Can AI Do a Construction Takeoff From Drawings? A Small Firm's Test

Sometimes, for some items. AI reads schedules and notes accurately and counts clearly drawn symbols well, but it can miscount crowded sheets and misjudge lengths and areas unless the scale is calibrated. A small contractor can price from it only after testing on a job already measured by hand, and keeping a manual check on the biggest cost lines.

The honest answer is therefore a method, not a yes or no. Vendors publish accuracy figures, and some are impressive, but they are the vendor's own measures on the vendor's chosen plans. Your plans are drawn by different offices, at different scales, with different symbols and varying neatness. The only accuracy that counts is the one you measure on your own drawings, against a count you trust.

Follow me on Instagram@sagnikteaches

The test below takes an afternoon and one plan set you already know well. What comes out is a single table that says, item type by item type, whether AI output can go into the price with a spot-check, needs a full recount, or stays manual. To decide which tools to put through it, start with the comparison of takeoff tool types by trade.

Connect on LinkedInSagnik Bhattacharya

Choosing the plan set that makes the test fair

Pick a job you have already priced and, ideally, built. You then know the real quantities from the site: the number of doors actually hung, the metres of partition actually erected. Failing that, pick a set you have measured by hand carefully and can re-measure.

Subscribe on YouTube@sagnikteaches

The set should be representative, not flattering. A clean, recent, single-storey set with a door schedule and a legend will make any tool look good. A mix is better: a clear set, a scruffy one scanned from paper, and one with several sheets at different scales. If you can only do one, use the kind you see most often.

Say your firm does small commercial fit-outs, and the test set is a ground-floor clinic: six sheets covering a floor plan, a reflected ceiling plan, a power and lighting plan, door and window schedules and two details. The numbers in this tutorial come from an imagined version of that set and are illustrative only. They show the shape of a result, not what any tool will do for you.

Step one: build the answer key first

Before you run any AI, produce a manual takeoff of the items you will test and lock it. Two people counting independently and reconciling the differences gives a stronger key than one. Record:

• Counts: doors, windows, sockets, light fittings, sprinkler heads, radiators, fire alarm devices.
• Linear items: partition length by type, skirting, cable tray, pipe runs.
• Areas: floor finishes by type, ceiling by type, wall finishes.
• Schedule items: anything stated in a table on the sheets.

Write down how long the manual takeoff took. The time saved is half of what you are trying to learn.

An illustrative answer key for the clinic: 18 doors, 9 windows, 74 double sockets, 52 light fittings, 88 m of partition, 139.6 square metres of floor. Manual time: 3 hours 40 minutes.

Step two: run the same plans three ways

A general chat assistant reading the sheets

Upload the PDFs to ChatGPT or Claude on a plan where you have turned off training, or on a business plan (for example ChatGPT Business at $20 per user per month billed annually, or Claude Team). Ask it to extract the door and window schedules into a table, then to count the symbols on the power plan. Tell it the sheet scale and give it one dimension you know so it can sanity-check. Expect it to be strongest on the schedules, middling on symbol counts and weakest on measured lengths and areas.

I will upload the power and lighting plan for a small clinic. Count only the double socket outlets, using the legend on the sheet for the symbol. For each room, give the count. Show the total. List any symbol you were unsure about and where on the sheet it is. Do not estimate: if you cannot see a symbol clearly, say so. Do not measure any lengths from this sheet.

A dedicated takeoff tool on a trial

Tools such as Togal.AI say they automatically detect, measure and compare directly from your drawings, and quote up to 98% accuracy on floor plans; that figure is the vendor's own claim, measured on drawings it chose. Togal's pricing page lists the Growth plan at $299 per user a month, billed yearly, and offers a demo rather than a stated free trial. Ask any vendor to run the demo on your test set, set the scale from a known dimension, and run the detection while you watch.

A hybrid: tool for areas, chat assistant for schedules

In practice, many firms end up here. The dedicated tool takes the areas and linear items; the chat assistant lifts schedules and notes into a spreadsheet; a person confirms the counts on the top cost lines. The test should include this combination because it is the realistic workflow.

