pdf2takeoff: Free Construction Takeoff Software, and What AI Still Gets Wrong

If you price construction work, you know the routine. A tender lands with forty sheets of drawings. Somebody opens each one, sets the scale, and starts clicking: every wall run, every floor area, every downlight, power point and door. It is careful work, and most of it is repetitive.

pdf2takeoff is free, open-source construction takeoff software built by Flowtivity, an AI and automation practice on the Gold Coast. It runs in your browser, and your drawings never leave your machine. If you run a construction business, this post covers the problem it solves, what AI still gets wrong about counting floor plans, and every benchmark behind the numbers we publish.

Key takeaways for construction business owners

• pdf2takeoff is free takeoff software that runs in your browser. There is no licence fee, no account, and drawings are never uploaded.
• AI cannot yet count a floor plan on its own. The best of 29 published AI models gets 54.3% of floor plan counts exactly right, and far fewer on busy sheets.
• The approach that works is a division of labour. Software does the measuring, AI proposes each item, and an estimator confirms every number.
• Asking AI for a list beats asking for a total. In our tests, an itemised approach scored 58.5% against the best published 54.3%.
• Every number we publish can be checked. Our scores, including the weak ones, regenerate from the open repository.

What is actually slowing takeoff down?

The estimator is rarely the problem. The tools around them are.

There are four bottlenecks we kept running into.

• Seat licences add up. Bluebeam, the tool most estimators use, costs AU$451 to AU$979 per user per year, depending on the tier (Bluebeam's Australian pricing page, checked 16 October 2026). A team of five on the top tier is close to AU$5,000 a year before anyone measures anything.
• Tender drawings get uploaded. Many newer cloud takeoff tools need your PDFs on their servers. For confidential tender documents, that is a real question for the client, not just for IT.
• Counting means clicking. A single electrical layout can hold hundreds of symbols. Each one is a click, on every sheet, on every revision.
• AI totals you can't check. "Auto takeoff" features hand you a number. When it is wrong, there is nothing to look at except the number.

Where does this show up in your own week? For most teams it is the counting, done again every time a revision arrives.

What is pdf2takeoff?

pdf2takeoff turns construction PDF drawings into measured quantities. You open a drawing at pdf2takeoff.com, and everything happens in the browser tab: rendering, measuring and counting. There is no account, no file limit and no paid tier. The code is open source under the Apache-2.0 licence, which means builders and software teams can use it, change it and build it into their own tools.

A takeoff follows the same five steps every time.

Here is what that looks like on the sample drawing that ships with the site: a small office plan at 1:100.

1. Set the scale. Click both ends of the printed 12.00 m dimension. The built-in text reader (OCR, the same technology that turns a scanned letter into text) reads "12.00" off the sheet, so you do not have to type it.
2. Measure. Trace the external wall: 12.00 m. Click inside the office and Dynamic Fill traces the walls for you: 32.00 m².
3. Count. Click one downlight. The scan finds every matching symbol on the sheet in about a tenth of a second and marks each one as a proposal for you to confirm or delete. On the sample it finds all 8.
4. Export. Download an Excel workbook, or save the markups back into the PDF as standard annotations that open in Bluebeam Revu and Adobe Acrobat.

Beyond those basics, it handles per-page scales, regions drawn at a different scale, perimeter, volume and angle, formula columns for waste factors and rates, layers, notes, a printable report, and an overlay that shows what changed between two revisions of a sheet. Against Bluebeam's own feature list, our scorecard covers every Basics and Core tier feature and most of Complete.

Can't AI just do the takeoff?

This was the question we most wanted answered, so we looked at the evidence instead of the demos.

The best public test is AECV-Bench, from AECFoundry. It asks AI models to count five things on 120 real floor plans: doors, windows, rooms, bedrooms and toilets. Results for 29 models are published, including the latest from Google, Anthropic and OpenAI.

The best model, Gemini 3.5 Flash, gets 54.3% of all counts exactly right. That sounds close to useful until you look at what it gets right. Bedrooms are easy, at 93%, because the model only has to read the word "Bedroom". Doors and windows are hard: the best door score of any model is 43%, and the best window score is 39%. Rooms are the hardest of all: no model gets more than 27% of room counts right.

It gets worse as drawings get busier, which is exactly where a takeoff tool needs to help.

On plans with five doors or fewer, the best model counts doors exactly right two times in three. On plans with fifteen doors or more, it is right about one time in ten. Windows show the same drop, from 73% to 12%. (We grouped the 120 published plans by how many doors or windows each one has.)

That matches what we saw when we tested 10 AI vision models on real construction plans: the models were never the whole problem. The way the work is set up around them matters more.

Why does a single AI number go wrong?

When we dug into the failures, they fell into three groups. Only one of them is about eyesight.

• Busy sheets. Holding forty symbols in mind and producing one total is hard for a model, the same way it is hard for a person who is not allowed to make marks on the page.
• Rules, not facts. Is a double door one door or two? Is a walk-in robe a room? Does a sliding glass door count as a window? People in the same office answer these differently. A model has to guess which convention you use, and the benchmark's answer key has its own.
• Same plan, new answer. Ask the same model the same question twice and the total can change.

What happens if you ask AI for a list instead?

So we tried a different question. Instead of "how many doors?", we asked the model to list every door, window and room it can see, with its position and type. Then the software counts the list under written rules.

Using OpenAI's GPT-6.1 Sol with this approach, our first full run got 58.5% of counts exactly right across all 120 plans, against 54.3% for the best published model. Four separate runs with the same fixed counting rules (one of them at a higher reasoning setting) landed between 58.0% and 61.3%, so every one of them was above the published best.

