How Does AI Improve Building Operations?
October 27, 2026
FOUNDATIONS FOR AI-POWERED FACILITIES
“Sorry, it’s my first day.”
We’ve all heard countless new hires say it. In facilities maintenance, this is no reflection on a person’s intelligence or capability. It only speaks to their lack of experience and context: how the job is done, the general strategies of the business, and the proven methods honed over years of experience.
New team members come to us with education, experience, and intelligence. What they don’t come with is the context critical to our businesses. So we begin the work of building that context around them so they can do real work.
In the new world of AI, there is an important limitation to understand: every conversation is a first day. Large language models (LLMs) are stateless. That is just a fancy way of saying they work a bit like a person with short-term memory loss. They only know what they were trained on before a certain date, and everything else must be looked up on the fly. That leaves a context gap, where key business questions may not get answered directly, completely, or in the way that matters most to your business.
In the new world of AI, there is an important limitation to understand: every conversation is a first day.
So if you can’t “train” a new AI hire the way you train a new team member, what do you do? You need technology around the AI that provides all the context and information it needs, on demand, to fill the gaps.
These solutions come in many forms, but the goal is always the same: convert the decades of experience, knowledge, and strategies you value into skills and tools the AI can easily reference.
How does AI improve building operations?
AI improves building operations by turning building automation system data into trusted diagnostics and prioritized action. The foundation is a standard information model that explains what each point, asset, and system relationship means. On top of that model, expert systems diagnose faults, machine learning and large language models speed up setup and data cleanup, and AI agents help summarize findings, route work, and confirm whether fixes held.
Why can’t AI make sense of raw BAS data?
Imagine you point your LLM of choice at the full output of your building automation system (BAS) and ask it a question: “What is the biggest problem I must address today?” The LLM takes in all of the raw data and starts trying to make sense of it. What does this data mean? What problems can I identify? How do I evaluate “biggest problem”? Every one of these questions requires assumptions, and some may be far from what the user actually meant.
It finds a folder of points labeled “AC1” with a point labeled “SAT” that has stayed between 72 and 74 degrees. It decides that an air conditioning unit should be sending out cooler air and flags this as an area for action. “AC1 is not cooling,” it tells you. “Check to see if there is a problem with its operation.”
But there’s a problem with this diagnosis. AC1 isn’t an air conditioning unit. It’s an anechoic chamber, a test room built to absorb echoes, and the first of five. And SAT in this case isn’t a discharge air temperature. It’s the space air temperature. The BAS engineer who designed the controls preferred “SAT” for space air temperature and used “DAT” for discharge air temperature instead.
These kinds of pitfalls are everywhere when an LLM tries to interpret BAS data. There are so many caveats, nuances, and exceptions that it becomes nearly impossible to make reliable judgments from raw data alone.
What happens when AI has a structured foundation?
Now imagine the scenario again. This time, when the user asks, “What is the biggest problem I must address today?” the agent isn’t given raw data. It gets a structured dataset of outputs generated by expert systems AI. It has predefined priorities, lists of problems and opportunities already identified, and the avoidable costs tied to those problems. For every piece of equipment in the building, it knows its name, purpose, sequence of operations, and all associated points. For every point, it knows exactly what that point means and the units applied. It can even see calculated values based on those points that are not in the BAS at all. The agent is now working from knowledge, not just data.
The agent’s response this time:
“The highest-value action you can take today is addressing Air Handler 4, which serves Anechoic Chamber 1. The temperature is holding steady, but the unit is running during unoccupied hours when no temperature control is needed. It is also heating and cooling at the same time. The preheat coil valve is stuck open, raising the temperature across the coil, and the cooling coil has to work to bring it back down. Together, these issues are costing you $10,000 a year in wasted energy. Resolving them today will stop that loss. Here are links to all of the supporting data, laid out in charts so you can review and validate them before taking next steps.”
The difference is enormous. AI with this kind of structure can summarize findings and drive outputs well beyond what even the best technicians can derive without hours of review and deduction. Better yet, these solutions can run 24/7/365 to identify and escalate the right actions.
24/7/365 to identify and escalate the right actions
Why does AI need an information model?
It’s clear now that AI is limited by the information it can access, just like a new hire on their first day. It’s not about how smart it is, but the context it can draw from to make decisions. It takes real work to build a system the AI can work within, so it fully understands the context of the questions it is asked. That starts with a universal semantic model, or information model.
Semantic models are just ways of describing what things really mean. Even our somewhat ambiguous SAT example is a kind of semantic model. But BAS follow many conventions, and meaning can get lost in those ambiguities. Structures like Project Haystack and Brick Schema have come about to fill these gaps. Whatever the method, the important thing is that we’ve turned data into information, with built-in meaning about what that data represents in the real world.
Data → Information → Knowledge → Wisdom
nearly impossible to make reliable judgments from raw data alone
built-in meaning about what that data represents in the real world
insights you can act on across your building or your portfolio
driving meaningful action faster than ever before
Expert systems or large language models: which is better for FDD?
Unlike LLMs, expert systems use deterministic methods. Determinism just means that given the same inputs, you can always expect the same outputs. So you can rely on expert systems to correctly and consistently diagnose the problems and opportunities in the now cleanly modeled data in your building. At this stage, information has become knowledge: insights you can act on across your building or your portfolio.
Want to see that handoff live? Join the live Sprocket Workshop, October 21–23, where [PRESENTERS] walk through how expert systems and LLMs work together on real building data, from the first fault to the finished work order. Save your seat → [URL TBD]
How does plain-English search work on building data?
