Can AI Do a Construction Takeoff? An Honest Test

No, not the part that matters: AI cannot reliably measure or count quantities off construction drawings, and any tool promising otherwise deserves your skepticism. That is not the whole answer though. This post comes from my video on AI quantity takeoffs, where I build an LLM takeoff workflow on a real set of mechanical drawings and show where AI genuinely helps around the measuring. It is written for estimators and contractors who keep getting pitched AI takeoff tools.

Key takeaways

• AI models cannot reliably take scaled measurements off drawings. On ClockBench, a visual reasoning benchmark, humans read analog clocks at 90.7 percent accuracy. The best AI model manages about 50 percent.
• The measuring and counting stay human. AI sets up the count, estimates the count for early budgets, or checks the count. It does not do the counting.
• If you run an LLM takeoff anyway, work one drawing at a time and count from the extracted vector data, not from AI vision.
• The three honest use cases: conceptual estimating, building your list of items and assemblies, and cross-checking a finished bill of quantities.
• Dedicated AI takeoff tools are expensive and none of them claim 100 percent accuracy. If takeoff volume is your real problem, outsourcing is usually cheaper and more accurate.

What is an AI quantity takeoff?

An AI quantity takeoff is using a large language model like ChatGPT, Claude or Gemini, or a dedicated AI tool, to pull quantities from construction drawings: counts, lengths, areas and volumes, returned as a bill of quantities. The promise is that you upload the drawings and the quantities come back done. The reality is more limited, and understanding why starts with how these models actually see a drawing.

There are two routes people try. The first is a general model in a chat interface: upload the PDF set, prompt it with your takeoff instructions, hope for the best. You cannot customize much beyond the prompt, and the results reflect that. The second route is dedicated AI takeoff software. I have looked at several of the well-known tools, and the pricing I have seen is astronomically high. None of them claim to be 100 percent accurate, which matters, because a bill of quantities you cannot trust still has to be checked line by line. If your goal is to hand over a drawing set and get back a finished BOQ in your exact format, the honest recommendation is to outsource the takeoff to a human. It is usually cheaper than the AI software and more accurate.

Why can’t AI measure off construction drawings?

Because current AI models are weak at precise visual tasks. They can tell you a sheet is an architectural layout, but reading a scale and taking a measurement from a precise section of a drawing is a different job entirely, and models fail at far simpler versions of it.

The cleanest evidence is ClockBench, a benchmark built by Alek Safar that tests whether AI models can read analog clock faces, a task trivial for most people. Humans average 90.7 percent on the current leaderboard. The best model, GPT-5.4 at high reasoning, scores about 50 percent. Gemini 3.1 Pro, the model I use for analyzing construction drawings, scores 32.2 percent. Most recent models, including Claude Opus 4.6, score under 10 percent. When the benchmark was first released in September 2025, the best of the 11 models tested managed 13.3 percent against an 89.1 percent human average, so the models are improving fast, but the gap is still enormous. If a model cannot reliably read a clock, I am not letting it measure ductwork.

Part of the problem is how models process a PDF. A drawing set has two layers. The vector data is the encoded text, lines and symbols, effectively instructions that say draw this box at these coordinates. Then, to see the sheet visually, the model converts it to an image and breaks it into overlapping tiles, usually compressed below the resolution you would view in a PDF reader. Every look the model takes is a compressed snapshot of one tile, never the full sheet at full detail. That is fine for context. It is not fine for measurement.

Building a more accurate LLM takeoff workflow

Stop relying on AI vision. Split the set into single drawings, extract the vector data, re-tile at high resolution, and give the model a written takeoff workflow with a confidence level on every line. The model becomes the brain that decides when to run deterministic scripts, and the scripts do the mechanical work.

In the video I build this in Google Antigravity, a coding environment where Gemini can write and run its own scripts against a folder of drawings. You could build the same pattern with Claude for construction; the workflow matters more than the tool. Mine looks like this:

1. Split the PDF set into individual drawings. Models have a limited context window. Processing 40 sheets one at a time is far more accurate than processing 40 at once, and splitting a PDF is a job for a script, not for AI.
2. Re-tile at high resolution. Instead of a few big low-resolution tiles, a script breaks each sheet into many small high-resolution ones. It costs more tokens and it reads more accurately.
3. Extract the vector data. All the text, tags and symbols come out as plain text. On my set, the split system units are tagged CU5 on the layout plans, so every CU5 shows up in that extraction. Counting text is something these models do very accurately. Counting pixels is not.
4. Write the workflow file. Mine tells the model to never use AI vision when it can count tags or use the vector data, to extract primary quantities first and derive secondaries after, to go one drawing at a time, and to state a confidence level for every item it returns.
5. Scan, propose, then execute. The model reads the set and proposes what items to take off. I approve the plan, then it runs the takeoff and writes the bill of quantities to Excel in my format.
6. Verify everything. The human check is written into the workflow because it is part of the workflow, not an afterthought.

The first version will not be very good. You iterate, feed the errors back, and get it to fix itself. That is the honest shape of these builds.

Where does AI actually fit in a takeoff workflow?

Around the measuring, not in it. AI is useful in three spots: rough quantities for conceptual estimating, building the list of items and assemblies before you measure, and cross-checking the finished bill of quantities after you measure.

Takeoff step	AI fit	Why
Working out what to measure	Good	Reads the set and returns the item list, often one of the slowest parts of a takeoff
Setting up assemblies	Good	Drafts the secondary logic from your existing data; the derivation itself is deterministic
Conceptual estimating	Good, within limits	For an order-of-magnitude budget price, a missed quantity here or there is acceptable
Counting tagged items from vector data	Usable with checks	Text counting is reliable; every count still carries a confidence level and a human check
Measuring and scaling dimensions	No	Stays human, in your takeoff tool. This is the line
Cross-checking a finished BOQ	Good	You measure, then AI hunts for errors. You are not relying on it, you are using it as a second pass

The middle of that table is where takeoff software already shines without any AI: once you have measured the primaries, simple assemblies derive all your secondary quantities. Rebar from slab volume, cable and glands from light counts. There is no AI involved and no risk, it is just deterministic logic. The same boundary runs through the whole estimate, and my Claude for construction estimating workflow follows it end to end: I keep the understanding, the measuring and the final number, and the model does the indexing, populating and cross-checking.

Common mistakes with AI takeoff tools

The expensive mistakes are all versions of the same one: trusting AI with the measurement instead of the work around it.

• Buying the accuracy claim. No tool I have seen claims 100 percent accuracy, and anything under that means every line still gets checked by a human. You did not remove the work, you moved it.
• Paying software prices for a labour problem. If takeoff volume is genuinely swamping you, outsourcing the takeoff is often cheaper than the AI tools currently on the market, and more accurate. Most of what AI usefully does around a takeoff runs on tools you may already pay for, and a fair chunk of it costs nothing. The free AI for construction estimating breakdown covers what is genuinely usable at zero cost.
• Sending the whole set in one prompt. Context windows are real. One drawing at a time, every time.
• Using AI vision when the vector data has the answer. If the item has a text tag, count the tags.
• Skipping the confidence levels. Every item the model returns should state how sure it is, and a human verifies it before it goes anywhere near a price.
• Forgetting the time maths. Once the drawings are in front of you and you know what you are measuring, the counting genuinely is not that long. From the video:

“I just think AI quantity take-off tools don’t offer the time savings or the benefits that people think they do.”

Tim Fairley, in the source video

Watch the full video above to see the workflow built end to end, from the Antigravity setup and the vector extraction through to the finished bill of quantities. The skills from this workflow, and the contractors running them on real bids, are in the ContractorOS community.