In this newsletter
This week’s developments
Source: AEC field notes
🏠 OpenAI’s Astra’s can 3D model: my thoughts
In summary: The internet is flooded with demos of OpenAI's latest model controlling 3D modelling software. The output isn't yet useful to AEC — but at this rate of improvement, AI may soon embed itself in workflows currently owned by qualified professionals.
The details:
OpenAI's Astra (GPT-6) can autonomously operate professional tools like Blender, Unreal Engine, Rhino, and Revit - software that takes young employees years to master
Today's demos are still primitive by AEC standards: convincing at first glance, but out of scale, riddled with nonsensical design decisions, non-compliant with building regulations, and far from coordinated
ChatGPT 3.5 launched in November 2022 and struggled badly with hallucinations. Given that pace of progress, GPT-6's successors becoming expert 3D modellers isn't far-fetched
Why it matters:
Today's shortcomings aren't insurmountable. Compliance, for instance, can be reduced to rules an AI can learn to check against - and specialized models for this are surely coming.
I can’t see Human relationships, liability, and accountability disappearing; construction is high-stakes, and law and insurance will keep requiring professional sign-off. But consider this: much of the built environment is simple, repeatable work, and the barrier to a first draft on these projects could drop to nearly zero.
If producing design information (the bulk of AEC firms' revenue) shrinks to reviewing and validating AI output, traditional fees shrink with it. Surviving that shift means changing how firms think about their value.
If this trajectory holds, the scarce resource won't be the ability to produce a design. It'll be the judgment, liability, relationships, and physical-world knowledge needed to turn that design into a building.
Source: AEC field notes
🚜 Bedrock's excavators run without operators onboard
In summary: Bedrock Robotics just confirmed its excavators are running fully autonomous on three live U.S. construction sites, two years after founding and only months after closing a $270M round at a reported $1.75B valuation.
The details:
The jobs are a Nevada water-treatment plant with Sundt, a Texas earthwork site with Champion Site Prep and a civil job for Zachry
The Bedrock Operator bolts a sensor-and-compute kit onto machines a contractor already owns, installed in a single day with no permanent changes
Bedrock has raised $350M total, including a $270M round in February 2026 at a reported $1.75B valuation, and now counts 15 partners across 23 states
Training data comes from contractors' own sites and workflows, built up over months of supervised deployments before this milestone
Why it matters: Contractors today have access today to an autonomy system without having to develop one themselves. Plus, the have the ability to use it on highly repetitive work, potential additional operating hours, and crucially an opportunity to keep their scarce experienced operators available for work that actually requires them.
AEC's likely new hire: fleet manager / autonomy supervisor, with one person overseeing several semi-autonomous machines rather than driving one.
Source: AEC field notes
🥽 NavigateAI whispers construction workers
In summary: Ex-Opendoor CEO Eric Wu just raised $25M for NavigateAI, a copilot that answers a construction crew's questions through a phone or Meta's glasses, and records every job it coaches as labelled video shot from the worker's eyes.
The details:
Using NavigateAI, a worker can point the camera at what they’re building and the system can retrieve, for example, building specs, manufacturer manuals, or company policies in real time
Every job also records labelled first-person video, which Wu believes will be worth as much to robotics firms as the software is to his contractors
NavigateAI aims to take roughly 20% of the value created. Lennar (contractors), an investor as well as a customer, spends about $9B a year on labour and installation, where a 5% gain is worth hundreds of millions
Why it matters: Tech like NavigateAI has the potential to reduce errors on site and let less experienced crews do specialized work under expert guidance. It also attacks the labour shortage from two directions: it makes junior workers productive faster, and it generates the training data needed to eventually put robots on a select amount of low-value repetitive tasks.
Source: AEC field notes
✏️ Increase your productivity in Microsoft 365
In summary: OpenAI and Anthropic have plugins for Microsoft 365 which can read and edit your files, negating the need to cut & paste or upload the documents you are working on manually - which I used to do all the time until I started using these plugins.
The details:
The key productivity gain is that you can build, analyse, summarise, and update documents without switching tools, and in plain language simply by installing the dedicated plug-ins
You own skills you’ve designed to tackle repetitive tasks are accessible via these plug-ins, and can be called as usual (using / in Claude, and @ in ChatGPT)
Getting Started:
Both OpenAI and Anthropic plugins are installed through the Microsoft Marketplace/AppSource: e.g. in Excel go to Home (or Insert) → Add-ins → search "Claude" or "ChatGPT" → add, then sign in with the relevant account.
Note, as of September 2026:
Anthropic (Claude): has plugins for Word, Excel, and Powerpoint
OpenAI (ChatGPT): has plugins for Excel and Powerpoint
Pro tip: Separately, Anthropic also ships a Microsoft connector (not the same as the Office add-ins) as part of Claude for Small Business, letting Claude read/act on Outlook, Word, Excel, Teams transcripts, SharePoint, and OneDrive from inside any Claude chat in the normal desktop app.
Haven’t started using Claude or ChatGPT? Check out our guide on how to start using frontier AI models for free: Link
Haven’t created your first skill set? Check out our crash course: Link
Trending AI Tools this week
DeepSeek-V4.1-Flash (DeepSeek): smallest model in DeepSeek's new architecture family, adds native multimodal visual understanding
AlphaGenome Atlas (Google DeepMind): precomputed predictions for all 9 billion possible single-letter DNA changes, 1-petabyte free research database
That’s it for today!
See you next week with more exciting updates,
Eduardo - The human behind AEC field notes
Cañadas Del Teide, Tenerife, Spain