Hey everybody, Realistically 2-3 people per week reach out to me about AI take-offs. Every video I do on the topic blows up on YouTube. It seems to be consistently the one thing that everybody wants. Maybe I'm strange, but I actually enjoy doing a take-off. I find it kinda relaxing. But unfortunately, most of the time you need to do one, it's one of 20 things you need to do. And it's probably the end of the day, so sitting measuring drawings is competing with going to the gym, watching Netflix or whatever else you do to relax. So I definitely understand the appeal. News: The take-off benchmarks are getting better. In the past I've been negative about AI take-offs, because I've tested them time and time again and been disappointed. The current figures are on the Quantity takeoff benchmarks. The test gives AI a set of drawings and gets it to prepare a BoQ. The best result was 59.2% (Claude Fable 5.2 in the Claude desktop app): 77 of 130 lines inside tolerance. On counts, 71% of 70. On measured lines, 41% of 51, scored inside ±5%. Every new model released climbs the benchmark. I think it’s only natural to assume this will eventually become a solved problem. Inform: What we’ve found improves take-offs There’s 5 things we’ve found improves quantity take-offs: 1. Drawing QA - Checking the drawings for size/scale, orientation, accuracy 2. Breaking the task into planning and doing. Identify what you need to measure then measure. 3. Using different approaches for counting, lengths, areas and complex geometry. I.e. complex geometry can be enhanced by getting the model to generate a 3D model first. 4. Adding in sanity checks. For example if the cable tray runs the length of the building it should roughly match the building length in the architectural drawings 5. Visual audit. We get AI to mark up a .html file with the counts. Both us and an independent sub-agent inspect this. See the skill produce a bill of quantities from a drawing set, and how I package it so you can run it again.