Use AI to speed counting and the first draft, then prove the output before it enters a cost plan: calibrate the scale against a known dimension, hand-check three or four quantities, and cross-check totals, units and rates in Excel. Treat anything outside your tolerance, say 5%, as unmeasured until a person has remeasured it.
AI take-off tools measure what they can read from a drawing: areas, lengths and counts. They do not know the measurement rules your contract names, what a drawing leaves implied, or whether the sheet in front of them is the current revision, and they return figures that look final. The work is therefore in the checks, and most of this tutorial is about those. Where I name tools, I have read the vendor's own page for the facts.
Where AI fits a QS workflow, and what each step needs checked
The same tool can be reliable at one stage and risky at the next, so attach the check to the stage rather than to the tool.
| Stage | What AI can do | Check before it moves on |
|---|---|---|
| Take-off | Count symbols; measure areas and lengths from a PDF or CAD export | Scale calibration, three or four hand-measured items, revision and sheet-count check |
| Item descriptions and grouping | Draft bill of quantities (BoQ) wording from take-off layers | Compare with the measurement rules your contract names; read every description |
| Rates | Match quantities to your own rate library | Rate within your own minimum and maximum; source recorded. Never rates from the model's memory |
| Cost plan | Assemble an elemental first draft from quantities and rates | Cost per square metre and element shares against your own past projects |
| Reporting | Draft commentary on variances | Every figure traced to the workbook |
Before you trust a take-off: calibrate, sample, compare revisions
Four quick steps stand between a tool's output and your workbook. None needs AI, and together they take perhaps 30 minutes on a modest drawing set.
- Calibrate the scale. Measure one printed dimension on the sheet, a 3,600 mm bay for example, with the tool. If it reads 3,510 mm, stop. The PDF has been printed to fit the page or the sheet is not to its stated scale, and every length from it will be out by the same proportion, with areas out by about double that.
- Sample by hand. Pick four items: a length, an area, a count and a derived quantity such as external wall area (length times height, less openings). Measure them yourself first, then look at the tool's figure. Looking first anchors you to its answer.
- Apply tolerance bands. As a starting rule: within 2% accept, 2% to 5% investigate, above 5% or any count error remeasure the whole element. Adjust the bands after the accuracy test described later.
- Confirm the revision. Check the title block against the drawing register and count the sheets in the set. A take-off tool will measure a superseded sheet as willingly as the current one, and most cannot tell you a sheet is missing; a few drawing-review features now claim to, but the drawing register is still the check.
Mixed scales are the commonest trap. One sheet often carries a plan at 1:100 and enlarged details at 1:20, and a tool that assumes the sheet scale will measure the lengths on a detail out by a factor of five. Calibrate each view you measure from, not just each sheet.
Schedules deserve the same suspicion. If the door and window schedules disagree with the counted symbols, one of them is wrong, and a person should find out which before pricing. What AI can and cannot take off goes through the categories that most often need hand measurement.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowA worked example: a two-storey extension, tool against hand check
Take a two-storey extension where an AI tool has produced the first-pass quantities and you have measured six items yourself. The comparison below is illustrative; the layout is the one to copy.
| Item | Tool | Hand check | Difference | Result |
|---|---|---|---|---|
| Ground floor slab area | 48.2 m² | 48.0 m² | +0.4% | OK |
| Floor finishes area | 90.7 m² | 91.0 m² | -0.3% | OK |
| Internal partition length | 62.4 m | 60.1 m | +3.8% | INVESTIGATE |
| External wall length | 31.6 m | 29.2 m | +8.2% | REMEASURE |
| Windows | 7 | 6 | +16.7% | REMEASURE |
| Roof area, sloping | 56.8 m² | 61.5 m² | -7.6% | REMEASURE |
In Excel the result column is one formula, with the tool's figure in column B and yours in column C, copied down:
=IF(ABS(B2-C2)/C2>0.05,"REMEASURE",IF(ABS(B2-C2)/C2>0.02,"INVESTIGATE","OK"))
For counts, use an exact match instead:
=IF(B2<>C2,"REMEASURE","OK")Each flag had a findable cause. The partition overshoot was a 2.3 m wall drawn on both the plan and an enlarged detail sheet and measured twice; removing it brought the tool to 60.1 m. The external wall figure included a 2.4 m boundary return that the drawings show as existing. One window was drawn on the plan and again on the elevation, and counted twice. The roof figure was plan area where the cost plan needs sloping area, a pitch factor of about 1.083 on 56.8 m² giving the 61.5 m² you measured.
