Understanding Agentic AI: The Next Leap in Building Intelligence

At NexusCon 2025, AI stories and applications will be peppered throughout numerous case studies. Additionally, the agenda contains two sessions filled with AI case studies targeted at specific stakeholders:

• “AI for Energy Management: Real Case Studies” will explore how AI is being applied to HVAC optimization, energy forecasting, building controls, and occupant-driven operations.
• “FM Meets AI: Case Studies in Improving Operational Efficiency” will showcase how AI is enhancing operational efficiency, reducing costs, and streamlining decision-making. 

If you’re a building owner or facilities manager, you’re likely too busy keeping tenants happy and chillers chilled to follow every twist in the AI revolution, so we figured a crash course prior to these sessions would be a benefit to our audience (and ourselves). Specifically, for the less technical smart buildings professionals, we aim to demystify agentic AI: what it is, how it works, and what it might mean for the future of our buildings. 

The Evolution of AI in Building Operations

A useful framework for breaking down the AI-in-buildings story is to think in terms of “versions.” We’ve got Version 1, Version 2, Version 3 (and a forecasted Version 4) of modern AI’s maturity in our industry. Each step represents a leap in capability and a shift in how much the human is in the driver’s seat.

While this article focuses on the recent wave of AI driven by large language models (LLMs), it’s important to recognize that AI in buildings didn’t start with ChatGPT and other LLMs. Techniques like regression models, decision trees, and other classical machine learning approaches have long been used in energy forecasting, system simulation, and supervisory control, often without the need for neural networks like we see in LLM based AI applications today. In fact, many advanced building analytics tools and optimization platforms today use a blend of these classical models alongside newer AI methods. Some even integrate with LLMs to deliver better context or interfaces. The lines between logic, machine learning, and AI are not always entirely clear

‍

Version 1: Generative AI in the Form of Generic LLMs (All Talk, No Action)

Imagine a facility manager tinkering with Google’s Gemini on her lunch break. She prompts Gemini, “Help me troubleshoot overheated rooms on the 5th floor of my building”.

Gemini suggests checking the setpoint, understanding the mapping of the HVAC zones, inspecting airflow, evaluating reheat coils, investigating the heating load, and gathering maintenance history. (This is actually what Gemini recommended as I gave it this prompt.)

Useful, yet generic, advice. It’s not tuned to the facility manager's specific building. It doesn’t know the sequence of operation of the air handler, the vault on the VFD, or the heat load of the space. In Version 1 of AI for buildings, we’re using generative AI in the form of large language models (LLMs). 

Generative AI develops content (written language in the case of an LLM) based on patterns in the training data it was provided by its developer. It operates as both a fancy encyclopedia and a brainstorming buddy. 

The LLM isn’t connected to any of the actual building systems or data, so its help is limited to general guidance. It might spark an idea for the FM, but it’s not always directly applicable to their situation. Humans still need to determine what advice to trust and how to effectively implement it. 

At this stage, the AI isn’t performing any real work on the building; it’s purely consultative, and sometimes the suggestions fall short.

Version 2: Generative AI Enchanced with Retrieval-Augmented Generation (AI with Your Building’s Data) 

Now let’s take our scenario a step further. Our FM gets access to a new tool—let’s call it BuildingGPT. This AI assistant looks like ChatGPT, but with a crucial upgrade: it’s connected to our FM’s own building data and documentation. 

In industry lingo, this is Retrieval Augmented Generation (RAG). Essentially, the AI can “retrieve” information from the building systems or knowledge base and use that to give much more relevant answers. Now when our FM requests, “Help me troubleshoot overheated rooms on the 5th floor of my building”, the LLM is not only trained on the knowledge of the internet, it also knows current readings and context: it might use BIM data to map the 5th floor to AHU-5, pull in yesterday’s BAS trend logs, and identify a fault in the VFD within the air handler serving the fifth floor. 

The answer it gives can now be more specific: “Floor 5 is served by AHU-5, which currently has a VFD fault; consider reviewing the VFD fault codes.” 

Many leading technology vendors in the building operations space are effectively implementing Version 2 (generative AI enhanced with RAG). 

One such example is 75F’s Saffron AI (winner of the Most-Innovative Demo Nexie Award at NexusCon 2024). Saffron layers an LLM with data from your building’s HVAC system, plus the ability to write data back to your HVAC systems, while keeping the human as the decision maker.

In the example above, Saffron’s LLM application was able to directly adjust the setpoint of an RTU with the user's permission. Some RAG-enhanced AI solutions, like Saffron, can help the user take action, although the human remains the decision maker. This is done by building workflows between your LLM and systems it can send actions to (more on the nuances of this in Version 4 below).

While Version 2 (Generative AI enhanced with RAG) does not make decisions on behalf of the building operator, many technology vendors are intentionally providing solutions at this stage as we await improvements in AI technology. There is significant change management and trust required to hand decisions off to the AI to make without human interaction. If the AI hallucinates or fails to make the correct decisions for your use case, not only can it cause damage and expense, but it will also erode the trust of many stakeholders affected by the system.

Version 3: Agentic AI (The Assistant Becomes the Decision Maker): 

In Version 3, our FM can communicate her goals to the AI, and then the AI agent handles the task. This is the self-driving mode of AI in buildings. Instead of just chatting and providing ideas, the AI can plan and execute changes across systems to meet the set goals. 

