The short version
- Togal.AI’s new benchmark tested general-purpose AI models on real takeoff tasks. They scored between 2% and 34%.
- Chatbots are good at reading text, not at measuring drawings to scale.
- Use tools built for takeoff, and use general AI for the work around the estimate: specs, scopes, bid emails and comparisons.
Every few weeks someone posts a demo of a chatbot “doing a takeoff.” In October, Togal.AI put that claim to a real test with a benchmark it calls Construction’s Last Exam.
What the test found
The benchmark gave frontier AI models from OpenAI, Anthropic, Google, xAI and others raw PDF drawings and asked them to perform takeoff tasks, 33 of them according to industry coverage. The models scored between 2% and 34%.
Worth knowing: Togal sells AI takeoff software, so it has a stake in showing that general chatbots fall short. That doesn’t make the result wrong. It matches what estimators see when they try it.
Want this working in your company? Book an AI discovery and we’ll show you where AI fits your office first.
Why chatbots struggle with takeoff
Takeoff isn’t reading. It’s measuring lines and areas at scale, counting symbols that look alike, and knowing which sheet overrides which. General AI models are trained mostly on text and photos. They can describe a floor plan, but they can’t reliably measure it.
Tools built for takeoff work differently. They’re trained on drawings, they read the scale, and they show you what they measured so a person can check it.
Where AI does help estimators
- Reading a spec book and listing the sections that apply to your trade.
- Turning a scope of work into a clean list of inclusions and exclusions.
- Comparing sub bids side by side and flagging what’s missing.
- Drafting bid emails, RFIs and clarifications.
- Using a dedicated AI takeoff tool, with a person checking the counts.
Questions
Can ChatGPT do my takeoffs?
Not reliably. In this benchmark the best general model scored 34%. Use a takeoff tool built for drawings and check its work.
Are AI takeoff tools accurate?
They’re much better than general chatbots, but you should still review what they measured, especially on complex or revised sets.
