Everyone's Talking About AI Agents in Construction. But What Do They Do?

Morning. Every AI tool in construction now has an “agent” or a “digital coworker” bolted on but no one actually knows what it means. Today we cut through it, explaining how they work and where to use them.

IN TODAY'S ISSUE

Today's read: 5 minutes

🚀 Meta just released Muse Wonder their first frontier class model

Why it matters: Frontier AI models are becoming more capable but they still haven’t walked on a construction site. They don’t understand your project or your experience. Paired with you, they’re powerful.

What helps is knowing what’s an ‘in distribution’ (ID) task and what’s ‘out of distribution’ (OOD).

An ID task matches the data and scenarios AI is trained on like writing text for a report. OOD is unfamiliar or novel, like understanding an inspection photo. For the latter your experience and judgement is needed to get a better result.

🏛️ FTC subpoenas OpenAI and Anthropic over AI safety incidents

Why it matters: AI models are becoming increasingly capable and the government is asking for details on security incidents and safeguarding practices.

For us in the industry, it’s something to be aware of. When you interact with AI, security and data privacy needs to always be top of mind especially when dealing with client data. As simple as copy pasting a client’s email to draft a response could be a breach.

And as these tools become more integrated at work, understanding how to stay safe and secure is becoming more important than ever.

💼 Viral AI personal assistant startup Instinct raised $350m in funding

Why it matters: Last month Instinct revealed it was growing its user base by 10% a day. The difference between chatbots like Microsoft Copilot or Claude is that their AI can complete multi-step tasks on your behalf.

That means it can open a browser, fill out forms, book trips or appointments, manage calendars and more. While it’s consumer facing today, there’s a future where it’s on your job site, checking the schedule, emailing the subbies and ensuring the next day’s activities are on track.

If you’ve been hearing Procore, Autodesk or really any technology vendors messaging recently they’re all about digital coworkers and agents. It’s cool and sounds exciting, but what does it actually do?

To understand, let's take a step back on how we got here.

When ChatGPT came out in late 2022, many of us had that holy sh*t moment. It could write my report or email for me. But the problem was that it wasn’t actually that good or accurate. While newer model releases have resolved these issues, ChatGPT still has limitations as it doesn’t live where the work is.

If I need to complete a proposal or write a report, it isn’t connected to my project folder with my geotechnical reports, design calculations or assumptions. What I would need to (if my data security policy allows me) is tell or teach AI all the relevant background context. I’d then work with it back and forth, reviewing the output, providing suggestions and more information until I got a report that I could copy in…and tell my boss I did it all myself.

It’s time consuming for two reasons:

But more than that, I need to sit in the driver's seat, directing the LLM and reviewing the output. AI Agents or Digital Coworkers are a way to fix this. Instead of me directing AI, the system can take a predefined input, work through a series of steps and give me the result.

For example, take a submittal. When I receive one I check against the spec to catch any risks or scope gaps. It’s the same process every time. A ‘submittal agent’ would ingest it and automatically cross-reference and flag gaps for my review.Instead of me prompting the LLM it works autonomously. And the idea that all these big platforms and companies have is to create agents for all of the most common workflows we have like RFIs, Bidding, Daily Logs or Inspection Reports.

They're most useful for repeatable tasks where there's a clear goal. The input can vary, but the desired outcome is well defined.

Want to learn more about Agents? Keep reading here

How do you become the AI expert?

Most of us don’t know AI but everyone wants to learn. One trick we learned was from a reader, Sammy who sent up AI focused Lunch & Learns. Every month, a different person from the business shares a presentation on how they use AI in their team. It goes for 45 minutes with questions at the end and as the host, she gets to talk to Senior Leadership about AI and how things are changing.

Well that was a lot. If you learned something valuable from us, please hit send and forward it to a mate. AI noise is everywhere right now and our hope is that this newsletter helps you to learn a little more about what it actually means for you and how it can help you do your work.

As always, hit reply if you have a question or want to chat. We reply to each one.

Till next week,
Sid & Bhragan