Step three: score the results in one table

Put every item on one row, with the manual key, each method's figure, the difference and a pass or fail against a tolerance you set before the run. Here is an illustrative result for the clinic set. These figures are invented to show the pattern, not measured from any product.

Read it row by row. The schedule items pass for the chat assistant. The socket count misses by 13, which fails any sensible tolerance. The takeoff tool passes on floor area and misses on partitions by 5 m, a 5.7% error, and its socket count is out by 4. Neither method gets you to a priced estimate without human work, but together they cover the ground quickly. The aim is to find which row types each method can be trusted on.

Why sockets go missing and partitions grow

Crowded sheets and symbol look-alikes

A double socket, a data point and a floor box can look nearly the same at small size. Counting errors cluster in rooms with the most services, such as plant rooms and reception, and where text overlaps symbols. The answer is to ask for counts room by room, so an error in one room is visible, and to spot-check the busiest room by hand.

Scale and measurement

If the sheet is plotted at a different size from what its scale bar says, any measurement is out by the same ratio. This is why you give the tool one known dimension and check a second. On scanned or shrunk sheets, treat measured lengths as suspect until proved.

Things not drawn

Plans show what the designer drew. A takeoff of a plan cannot count items in the specification only, such as firestopping at penetrations, sleeves through walls or fixings. Whether the tool or the person does the count, those lines come from the specification and your experience.

Merged or duplicate detections

A wall shared by two rooms may be counted twice or once, depending on the tool, and a door drawn on two sheets may be listed twice. Reviewing the detected objects on the drawing, rather than only reading the totals, is the safeguard. This is the main advantage of dedicated tools over a chat window.

What the error costs once it reaches the price

Run a quick sum to decide how strict to be. Say the electrical sockets cost $145 each installed, and the AI count is 13 short at 61 against 74: that is 13 x $145 = $1,885 left out of the price, before overhead and margin. On a $180,000 fit-out that is about 1%, which a small contractor might absorb, but the same slip repeated across lights, data points and devices can add up to several percent. Margins in this work are often tighter than that. Put differently: an error that looks small as a count can be big as a share of your profit.

A second sum goes the other way. If manual counting takes 3 hours 40 minutes and the checked AI workflow takes 1 hour 30 minutes, you save about 2 hours 10 minutes per job. At 40 estimates a year, that is roughly 87 hours, or two working weeks, which is worth having as long as the checks stay in. If you skip the checks, the saving is false.

Setting tolerances that suit the item

1. Zero tolerance for schedule-driven items. Doors, windows, equipment: the schedule is text, so any difference is an error to find.
2. Small tolerance for counted items. One or two on a few dozen, or a few percent on hundreds. Anything beyond means a manual recount of that item.
3. A percentage for measured items. Two to three percent for areas, a little more for linear runs, set from what your pricing can absorb.
4. Manual check on the top five cost lines whatever the tool says, because the biggest lines hide the biggest errors.

Write your thresholds on the test sheet before you run the tool. Choosing them afterwards, to fit the result, defeats the purpose.

Three situations, three decisions

The same test gives different answers in different trades, because each trade prices from a different kind of item. Three illustrative outcomes:

A small joinery workshop pricing from architectural plans

Most of the value is in schedules (doors, ironmongery, frames), which the chat assistant lifts well. A person counts the unusual items. A plausible result: AI for schedules, manual counts for custom items, and perhaps an hour saved a job.

A flooring contractor with room-by-room finishes

Areas are the key figure, so the dedicated tool matters more than the chat assistant, and the room-by-room review is quick. A plausible result: trust areas within 2% after a calibrated scale check, and still measure irregular rooms by hand. The flooring-specific test goes further on this.

An electrician pricing a larger job

Symbol counts decide the price. Crowded plans caused misses, so the rule is room-by-room counts with a manual recount of the three busiest rooms. How electricians count fittings from drawings with AI has a prompt for this.

Lifting a schedule: a before and after

Schedules are the safest place to start, because they are text. Say the door schedule arrives as a PDF table that copies out as a jumble: "D07 900x2100 FD30S single leaf, hinges 3, closer Y, vision panel N, D08 ..." with the columns merged. Ask the assistant to rebuild it, and add a rule that stops it guessing.