Now price the errors, using the rates assumed in the cost plan: $190 per m² of external wall, $240 per m² of roof and $850 per window. The extra 2.4 m of wall at a 5.6 m height is 13.44 m², or $2,553.60 too much. The extra window is $850 too much. The missing 4.7 m² of roof is $1,128 too little. The net error is $2,275.60 over, which looks tolerable, but the gross error ignoring direction is $4,531.60. On a $168,000 package that is 1.4% net and 2.7% gross, and errors that cancel this week will not cancel next week.
BoQ cross-checks in Excel: six tests that catch most errors
Whatever produced the bill of quantities, a handful of columns of formulas will catch most mechanical errors before a client or contractor does. Put them beside the BoQ so a flag is impossible to miss. A BoQ template checked with AI shows the layout these formulas sit in.
| Test | Formula | What a flag means |
|---|---|---|
| Extension | =ROUND(D2*E2,2)-F2 | Quantity times rate does not equal the amount: someone typed over a formula or changed a quantity after pricing |
| Section sum | =SUM(F2:F40)-F41 | A subtotal is hard-coded, or a row sits outside the summed range |
| Unit list | =IF(ISNA(MATCH(C2,Units!$A$2:$A$30,0)),"UNIT?","") | The unit is not on your approved list, so the quantity may be in the wrong basis |
| Rate range | =IF(OR(E2<G2,E2>H2),"RATE?","") | The rate is outside the minimum (column G) and maximum (column H) in your rate library |
| Duplicates | =IF(COUNTIFS($B$2:$B$300,B2,$C$2:$C$300,C2)>1,"DUP?","") | Same description and unit appear twice: double-measured or double-listed |
| Blank or zero | =IF(OR(D2="",D2=0),"QTY?","") | An item carries no quantity, so a measurement was dropped |
A flagged row looks like this: "Blockwork inner leaf, 142.5 m², rate $48.00, amount $6,480.00". The extension test returns 360, because 142.5 x $48.00 is $6,840.00 and the amount says $6,480.00. The amount still reflects 135.0 m², the quantity before someone revised the take-off. Without the test, the BoQ would have gone out $360 light on one line and the same slip on twenty lines would be a pricing hole.
If you are unsure how to build any of these, an assistant is a sensible helper for the formulas themselves, and the ChatGPT guide for Excel covers the prompts. The formulas are then checked by pasting known good and known bad rows through them, which takes two minutes and is the only test that matters.
A second reader for the BoQ: a prompt that only asks questions
Excel catches arithmetic. It cannot tell that a description is ambiguous or that something usually priced alongside blockwork is missing. An assistant is a reasonable second reader for that, provided it asks questions and changes nothing.
You are a second reader for a bill of quantities extract. Use only the extract below. Do not change any quantity, unit or rate, and do not add items.
List as questions for the surveyor:
1. Items whose unit looks wrong for the description.
2. Descriptions that leave out a thickness, grade, finish or location.
3. Pairs of items that may be duplicates.
4. Work that is normally priced with an item in this extract but does not appear in it.
Number each question and quote the item reference.
EXTRACT
[paste rows: ref, description, unit, quantity]On a 40-row extract, illustrative output might include: "Item 4.12, blockwork inner leaf, is in m² with no thickness stated; what thickness?"; "Items 4.12 and 4.19 share a description and unit; is 4.19 a duplicate or a different location?"; "Item 5.03, insulation, is measured in nr; should the unit be m²?"; and "No damp-proof course appears alongside the blockwork items; is it measured elsewhere?" Answer each question in the workbook, not in the chat, and drop any you cannot justify. The fourth kind is the most useful and the least reliable, because the assistant may ask for items your project does not need, so treat it as a prompt for your own judgement and never as a list to add.
Cost plan checks: ratios from your own projects, not the model's
A cost plan assembled by an assistant from quantities and your rate library is a legitimate first draft. The rates must come from your library; turning supplier quotes into a rate sheet shows how to build one. The test of the draft is a reasonableness pass against jobs you have actually priced.
