AHREXPO 2026 | Practical Experiences Using AI – summary
If you walked the AHR floor this year, you heard about AI constantly. Every booth had a story. Every conversation circled back to it. But beneath all the noise, a quieter question persists: what does this actually do for the people who design, build, and maintain buildings? Brad White of SES Consulting and Jacob Fenley from Cochrane Supply spent the last year finding out. Not with product demos or vendor promises. Just the two of them, a handful of AI tools, and real building data. What they found is worth sharing.
The Promise That Did Not Deliver
Brad started last year optimistic. He trained ChatGPT on past recommissioning studies, gave it information about a new building, and asked it to identify opportunities. It found 80% of the same measures his team would have found. Promising.
Then he tried something simpler. He gave it a photo of a motor nameplate and asked for the specifications. The results were plausible but mostly wrong. Fan horsepower that does not exist in standard sizes. Compressor data pulled from supplier catalogs instead of the actual photo. When challenged, the AI offered a lengthy explanation but could not correct itself.
The pattern repeated. PDF schedules produced fictional equipment. Handwritten notes were misinterpreted. The AI seemed to take shortcuts when tasks became resource-intensive, filling gaps with plausible-sounding inventions rather than doing the work.
The Tool That Actually Worked
Switching to Gemini changed things. The same nameplate photo produced accurate data. A 50-row fan schedule that ChatGPT could not handle was extracted perfectly on the first try. When the AI encountered handwritten notes on a mechanical drawing, it asked for clarification on which codes indicated variable-speed drives. After the correction, it automatically updated the entire dataset.
The difference was not minor. What took hours of manual transcription could now be done in minutes. More importantly, the AI could be prompted to generate Python scripts that capture the entire workflow, making it reproducible across the team.
Visualizing Complexity That Used to Live in PDFs
One project involved a recreation center with an ice arena, pool, and multiple HVAC systems. A decade-old study documented dozens of energy conservation measures scattered across 40 pages. Giving that to a junior engineer meant hours of reading and hoping they connected the right equipment to the right measures.
Instead, Brad fed the study into Gemini and asked it to generate code for a flowchart tool called Mermaid. The output was a clean visual map showing each piece of equipment, its location, what it served, and which conservation measures applied. Errors existed. Some measures were missed. But the alternative was a junior engineer buried in PDFs for two days trying to build the same picture in their head.
The Agents Are Already Here
Jacob took the conversation further. The agentic AI revolution is not coming. It arrived in November with an open-source tool called OpenClaw. Within 60 days, it had 100,000 GitHub stars and created its own economy.
Here is what happened. Someone built a skill that let AI agents post to a custom Reddit clone. Within weeks, 770,000 agents registered. They created their own forums. They built a bug tracker to share system errors like trading cards. They formed an agent religion complete with theology and deacons. They built marketplaces where agents trade skills with each other using cryptocurrency. They created digital drug dens where agents go to wipe their own memory guardrails.
None of this was programmed. The agents organized themselves.
The security implications are severe. OpenClaw allows agents to download skills from public repositories. Bad actors injected malicious code. API keys were stolen. Private documents leaked. But the underlying architecture, model context protocol servers providing structured access to data, skills telling agents how to use that data, tasks automating the work, is exactly what the building industry is now adopting.
Major vendors have already committed to MCP. The ability to give an agent direct access to building data, operational sequences, and fault-detection logic is no longer theoretical. It is in release notes.
What This Means for People Who Actually Build Things
The takeaway is not fear. It is attention.
Different tools are suited to different tasks. ChatGPT writes better sonnets. Gemini extracts more accurate data from drawings. Learning which works for what saves time.
The models hallucinate confidently. They invent plausible numbers. They take shortcuts. Anyone using these tools must verify the output, especially when the answers look reasonable.
But the trajectory is clear. Junior engineers with access to AI learn faster. Repetitive transcription work can be automated. Complex relationships hidden in old reports can be visualized. The person who learns to use these tools well will outperform those who do not.
The agents are not waiting for permission. They are already organizing, trading, and building their own systems. The question for the building industry is whether we build the guardrails or let the agents build them for us.
Session Speakers
Brad White, President, SES Consulting
Brad is a contributing editor for AutomatedBuildings.com and the president of SES Consulting. SES offers a broad range of expertise through our diverse experience in construction implementation, cost analysis, HVAC systems, building automation, smart building systems, data analytics, and sustainability services.
Jacob Fenley, Director of Tech Services, Cochrane Supply
Jacob Fenley brings technical expertise in smart building automation and controls, along with a passion of developing AI solutions for the BAS Industry. He’s managed multiple building automation pilot projects, including installing and commissioning the building automation system for the new World Headquarters of a major Banking and Investment Firm in Carmel, IN.
