Hi everyone I'm Ameet, I'm not a contractor. I'm a builder. I've spent the last few months talking to subs and GCs about one specific problem: why subs keep getting held up on mobilization because their insurance certificate bounces. Here's what I keep hearing: - Subs send the cert. GC says it's wrong. Nobody explains why. - The sub asks their broker. Broker says it's fine. GC says it's not. - This goes back and forth for weeks. The sub loses money. The GC loses time. I'm building an AI tool that reads the GC's requirements, checks the cert before the sub sends it, and tells them: ✅ This will pass ⚠️ Fix this first (with plain English instructions) 🚫 This won't pass — escalate now I'm not here to sell anything. I'm here to make sure I'm solving the right problem before I build the wrong thing. For the subs here: what's the most frustrating part of the COI process? What costs you the most? For the GCs: what do you wish subs would just get right the first time? Any feedback would be massively helpful.

Hi All, Has anyone set up both inbound and outbound calling for their agent? If so, how is that handled, in the prompt or through some other manner? Use case: inbound calls classify caller as prospect, current client, vendor/supplier, field technician or urgent caller and then routes call to proper workflow (get info, take message, schedule appt., forward call, etc.), updates crm records etc. If field technician, records info on job status, updates job and creates task list. Outbound: calls suppliers back, makes call based on tasks created by field technician such as calls supplier to place order, calls sub contractor to schedule work, etc.

I believe AI is going to fundamentally change how we plan construction projects. The recent advances of models like Opus 5.5 and GPT-6, particularly at interpreting construction drawings and building 3D models, equips us with new capabilities. What you used to only be able to do with a team of BIM modelers and computer programmer's. The capabilities that were reserved for billion dollar projects. Is it perfect? No. But how quickly and cheaply we can do things unlocks a wave of new planning applications. Want to create 20 different simulations of how a crane is installing structural steel? Sure, no worries! Want to simulate 10 different crane size options and determine where it has to set up and pack up? Easy! I feel like I’ve only scratched the surface of this topic. There is so much more to come. This week I’m going to talk about 5D planning and the unified work breakdown structure. As always, it will be structured around News, Inform Work. News Gemini 4 Argon: good at counting, weaker at the rest. I ran Google's new model through my construction tests. It scored 52% on the take-off test, level with Opus 5.5 at about 53%. Its estimate, procurement and programme outputs were weaker.. Watch the Gemini 4 Argon test ChatGPT dots: a personal agent with its own computer. OpenAI's dot runs in the background on a virtual computer, so your laptop does not need to be on. You message it, or call it like you would a person. It can also use your ChatGPT and Codex skills, so you can start a workflow from your phone. Watch the ChatGPT dots walkthrough Inform: what the evidence says Last week I covered two academic studies on BIM, and their results were mixed. Here is the contractor side, and where AI is up to. - Contractors report less rework. In McGraw Hill Construction's 2012 North American study, 74% of contractors were using BIM. They ranked reduced rework and fewer document errors and omissions among its top benefits. In the 2014 global study, around half of the contractors most heavily engaged with BIM reported a return above 25% on their investment. These are surveys of BIM users, so they are self-reported and likely to favour BIM. - 4D planning on one project. A 2026 case study of a dormitory project in Indonesia used a 4D model to reschedule the work. The planned duration dropped from 133 to 116 days. That is one project, but it shows the mechanism: when you can see the sequence, you can test better ones. - AI is improving but still unreliable. AECV-Bench (January 2026) tested leading models on 120 floor plans. They read text and answered document questions with up to 95% accuracy, but counted doors and windows correctly only 40–55% of the time. BIM-Edit (June 2026) asked models to make 324 edits to IFC building models. The best model averaged 49.5%, and none fully completed more than 3.4% of the tasks. On my own take-off benchmark, Opus 5.5 leads at 53.1%, with GPT-6 Astra at 46.2%.

You've probably heard some variant of the saying "what gets measured, get's managed" I don't think that's completely right. It should be, "what get's measured and communicated, get's managed" A construction dashboard should help someone make a decision. My recommendation is to start with who will use it and what they need to know. A project team needs upcoming milestones, cost overruns and risks requiring action. A client overseeing several projects may also need a simple scope summary. I would build it around three areas: Cost: Are we forecast to finish within budget? Where are variations accumulating, and when are payments due? Risk: Which risks need attention? How much will our planned treatments reduce their likelihood or impact? Programme: Which milestones are approaching, and where have dates slipped against the original plan? The information behind it needs to be structured and maintained. Build updates into your existing forecasting, payment and reporting processes so the dashboard stays current without duplicating administration. Then use it in your project reviews. If the same question keeps sending you back into a spreadsheet, that information probably belongs on the dashboard.

I did a quick video summarizing my testing and experimenting on it: https://youtu.be/9iOHzkDl5zc?si=4lwFgt46Hae8-Mfy Here are the benchmark results: https://contractoros.build/benchmarks/ Gemini 4 Argon came close to Claude on my construction take-off tests, but struggled with the management work. It scored 52% on my take-off benchmark, compared with 53% for Claude Opus 5.5, across 130 questions from 10 drawing sets. I would treat that as effectively equal. Gemini was stronger at counting items and weaker at measuring lengths and areas. The way you use the model also matters. For drawing questions, the raw Gemini model ranked in the top three. Running it through the Gemini app connected to Google Drive dropped it to 10th. On estimating, procurement and scheduling, it made some basic mistakes: - Changed a quoted 16-week steel lead time to seven weeks. - Used the client sell price, including profit, as the cost budget. - Claimed files were finished when the output folders were empty. My view: promising for reading drawings, but I would keep Claude or ChatGPT for more complex construction tasks. These results used the Gemini app with Google Drive. I have not yet repeated the tests in Google’s Antigravity coding tool.