Room counting rules turned out to be a matter of convention, so we also tested setting them from half the plans and scoring the other half. On those held-out plans we scored 62.8%, against 56.9% for Gemini 3.5 Flash on the same plans. That is a gap of about six points, and the statistical margin on it is wide (roughly one to eleven points), so we treat it as promising rather than settled.

The most useful finding was not a score at all. When three runs agreed on a count, it was right 78% of the time. When they disagreed, it was right 30% of the time. Flagging only the counts where runs disagree would point a reviewer at about two thirds of the errors while asking them to check just over a third of the counts.

Some honest caveats, because they matter:

• GPT-6.1 Sol is newer than any model in the published results. We have not yet run it on the benchmark's standard prompt, so we cannot yet say how much of the gain is the method and how much is the model.
• This is research, not a feature in the app yet.
• The benchmark's plans come mostly from CubiCasa5K, which is licensed for non-commercial use, so we publish numbers only, never images, and we do not train on them.

So who does what?

Putting it together gives a clear division of labour. Software is good at exact maths and tireless searching. People are good at judgement and accountability. Neither should do the other's job.

The rule we hold to is simple: no number without a marker. Every quantity in pdf2takeoff traces back to something on the drawing you can click, check and delete.

How do we test every number?

Takeoff tools are full of claims. We wanted ours to be checkable, so every score we publish regenerates from the open repository with one command, and a drop in any of them fails the build.

We use three benchmarks.

• ScaledDrawingBench, which we built because no public test checked measuring lengths and areas from scaled drawings. It has 200 synthetic floor-plan sheets, half clean digital drawings and half degraded scans, with about 2,000 estimator-style questions. We released it under CC BY 4.0, so anyone can use it, including to beat us.
• AEC-Geometric-Bench, an external test of finding walls, rooms and symbols on 15 real sheets.
• AECV-Bench, the external counting test described above.

Here is how the engine scores today, without any AI model.

• 99.8% on clean, born-digital sheets in ScaledDrawingBench.
• 37.3% on degraded scans. That is up from 23.0% when we started, and it is the area we are working on hardest.
• Zero made-up answers on 92 questions designed to have no answer. When the engine cannot read something, it says so.
• 0.519 wall-finding score on AEC-Geometric-Bench, about 1.8 times the best AI model tested (Claude Fable 5 at 0.287). A purpose-built commercial pipeline scores 0.931, so there is plenty of room above us. On whole-room areas, the best AI models still beat our engine.
• 0.295 on AECV-Bench with no AI model at all. Its bedroom and toilet counts come from reading the room labels with a dictionary and OCR.

We publish the weak numbers next to the strong ones on purpose. A score you cannot see the failures behind is just marketing.

What is coming next?

The next step is AI counting that works the way the research above suggests: the model proposes each door, window and room as a flagged marker, the software counts them under the counting rules you choose, and you confirm.

• Bring your own model. It will work with an OpenRouter key, which gives one key for hundreds of models, or with an OpenAI, Gemini or Z.ai key directly.
• Or keep it fully local. Point it at a vision model running on your own computer through Ollama or LM Studio, and nothing leaves the machine at all. If private AI is new to you, our private AI guide explains how it works.
• Off by default. When you turn it on, the sheet goes only to the provider you pick, using your own key. It never goes through us.
• Counting rules as a setting. Switch between the benchmark's rules and a schedule-style count (one per doorset, every window unit separate) and the totals update instantly, with no new AI call.

The first piece, which lets AI agents propose items for review, is in review on GitHub now.

What does this mean for a construction business?

For a managing director, the short answer is: lower software costs, drawings that stay private, and estimators who spend their time checking and pricing instead of clicking. The person who signs off the price does not change.

• Software spend. A takeoff seat costs AU$451 to AU$979 per estimator per year on Bluebeam. pdf2takeoff costs nothing, and because it is open source, there is no vendor who can raise the price later.
• Tender confidentiality. When a client asks where their drawings go, "they stay on our computers" is the simplest possible answer.
• Estimator time. Counting is the most repetitive part of a takeoff. When the software proposes and the estimator reviews, the same team can get through more drawings.
• Accountability. If a quantity is questioned after the job is won, you can show exactly which markers it came from.

Takeoff is one part of a bigger picture. The same pattern, where software does the repetitive work and an estimator makes the call, is how we build AI for construction businesses, from an agent that reviews every estimate before it goes out to safe work method statements in hours instead of days. If you want to see how an engagement like that runs, our construction AI case study walks through one from the first conversation to the rollout. And if you would rather just talk it through, you can book a free consult.

Old way vs new way

Try it: a 20-minute experiment

You do not need to change your process to test this. Try one small experiment.

1. Pick one sheet from a job you have already priced, ideally a lighting or power layout.
2. Open it at pdf2takeoff.com and set the scale from a printed dimension.
3. Click one symbol type and let the scan count it. Confirm or delete each proposal.
4. Compare the count and the time against your original takeoff.

What would you do with the hours you get back if every symbol on every revision counted itself, and you only had to check the ones it was unsure about?

If you want to go further, the same engine runs from the command line and as an MCP server, a standard way for AI agents to use tools, so an agent can run a takeoff and hand you a marked-up PDF to review. Our guide to AI for construction estimating covers where AI fits in the wider estimating workflow.

Takeoff does not have to be expensive, opaque or a week of clicking. The tools are free, the drawings stay yours, and every number can be checked. That is a good place to build from.