
ClearOak & Co. is a software and operations partner for trades businesses. We start with how your estimators work today, not with software. I would sit down with your estimating team, look at a few recent bids, and find where the hours actually went and where the counts went wrong.
From there, we bring in the right builders to set up a takeoff tool that fits your drawings and your estimating software, run the one-bid test with your estimator, write the check-by-risk list for your scope, and train your team to own it. What you get back is your estimator's time, and more bids out the door.
If your estimators are still spending the bid week counting, schedule a call and we will look at one of your recent bids together.
Prefer email? Reach me at john@clearoakconsulting.com and tell me which count cost you the most time on your last bid.
Sources
- Testing AI Accuracy in Quantity Takeoff: A Methodological Case Study in Commercial Construction, Sadeh, Geloso and Todorov, Utica University, Proceedings of the Associated Schools of Construction 62nd Annual International Conference, EPiC Series in Built Environment vol. 7, pp. 753 to 762, June 2026 (single 16,000 square foot building, 29 line items, one tool)
- Actual results MEP estimators are getting with AI-assisted takeoff, Brooke Stewart, Trimble, August 7, 2026 (vendor data, more than 4,000 U.S. contractors, no accuracy figure published)
- AECV-Bench: Benchmarking Multimodal Models on Architectural and Engineering Drawings Understanding, Kondratenko, Birhane, Hsain and Maciocci, arXiv preprint, January 2026 (not peer reviewed)
- Occupational Outlook Handbook: Cost Estimators, U.S. Bureau of Labor Statistics, last modified August 27, 2026
Keep reading
- The 30-Day Clock in the Service Ticket
Since 2026, a refrigerant top-off can start a federal 30-day repair clock. The numbers are on your tickets, if someone reads them.
- The Permit Packet That Keeps Coming Back
Each permit correction cycle costs about a month. Cities now use AI to catch the repeat misses, and you can check first.
- The Spec Book Nobody Finished Reading
Subs build submittal registers by hand from 1,000-page specs. AI can draft the register in minutes so the PM checks a list instead.