That depends on everything upstream. An assistant pointed at a raw BAS has little context. Pointed at a standardized model and a set of diagnosed, costed, prioritized findings, it becomes useful to people who need answers but do not live in trend charts. Results alone are not enough, though. An assistant also needs context about what the data means, how the model is structured, and how each diagnostic was calculated. Clockworks supplies that context through purpose-built AI skills and tools that give assistants structured access to the information model, diagnostic results, and portfolio KPIs, so clients get grounded answers without building that plumbing themselves. The person who needs the answer no longer has to know how to dig it out of the BAS. In Sprocket, the Clockworks built-in AI assistant, answers can include interactive charts and links straight to the diagnostic, task, or data behind them, so a question can turn into action in a few clicks.
The final step, and the one LLMs now make possible, is pulling all of this together and summarizing it across every building you manage. Gone are the days of needing to learn a specific software’s interface or build a dashboard. Now you just ask the questions you want answered in plain language, or describe the outcomes you want to reach.
“Show me the top 5 highest room temperatures.” “Are there any silent issues across my buildings where no one is complaining, but it’s costing me money?” “What is the best action I can take today to meet my critical compliance requirements?” All of these questions can now be asked and answered with real supporting data that proves the answers are correct. This is where we cross the final horizon: turning knowledge into real wisdom and driving meaningful action faster than ever before.
What changes for facilities teams with AI?
The result of all this is a new way to provide facilities maintenance, one that requires far less investigation and leads to far more proactive outcomes. Engineers can start their day with a command like, “Give me a prioritized list of problems to solve today, in descending order of comfort impact.” Rather than waiting for a phone call about a cold room or an angry email from management about an out-of-control energy bill, teams can handle issues as they arise and act quickly. And it all starts with the right kind of intelligence.
To get a program like this off the ground, start with the practical guide to building a sustainable FDD program, along with sector-specific write-ups for healthcare and higher education.
SEE WHAT YOUR BUILDINGS ARE ALREADY TELLING YOU
Walk through the Information Model, diagnostics, and AI workflows on real building data with a Clockworks expert.
Want to look around first? Start with the recorded demo and Q&A, the live demo overview, or the building analytics comparison guide.
FREQUENTLY ASKED QUESTIONS
What is fault detection and diagnostics?
Fault detection and diagnostics, or FDD, is software that continuously analyzes building automation system data to find equipment problems, explain why they are happening, estimate their cost, and recommend what to do next.
How is an FDD diagnostic different from a BAS alarm?
An alarm fires when one reading crosses a limit, with no sense of cause, cost, or priority. A diagnostic compares many readings across related equipment over time, works out the underlying cause, estimates the impact, and ranks it against everything else in the portfolio.
Can an AI chatbot read my BAS data directly?
It can read the data, but it can’t reliably tell what the data means. Point names follow whatever convention the controls engineer chose, so “SAT” might mean discharge air temperature in one building and space air temperature in another. Without a semantic model and expert systems behind it, a chatbot has to guess, and some of those guesses will be wrong.
What is an information model in building analytics?
An information model is a standard structure that defines equipment, data points, the relationships between systems, and the engineering values behind them, so the same analysis can run on any building. The Clockworks Information Model is built on equipment types, equipment variables, building variables, point types, inferred relationships, and calculated data.
What is an expert system in AI?
An expert system is a type of AI that encodes human expertise as explicit, traceable logic. Unlike a large language model, which predicts likely answers from patterns in text, it applies established engineering reasoning consistently everywhere and shows the steps it followed.
What is an agentic workflow?
An agentic workflow is AI that completes a multi-step process instead of only returning an answer. In a building, that might mean reading a diagnosis, creating a prioritized task, assigning it, and later confirming the problem is gone.
Can I ask questions about my building data in plain English?
Yes. Natural language data exploration lets users type questions the way they would ask a colleague, rather than building a report or a query. Answers come from real diagnostic results, so they are grounded in what the expert system already found.
Does Clockworks use generative AI or large language models?
Yes, in a defined role. The diagnostics themselves come from the expert system, which is consistent and traceable. Generative AI and agents sit on top of those results to summarize, answer questions, report, and drive action. Keeping the two layers separate is what keeps the output reliable. In the platform, that layer is Sprocket, Clockworks’ AI chat.
Will this work with our building automation system?
Yes. Clockworks is vendor neutral. Because equipment and points are translated into a standard information model, the same diagnostics can run across Johnson Controls, Siemens, Automated Logic, Niagara, Trane, Honeywell, and mixed-system portfolios.
Do we have to write custom rules for every building?
No. Because equipment and points are described in a standard information model, one global expert system applies across building types and portfolios without custom rule writing. Each site does supply its own engineering details, such as equipment ratings, utility rates, and schedules.
How do we know an AI recommendation can be trusted?
By tracing it back to the source. Every Clockworks diagnostic shows the logic, underlying data, and cost calculation behind the recommendation, with links to the source data in the platform, so engineers can verify it before acting.
How much energy can this actually save?
It depends on the portfolio and how consistently the findings are fixed. Published results range from a 46 percent cut in electric reheat at one surgery center to roughly $2.5 million a year across 47 buildings at the University of Iowa. Results depend less on the software than on whether someone owns the follow-through.
What kinds of organizations use this?
Health systems, universities and research institutions, life science and pharmaceutical facilities, and corporate real estate portfolios. These are places where equipment reliability, environmental conditions, and energy cost all carry real consequences.