AI construction estimating: a builder’s guide
AI estimating splits into two halves that get sold as one. The takeoff half reads a drawing set and produces quantities; the pricing half turns those quantities into money. The takeoff half is largely solved and is the part every vendor demonstrates. The pricing half is where estimates are actually won and lost, and it depends entirely on whose rates are used — yours, built from jobs you have closed out, or an industry average that has never seen your workshop, your crew or your suppliers. A tool with a brilliant takeoff and a generic rate book will produce a fast, confident, wrong number.
Estimating is where AI has landed hardest in construction, and for a good reason: it is the one process in the business where the inputs are documents, the output is a number, and being slow costs you bids you never submitted. ServiceTitan's 2026 report put cost estimating at the top of contractor AI use at 24%, ahead of bid management at 22%.
It is also where the marketing is least useful, because almost all of it demonstrates the easy half.
The two halves, and which one is actually hard
| Takeoff | Pricing | |
|---|---|---|
| Input | The drawing set | Quantities + your history |
| Output | Quantities, areas, counts | A defensible number |
| What makes it hard | Drawing quality, revisions, scale | Whose rates, and how honest they are |
| State of the art | Largely solved, improving fast | Entirely dependent on your data |
| What vendors demo | This | Rarely this |
Every AI estimating demo you will see is a takeoff demo. A set goes in, coloured polygons appear over the joinery runs, a quantity table falls out. It is genuinely impressive and it is the half that does not decide whether you make money.
The pricing half decides that, and it has nothing to do with how good the model is.
Why the rate book beats the model
An estimate is quantities multiplied by rates. The model supplies quantities. The rates come from somewhere else, and there are only two somewheres.
An industry average. Fast to set up, wrong in a way you cannot see. It has never seen your workshop, your crew's productivity, your suppliers' pricing or the fact that you always run a tradesman and an apprentice rather than two tradesmen. It is an average of businesses that are not yours, including the ones that went broke.
Your own closed-out jobs. Slower to build, and the only version that is actually yours. A rate derived from three comparable jobs you have completed carries something an average cannot: a spread. You can see how tight the band is, and therefore how much you should trust it.
This is the part worth being stubborn about when you are being sold to. Ask where a rate came from. The good answer names jobs. The bad answer names a database.
There is a second-order effect here that matters more over time. A rate book built from your own closed-out work gets better every job, because each completed job is another observation and the band narrows. An average does not improve, ever. Two years in, the gap between the two approaches is not a rounding difference.
What "accurate" should mean
Accuracy claims in this category are close to meaningless — 95%, 98%, against what baseline, on whose drawings, measured by whom. Replace the question.
The thing that actually matters is traceability: can you click any number in the estimate and see where it came from?
- A quantity should name the sheet and the detail it was measured off.
- A rate should name the jobs it was derived from, and how wide the spread was.
- A labour figure should name the hourly rate and the productivity assumption, separately, so you can argue with either.
An estimate you can audit in ten seconds a line is one you can defend in a meeting. An estimate that is a list of confident numbers is one you have to re-do to trust, which means the software saved you nothing.
The checking does not go away
This is the part the time-saving claims quietly omit. AI shortens the measuring. It does not shorten the judgement, and on a real job the judgement is:
- What has the architect drawn that nobody has priced? (Access. Protection. Out-of-hours work. Making good.)
- What will each sub exclude, and who ends up wearing it?
- Which of these allowances is fiction?
- What did we get wrong on the last one of these?
None of that is in the drawings, so none of it is in the takeoff. Expect the total time on an estimate to fall by half to two-thirds, not to fall to zero. That is still an enormous number — it is the difference between pricing four jobs a month and pricing ten — but budget for the honest version so the first estimate does not feel like a failure.
The one thing that breaks estimates after the job starts
Worth flagging because it is invisible until it costs you: what happens to the estimate once you win.
On most systems, the estimate stays editable forever. Somebody adjusts a line in month three — reasonably, because the scope changed — and the estimate now silently disagrees with the budget, the schedule of values and the contract sum that were all built from it at award. Nothing errors. The numbers just stop matching, and nobody finds out until the final claim.