Say the draft totals $412,000 for 1,180 m² of gross internal floor area, which is $349 per m². Your last five similar jobs ranged from $410 to $455 per m² (invented for the example). The draft sits at least $60 per m² below your lowest, and 1,180 x $410 is $483,800, so it is at least $71,800 short. Break it down by element to find where:
| Element | Draft | Share of draft | Your usual share (illustrative) |
|---|---|---|---|
| Substructure | $38,000 | 9.2% | 8% to 11% |
| Superstructure | $131,000 | 31.8% | 28% to 34% |
| Finishes | $61,000 | 14.8% | 12% to 16% |
| Services | $79,000 | 19.2% | 20% to 25% |
| External works | $0 | 0% | 4% to 7% |
| Preliminaries | $66,000 | 16.0% | 12% to 16% |
| Contingency | $37,000 | 9.0% | 5% to 10% |
External works at $0 is the obvious hole: nobody had measured them, so the draft contained nothing. Adding them at a typical 5% would add roughly $22,000 and move the cost per m² towards your usual range, which tells you the gap is partly a missing element rather than a pricing error. Services at 19.2% is a smaller prompt to look again. Preliminaries, contingency and provisional sums are judgement calls tied to your risk allocation and the client's procurement route, so set them yourself and do not accept a figure the assistant volunteers.
Tools: what the vendor pages say, and what they leave out
I read each vendor's own pricing or product page in November 2026. Prices are USD list prices and change, and the right-hand column matters as much as the left.
| Tool | Price as published | AI features named | What the page does not say |
|---|---|---|---|
| Kreo | Essential $100 (no AI), Pro $175 for 1 to 4 users, Business $250 from five seats (at least $15,000 a year); per user per month, billed annually; 7-day free trials | AI scale detection, one-click room areas and wall lengths with AI completing the outline, Auto Count for repeated symbols, batch AI take-off across a sheet set (Pro and above) | No quantified accuracy figure |
| Togal.AI | Growth $299 per user per month billed yearly; Business by quotation | "Unlimited automated takeoffs" and symbol searches | Any accuracy data; "Try Now" books a demo rather than a self-serve trial |
| Bluebeam Revu | Basics $260, Core $330, Complete $440, Max $590 per user per year (Max labelled introductory) | Listed under Max only: AI drawing reviews (Smart Review checks such as missing sheets, missing door or plumbing fixture tags and gridline consistency), AI drawing comparisons (Smart Overlay), MCP integration and multi-view stitching | Any accuracy measure; none of these features measures quantities for you |
| STACK | Takeoff and Estimate "as low as" $299 per user per month, with a free version | "AI Takeoff" and "STACK IQ" appear in the site menu | What they do or which plan includes them |
| PlanSwift | Essential $2,000 and Core $3,000 a year for one seat; 14-day trial | Auto Count, Auto Scale and Auto Bookmark (Essential); Auto Takeoff first pass (Core) | Accuracy |
| iBeam (Beam AI) | Done-for-you take-offs from $25,000 a year (by trade and volume); do-it-yourself licence from $8,000 a year per trade, for HVAC, plumbing and steel only | Reads drawings and specifications; returns QA-checked take-offs in 24 to 72 hours as structured Excel outputs | The basis of its "99%+" claim |
The two accuracy statements I found are worth reading closely. iBeam says "99%+ accurate, QA-checked" and that every take-off is QA-checked before delivery, so the figure includes a checking step on the vendor's side, not the raw output of the machine. Togal's FAQ calls its product "the most accurate estimating solution on the market" without data. Neither tells you how accuracy was measured, on which drawings, or for which elements. Treat such figures as vendor claims and substitute your own test, described next. For a map of which tool type suits which job, see five tool types compared by job.
