How to Use AI in Construction Estimating Without Losing Control of the Bid A construction estimate is a commitment: we can build this scope for this price, in this time, from documents we’ve often only had for a couple of weeks. That’s why handing the whole bid to AI and hoping for the best is a terrible idea. I think there are three sensible levels: - Start with risk reduction. Use AI to check your estimate for missed requirements, gaps, double-counting and formula errors. - Use it to structure information. Build requirements and clarifications registers, then review them before they feed into the next step. - Once your data is organised, use it for more of the heavy lifting like assemblies, line-item pricing, subcontractor quote comparisons and submission documents. The important part is keeping human review between each workflow so errors don’t compound. Where are you currently using AI in your estimating process?

I hear a lot of confusion about the term “AI agent” in construction. To me, an agent pulls together a few different concepts: - Workflows that explain how to complete repeatable tasks - Context from your project documents, registers and correspondence - Memory about the project, your business and how you prefer to work - Tools and permissions that control what it can access and what actions it can take We then package this into an interface that we or our team can call when we need work done. The main difference between a workflow and an agent is responsibility. A workflow completes a defined task. An agent owns a narrow area of responsibility, chooses which workflows to run, checks the result and keeps a record of what it has done. Would I let an AI agent autonomously build a programme or send contractual notices? No. But I would let it monitor correspondence, identify a potential variation, update the register and draft a notice for review. That is where agents are useful today: bounded responsibilities, clear permissions and human approval. I explain the full model here: https://youtu.be/hnGixjVMD2g

Registers are often the first thing to slip when things get busy. That was my experience as a PM, and an out of date register is essentially useless. Scheduled tasks are great for making sure they're kept current. I don't schedule a prompt, I schedule a skill. The instruction in the task is one line, run the document controller, and the skill carries all the detail. Three things to work out before you build one: - Where the information comes from. Project inbox, registers, daily site reports, meeting minutes. - Where it ends up, and whether you review it first or it writes straight to the register. - How often it runs, because when you open the register you need to know what window it covers. Then run it manually for a few days. Once you stop tweaking it, that is when it goes on a schedule. You still want to check in on the outputs though. I went into more detail in this video.

Which model is best at reading and understanding construction drawings? @Josh Turner and I just launched an AI benchmark for construction tasks. If you've not come across a benchmark, it's a way to score AI models. There's benchmarks for software engineering, legal, maths, etc. For our benchmark, we went through multiple sets of large construction drawings and created a database of questions and answers (Very time-consuming). Questions like "how many footings are in the slab" or "what are the cable tray specifications. We then scored each model based on what percentage of questions they answered correctly and the cost of the analysis. So far, we've tested 9 recent mods. GPT 5.6 won. Surprisingly, Opus 5 and 4.8 both beat Fable 5. One caveat: we tested the models via the API, so you would likely get different performance (almost certainly better results) if you used the harness (Claude Code, Codex, Cursor, etc.) We'll be running and updating this for every new model that comes out. We will also be adding more domains to our analysis - take-offs, estimating, scheduling, etc. We've got come up with some pretty interesting ways to test them. Check out the results here: https://contractoros.build/benchmarks/

Hey everybody, thought I would kick off a new post for August. What are your goals for this month and what are you trying to get out of the community? Mine are pretty simple and straightforward 1. Keep experimenting with AI and sharing what I'm learning 2. Keep making a ton of youtube videos 3. Keep learning about construction project management and working out the best way to deliver projects and run a construction business Keen to hear what you're goals are and what you want to get out of the community