One easy way to grasp agentic AI today is to see how it is being supplied to the consumer industry. OpenAI has developed an agentic AI called Operator. In this video, you can watch Operator search the web to find a recipe for a pasta dish, and automatically add all the correct ingredients to an Instacart order, based only on a prompt. Operator had to make decisions about which recipe to choose and what items to add to the cart.

This OpenAI technology is finding its way into our industry. Neeve.ai, a secure edge cloud infrastructure provider, is partnering with OpenAI’s Operator to bring this technology to the forefront of building operations. Neeve is uniquely well-positioned to succeed with agentic AI due to its positioning in the technology stack: it resides at the network layer, providing secure access to nearly all OT systems, enabling both the pull of information (RAG) and the creation of workflows to and from these systems. Neeve will be taking the stage at NexusCon to share how they're putting this technology into practice.

The most mature examples of agentic AI in buildings today support the use case of energy managers. Advanced supervisory control providers like Brainbox AI and Facil.ai are utilizing agentic AI to optimize chiller and RTU setpoints based on information from multiple sources, including electrical meter data, weather data, comfort data, and more. While the use case may be narrow today, these AI agents use data to decide setpoints that optimize systems beyond what humans are capable of. Both Brainbox and Facil will showcase these success stories with their customers during the AI for Energy Management session at NexusCon.

Beyond the energy management advanced supervisory control use case, we’re starting to see use cases emerge for the facility management role as well. One such example is from Xempla, a technology provider based in Australia. Xempla’s Autonomous Maintenance Operating Center is a family of AI agents that integrates with your existing systems (BMS, CMMS, IWMS, IoT sensors, etc.). 

Xempla’s agents are designed to continuously contextualize asset data, identify patterns or anomalies, and then think, plan, and act based on these anomalies, where normally a human would be required to do so. The key “decisions” that Xempla’s AI agents are making are the prioritization of issues, whether to escalate issues to humans, generation of maintenance work orders with relevant insights, and validation after work is complete that the issue is fully resolved. They’ve deployed this technology in over 75 buildings across various sectors, including commercial real estate, healthcare, and retail, and are rolling it out across a portfolio of more than 150 buildings for a large real estate developer in the Middle East. 

A key aspect of an agent is a framework that must be provided in order for it to make decisions. The typical agentic AI operates on the ReAct framework: reason, then act. Simply put, the agent will use an LLM, trained on generic data and RAG information unique to your building, to reason. It will then act on that reason by triggering decisions in other systems through the workflows that enable it to do so. Similarly, Xempla’s proprietary agent framework is called DIIV: the agent is trained to detect issues, investigate their causes, implement solutions, and verify results. 

For facility management, the challenging aspect of agent implementation (or decision-making) is that most issues arise from the physical environment and therefore require physical solutions. Thus, decision-making and implementation for Xempla and similar technologies often involve automatically pushing work orders to the CMMS of the maintenance team, which then requires a human to respond to the work order and resolve the issue. This may be as close as we get to a fully agentic solution until our buildings are filled with maintenance R2D2s 🤖. 

A Potential Version 4: The Model Context Protocol (MCP): 

Log in to read how MCP might be the answer to the rapid adoption of AI agents in buildings.

If you’ve ever tried to integrate one building application with another, you know the pain: custom interfaces, drivers, APIs, and never-ending firmware mismatches. For an AI agent (Version 3 above) to function, it must both consume data from multiple sources and send commands to various machines or software. Traditionally, that would require bespoke integration for each system. That’s a scaling nightmare, and it’s one big reason truly agentic AI hasn’t permeated buildings yet—the plumbing is hard. 

Many are calling out Model Context Protocol (MCP), an emerging open standard that could be a game-changer for connecting AI to our building tech. MCP was initially championed by Anthropic and has since been embraced by OpenAI as well. MCP creates a layer between the AI agent and the many services/tools in a building, so you don’t have to custom-build a new integration for every single tool. 

Our industry has seen the adoption of BACnet as the open serial communication protocol, and REST API as the application-to-application web communication framework. In a similar way, the support from large AI providers like Anthropic and OpenAI suggests that MCP may gain a network effect, becoming the framework of choice for standardizing communication between AI and services and tools.

This two-way, open-source communication protocol consists of an MCP server and an MCP client. A developer can set up an MCP server for a given tool or data stream (think meter data, CMMS applications, FDD applications, etc). On the other hand, the AI agent (the MCP client) can query these servers and retrieve information or issue commands in a uniform manner. The protocol handles the translation. This trend is prompting vendors to standardize their support for MCP, enabling their applications and data to be more easily accessed by AI agents. 

Why does this matter to a building operator or facility manager? Because it could dramatically lower the barrier to deploying AI agents in our buildings. If we had to involve a software developer every time we wanted to add an AI skill (such as a fault detection agent or an energy optimization agent) to establish tool-specific connections, it’d be slow and expensive. With something like MCP, adding a new AI agent to your tech stack might one day be as simple as installing an app. 

Conclusion: A Practical Path Forward

If you take anything away from this, we hope that you, as a building owner, remain both excited and skeptical about AI. When you hear the term “AI” in your next lunch and learn demo, recognize the vast spectrum of what that might mean. While agentic use cases could change the way we operate our buildings, the industry is still littered with claims of AI that aren’t much more than an LLM overlay. Conversely, we’ll likely see many players jump to agentic AI solutions while its decision-making capabilities are still questionable. 

Whether you’re a skeptic or an optimist, we hope you’ll bring your ideas and opinions to NexusCon as we try to sort out what’s real with AI in buildings and where we’re headed.