This post is based on the "Agentic AI in AEC: Where Firms are Slashing Project Time and Seeing Results" webinar hosted on October 27 by Cameron Duncan of Hallian Technologies.
Cameron Duncan, Co-Founder of Hallian Technologies, answered six practical questions about AI agents during the Q&A portion of a recent webinar.
Attendees wanted to know how to protect contract privacy, what a "skill" actually is, which documents to hand an agent, and how to train a team. His answers are condensed below.
Q: How do we maintain the privacy provisions in our contracts while giving an AI agent project history and context?
Cameron said there are a few ways to answer this, starting with time spent reviewing the terms of every tool you consider. Terms for tools like Claude, OpenAI, and Groq are changing seemingly weekly, he noted.
He called Microsoft Copilot a good option for many firms. He said it's inside the Microsoft tenant and carries the customer copyright commitments you'd expect from a company built on serving enterprise customers.
A third route is private, self-hosted solutions, which is what Cameron's firm builds. He was upfront about the plug. The appeal, in his words, is full control and ownership of everything going into the system.
Which path fits depends on your scenario and on what your contracts and projects require. And for government work, CMMC concerns can change the answers.
Q: What is a "skill" for an agent? Is it just an effective set of instructions from the user?
At its most basic, yes. Cameron described a skill as a set of instructions, and he often likens it to a standard operating procedure handed to an agent so that a task is done the same way every time.
His example was a company memo template. A skill can explain how to build memos in your firm's format. Skills can also embed reference material, such as the template itself, and even scripts, meaning code snippets the agent can run on your behalf.
So how does it differ from a prompt? Slightly, Cameron said, though the two are similar in some ways. At the highest level, a skill is a work instruction for the agent.
Q: What documents do you share with the LLM when developing a PM assist agent?
Cameron's demonstration used a standard project folder from a geotechnical engineering firm. Inside were the contract, the technical scope, drawings showing where the work would happen, transcripts, memos, meeting minutes, and boring logs.
His recommendation was straightforward: whatever you normally use to manage a project is what the agent should be able to access.
Q: Do you build a new agent for each project, or adapt one agent to the parameters of each new project?
It depends on how different the projects are. Cameron's team recommends a standard set of instructions for any project agent, so you can essentially copy and paste it, then adjust certain details in the description.
In his experience, once a robust project agent exists, setting one up for another project takes about 15 minutes. You change the project folder, the project name, the client name, and, if relevant, which project the agent looks at in Unanet.
The approach breaks down when projects are wildly different. If a firm does geotechnical work and also environmental Phase I assessments, Cameron would build a separate agent for the latter, because it needs to do slightly different work.
Q: How are AE firms training their workforce on these fast-changing tools, and is there data on how effective the training is?
Cameron answered from what he has seen rather than from data, and he didn't cite effectiveness statistics. Success, he said, starts at the top: an organizational commitment that the technology is here to stay, backed by a push from ownership or leadership to set a firm direction.
The next ingredient is a robust AI committee. It decides which projects to pursue, what to build, and where the opportunities lie, then carries those plans out. As part of its mandate, the committee should run an actual training regimen, such as lunch-and-learns. One client treated its launch like an internal marketing campaign and produced a full marketing video.
His summary: training is never one-and-done. It's a slow drip over time, tailored to what your firm needs.
Q: Most project folders are messy, with irrelevant data and draft versions of files. Do your recommended solutions rely on disciplined use of document metadata?
Cameron agreed that many project folders are messy and said there are several ways to counteract it. The first is to design folder structures knowing that more than one person will use them. If someone who isn't on your project can't open the folder and understand what's going on, you probably need firmer standards. He acknowledged that's hard to do, but as these tools get more powerful, better data produces more robust results.
The second is something simple that has worked for his team: an archive folder inside the project folder, with the agent instructed not to look there. Anything that becomes a draft moves to the archive. The history stays, while the agent focuses on what is actually live on the project.
The last piece is making sure your teams follow the guidance you set.
The takeaway
Across all six answers, one message keeps surfacing: agents perform best at firms that have done the unglamorous groundwork first. That means clear contract terms, clean folders, standard instructions, and team habits that stick.
Put these ideas to work at your firm
Cameron Duncan leads PSMJ's two-day, in-person AEC AI Acceleration Workshop, built for the people who own the AI decision at AEC firms. Expect live demonstrations rather than slide-reading: a live build of a working AI agent, an AI-powered RFP analysis and go/no-go workflow, and a tiered roadmap for your own firm. You also leave with a Monday-morning action plan.
The workshop carries up to 13 LU hours and comes with a satisfaction guarantee. Register for the AEC AI Acceleration Workshop or view the full agenda.