Rebuild the door schedule from the text below as a table with columns: Door ref, Width, Height, Fire rating, Leaves, Closer (Y/N), Vision panel (Y/N), Notes. If a value is missing or ambiguous, write CHECK instead of guessing. At the end give the total number of doors and the number of doors with each fire rating.

An illustrative result: a clean 18-row table, two cells marked CHECK (a door where "FD30S" ran into the next reference), and totals of 18 doors, 6 with a fire rating and 12 without. What you fix: open the PDF and resolve the two CHECK cells yourself, then compare the total with the number of door tags on the floor plan. If the schedule says 18 and the plan shows 19 tags, you have found a drawing discrepancy worth a query to the designer, which is a useful by-product no tool volunteers.

Cross-checking the plan against the schedule

A cheap and powerful test is to make the sources disagree on purpose. Ask the assistant to list every door reference it can see on the floor plan, then compare that list with the schedule rows. Anything on one list and not the other is a discrepancy to run down. In the clinic example, the plan showed a door reference D19 that the schedule did not list, and the schedule listed D11 that the plan never showed. Neither is a counting error by the tool; both are real problems in the drawing set, found in minutes. Doing the same for window references, equipment tags and room names tends to turn up the unclear scope items that later become variations.

Air handling units counted once per sheet

Picture a small mechanical contractor running a takeoff tool on a set where the ground-floor plan and the services plan both show the same air handling units. The tool detects each unit on each sheet, and the total comes out at twice the real number. The error is caught only because the manual key says 4 units and the tool says 8, so the row fails its zero tolerance. Without an answer key the estimate would have gone out at double. Duplicates across sheets are a standard failure for any method, and a pass-or-fail table finds them at once. Put "same item on more than one sheet" on the list of things your review looks for.

A filled-in test record to copy

Keep each test on one page so the results can be compared and shown to a partner or a lender if they ask how you price. Here is a filled-in example, again illustrative:

• Job and sheet set: clinic fit-out, six sheets, mixed scales (1:50 and 1:100), recent CAD output.
• Methods tested: chat assistant on schedules and power plan; takeoff tool on floor plan.
• Scale calibration: set from a 3,600 mm dimension string on the floor plan, checked against a 2,400 mm door-to-door dimension; both within 1%.
• Passes: window schedule, floor area, light fittings within tolerance.
• Fails: sockets (13 under), partition length (5 m over), door count in tool (1 under).
• Causes found: sockets under-counted in two crowded rooms; partitions over-measured where a dual-sided wall was counted per side; one door drawn with an unusual symbol.
• Decision: use the assistant for schedules, the tool for areas; count sockets and partitions by hand; retest after ten jobs.
• Time: manual 3 h 40 min against 1 h 30 min for the combined workflow including review.

A page like this also makes the tool-buying decision easy. If a product fails on the rows that drive your price and the vendor cannot show you why, you have not lost money on a licence.

A routine for the first ten live jobs

Run both methods on the next ten estimates, with the manual takeoff as the price basis. Log the difference on every row type in a simple sheet: job, item, manual, AI, error, cause. By job six or seven you will see where the errors repeat. Then move one row type at a time to AI-with-check, starting with the most reliable (schedules), and only move the next type when the log supports it. Keep the log after you switch, adding a monthly spot-check on one job so a software update or a new drawing style does not catch you out.

When you move on to pricing, the takeoff is only the quantities. Bid checks after the takeoff covers rates, allowances and the final read-through. And if the quantities feed subcontractor quotes you receive, levelling those bids with AI uses the same habit of checking against the drawings.

When not to use AI on a takeoff

• Hand-drawn or very poor scans. Detection quality drops, and so does a chat model's reading.
• Heavily renovated buildings. Existing and new work on the same sheet confuses detection; a person separates them.
• One-off bespoke items. The unusual item is exactly what a tool has not seen.
• Tenders with strict measurement rules. If the contract specifies a method of measurement, follow that method by hand or with a tool configured for it.