Do the licences pay for themselves? A break-even in hours
With published prices you can work out how many hours a tool must save before it earns its keep. Assume an internal cost of $55 an hour (replace it with yours) and divide the annual price by it:
| Tool and plan | Annual cost per user | Break-even hours a year |
|---|---|---|
| Kreo Pro | $175 x 12 = $2,100 | 38 |
| PlanSwift Core | $3,000 | 55 |
| Togal.AI Growth | $299 x 12 = $3,588 | 65 |
| STACK Takeoff and Estimate, lowest listed price | $299 x 12 = $3,588 | 65 |
| Bluebeam Revu Max (a wider tool, with AI review) | $590 | 11 |
The hours that count are net of checking. Suppose a practice does 40 take-offs a year at 18 hours each by hand. With an AI first pass of 5 hours plus 4 hours of checking (the calibration, hand samples and Excel tests above, with time to fix what they flag), each job takes 9 hours, saving 9 hours. At that rate Kreo Pro's 38 break-even hours are passed on the fifth take-off, and the 40 jobs save 360 hours in all. The timings are assumptions, so replace them with the ones from your accuracy test; if the checks take as long as measuring by hand, the sum comes out at zero and you have learned that cheaply.
Run your own accuracy test on three finished projects
The test takes about a day per tool and turns a vendor's claim into your own tolerance rule. Choose three completed jobs of the kind you do most, where the final measured quantities are known. Run each through the tool without looking at your numbers, then record the percentage error by element. An illustrative result:
| Element | Project 1 | Project 2 | Project 3 | Worst case |
|---|---|---|---|---|
| Slab area | +0.4% | -0.8% | +1.1% | 1.1% |
| External wall length | +8.2% | +3.1% | +5.6% | 8.2% |
| Windows and doors (miscounts) | 1 | 0 | 2 | 2 |
| Roof area | -7.6% | -2.2% | -4.0% | 7.6% |
The worst-case column writes your rules. Slab area never exceeded 1.1%, so a one-in-four spot check is enough. Wall length reached 8.2% and roof area 7.6%, so both are measured by hand every time, with the tool's figure used only as a second opinion. Counts went wrong on two of three jobs, so every count is verified against the schedules. If a tool will not give you a usable rule for an element, take that element off the AI list and say so in your procedures. A small firm's test of AI take-offs on real drawings is a good companion for the practicalities.
A review sheet that makes sign-off quick
One page per take-off keeps the audit trail short and the reviewer's job to ten minutes. Record these and nothing else:
- Drawing set: sheet list and revision letters, with the date of issue.
- Tool and settings: product, version, and the scale calibration reading (for example "3,600 mm bay read as 3,598 mm").
- Hand-checked items: the four or more items, your figure, the tool's figure and the percentage difference.
- Flags and what you did: each INVESTIGATE or REMEASURE line, the cause found and the corrected quantity.
- Excel tests: all six clear, or the exceptions and why they stand.
- Judgement items: who set preliminaries, contingency and provisional sums, and on what basis.
- Sign-off: preparer, reviewer and date.
Add one standing rule: two REMEASURE results in the same element across two projects take that element off AI until the tool or the drawing practice changes. If your appointment terms or professional body expect you to tell the client that AI was used, say so plainly; whatever the rule, you stay responsible for the quantities, which is why the review sheet exists. If you would rather work out where the first pass belongs in your own process, that is the kind of job my AI implementation consultation covers.
Further reads
- AI Tools for Construction Estimating: Bid Checks After the Takeoff — Bid checks to run after the take-off is done.
- How Can AI Help Quantity Surveyors? BoQ Checks, Takeoffs and Cost Reports — The wider view: BoQ checks, take-offs and cost reports.
- Can AI Check an Interim Valuation? A Quantity Surveyor's Test — Using the same checking habits on an interim valuation.
- Use of AI in Quantity Surveying: Comparing Subcontractor Quotes — Comparing subcontractor quotes line by line with AI help.
- ChatGPT for Construction Estimating: Check a Quote, Do Not Write It — Checking a quote with an assistant instead of writing it.
- AI Take Off Software: Trial Three Tools on One Real Plan Set — How to trial three take-off tools on one real plan set.
- RICS AI Standard for Small Firms: Register, Assessment, Client Notice — Turn the effective RICS standard into a written pre-use assessment, two practical registers and client wording that can be checked.
- Using AI in Surveying: Quotes, Reports and Client Updates — Quote templates, report first drafts and status emails for a small land surveying firm, with the measurements left to the surveyor.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: vendor pricing and product pages read in November 2026 (Kreo, Togal.AI, Bluebeam, STACK, PlanSwift, iBeam). Prices are USD list prices and change; accuracy statements are vendor claims. The worked figures are illustrative.