The correct behaviour is that the estimate freezes when the job is marked won. From that point a cost is either a variation the client pays for, or something the builder wears against the budget — and both are recorded as such. The estimate is the contract's basis and it does not move.
If you are evaluating tools, this is a five-minute test. Win a test job. Try to edit a priced line. If it lets you without comment, you have found a future argument with a client.
What to ask a vendor
A short list that separates the real ones fast:
- "Show me a takeoff off a real PDF — a bad one." Not their demo set. A late revision with a title block that does not match. This is what you will actually feed it.
- "Click that number and show me the drawing." Traceability, in one gesture.
- "Where did this rate come from?" If the answer is not "jobs you have closed out", ask what happens as you complete more work. If the answer is "nothing", the tool does not improve.
- "What happens to the estimate when I win the job?" See above.
- "What can the AI change without me?" Get the list. Dates, prices and payments should not be on it.
- "How do I get my data out?" Excel, CSV, and ideally a straight push to your accounting package. A rate book you cannot export is a hostage.
Where this is going
Two things are worth watching over the next year or two.
The first is that takeoff quality stops being a differentiator. It is converging fast, and within a couple of cycles every serious tool will read a drawing set competently. What will not converge is whose data does the pricing — that is a moat made of your own history, and it belongs to whoever bothered to capture it.
The second is that the estimate stops being a document and becomes the spine of the job. The value was never the spreadsheet; it was the thread from the estimate line to the committed order to the invoice to the claim. Once that thread exists, variance shows up while you can still do something about it, which — given what ASIC's insolvency figures say about how construction businesses actually fail — is the whole game.
The short version
Buy for the rate book and the traceability, not the takeoff demo. Assume the checking stays. Make sure the estimate freezes at award. And be suspicious of any accuracy number that cannot be clicked.
Base does the takeoff off the PDF the architect sent, prices it from a rate book built out of your own closed-out jobs, and shows you every line a rate change would touch before it touches any of them. The wider argument about where AI belongs on a job is in this article; the business case is here.
Common questions
- How accurate is AI construction estimating?
- Accuracy is the wrong question, because it is two questions. Measurement accuracy off a clean drawing set is high and improving. Price accuracy is a property of your rate book, not the software — if the rates come from a national average, the estimate will be wrong by whatever the gap is between that average and your actual costs, and no amount of model quality closes it. Ask about traceability instead of accuracy: can you click any number and see the drawing and the past jobs it came from.
- Do I need CAD files or a BIM model for AI takeoff?
- No, and you should be suspicious of any tool that requires them. The set you are actually sent by an architect is a PDF, usually late, often a revision behind. A takeoff that only works off a model is a takeoff that does not work on most of the jobs you price.
- How long does an AI estimate take?
- The first pass on a typical commercial fitout set takes minutes rather than the day or two of measuring it replaces. The honest total, including the checking that must still happen, is usually half to a third of the original. Vendors quote the first number. Budget for the second.
- What is a rate book and why does it matter more than the AI?
- A rate book is your own record of what things have cost you — unit rates, material costs, subcontractor prices, labour hourly rates — built from jobs you have actually closed out. It matters more than the model because the model only supplies quantities. The money comes from the rates. Two builders running the same AI over the same drawings will get different, correct answers if their rate books are honest.
- Will AI estimating replace estimators?
- It replaces measuring and typing, which is most of the hours. It does not touch the part that is actually hard: knowing what the drawings leave out, what a sub will exclude, and which allowance is going to get eaten. Firms that have adopted it well have not cut estimators; they have submitted more bids with the same ones.
Rou runs commercial fitout and shopfitting projects in Australia and built Base to run his own jobs — the estimating, the programme and the money — before it was a product. He writes about what actually changed on site, not what a vendor deck says should.
One system, tender to handover.
Base reads the drawing set, prices it from your own rate book, and keeps the estimate threaded to the budget, the claims and the closeout.
Read next
- AI in construction
AI in construction: what actually works on a live job
Where AI genuinely earns its place on a building job in 2026 — estimating, document review, chasing subbies — and the three places it should never be allowed near.
- Running the business
Where AI actually saves a builder money
The four places AI changes a builder’s numbers — bid volume, margin leakage, cash timing and admin drag — and how to work out what each one is actually worth to you.