The practical stance: treat AI output like a keen new estimator's first draft. It is quick and often right, it will occasionally be confidently wrong, and everything priced from it gets checked by someone who has seen the building go up. How quantity surveyors use AI on takeoffs and reports shows the same approach from the surveyor's side.

Questions about AI takeoffs for small contractors

How accurate must a takeoff be to price from?

That is your call, but set the tolerance before the test. Many small firms accept a few percent on high-volume items and none on unique or expensive ones. Count errors on doors, fixtures or devices hurt more than a small area difference because each item carries a fixed cost.

Can ChatGPT or Claude measure from a PDF plan?

They can read schedules, notes and dimension strings well. Measuring lengths and areas from a drawing is unreliable unless the scale is stated and calibrated, so check any measured figure against a known dimension on the same sheet.

Is a dedicated takeoff tool better than a chat assistant?

Dedicated tools measure on the drawing at a scale you set, so you can see what was counted where, which is the safer workflow. Vendors publish accuracy figures but these are their own claims, so run your own test on your own plans, ideally during a demo or trial.

Should I stop doing manual takeoffs?

Not at first. Run both on a handful of live jobs until the differences are small and predictable. Keep a manual check on the top few cost lines even afterwards, because those are where an error costs real money.

Further reads

• AI Take Off Software: Trial Three Tools on One Real Plan Set — Trial three takeoff tools side by side on one real plan set.
• AI Estimating Software for Plumbing: Can It Do a Takeoff From Drawings? — Takeoff from drawings for plumbing work.
• AI Drywall Estimating Software: Plan or Photo Takeoff for a Small Crew? — Plan or photo takeoff for a small drywall crew.
• Can AI Produce Accurate Estimates for a Small Building Firm? — Whether AI estimates hold up for a small building firm.
• Best AI for Quantity Surveyors: Five Tool Types Compared by Job — Five AI tool types for surveyors, compared by job.
• AI Software for Architectural Drawings: Notes, Schedules and Sheet Checks — Catch schedule and sheet errors in the drawings themselves.
• AI Estimating Software for HVAC: Replacement Quotes and Duct Takeoffs — Duct takeoff tools and replacement quote builders do different jobs. See what each does, what it misses, and how to test one on a finished duct job.
• AI Estimator for Contractors or a Human Estimator? How to Choose — A cost sum, a volume-and-complexity grid and a hybrid workflow for choosing between an AI estimating tool and a part-time estimator, with a named sign-off.
• How Much Does Estimating Software Cost? Price Bands With and Without AI — Price bands for a small contractor's estimating software, what AI adds plan by plan, and the break-even hours that justify the spend.
• Which AI Tools Should a General Contractor Buy? What to Skip — A buy-first order, a skip list with triggers, and a hours-it-must-save test for a general contractor, priced for a six-person remodelling firm.
• ChatGPT for Construction Estimating: Check a Quote, Do Not Write It — A worked check of a small extension quote: arithmetic, scope gaps and omissions found with ChatGPT, every change left to the estimator.
• Construction Estimating With AI: Turn Supplier Quotes Into a Rate Sheet — Turn supplier quote PDFs into checked rates for your estimates: the extraction prompt, unit traps, five tests, an addenda log and a scope-gap review.
• Data Cabling Quote: Count Drops, Cable Runs and Test Lines With AI — A contractor-side data cabling quote for a 12-drop office: drop schedule, cable metres, test and label lines, exclusions and a count check done with AI.
• Masonry Bidding: Review Plans and Specs for Scope Gaps With AI — A bid-review routine for masons: AI tables each requirement, flags who supplies and who installs, and turns gaps into RFIs, exclusions and clarifications.
• AI Software for Construction Drawings: Catch Changes Between Revisions — How to compare two construction drawing revisions with overlay and AI tools, read the change list, and catch what the software cannot see.
• Can AI Tools Read Construction Drawings? What to Check Afterwards — What AI gets right and wrong on a plan PDF, with a scale check, a schedule cross-check and a worked room-by-room example.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Togal.AI homepage and pricing page (accuracy figures are vendor claims); ChatGPT Business pricing page; general estimating practice.