BMS reConnect #52: The "AI in BMS" Issue
AI use cases are (finally) coming into focus. Also: lost gold and peaches.
DISCLAIMER
All opinions are my own. Hit the subscribe button if you haven’t yet.
SPANISH GOLD
I’m a click-bait-sucker when headlines pop up about the discovery of sunken treasure. Shipwrecks and Spanish gold…something romantic about the notion, I guess. And we got a good one last week:
You are reading that right: perhaps $20 BILLION worth of Spanish gold is on the sea floor near Cartegena. Well…technically, it was Colombian gold. Or Bolivian gold. Or Mexican gold. Regardless, it was somebody else’s gold, but the Spanish had taken it, minted much of it, thrown in a few tons of precious gems, and loaded it onto one of the famed Spanish galleons to take back home. They never made it.
As it were, the Spanish had a pretty poor understanding of Atlantic weather patterns and ocean floor topography, continuously losing ships to storms and reefs. And they got worked by the British in a bunch of naval engagements in the post-Spanish Armada era, with many a decisive defeat foisted upon them by El Draque, aka Sir Francis Drake. (The History of the Americans podcast I’ve referred to here a few times does numerous episodes on Drake, who by all accounts was a verified badass).
During the period of 1566-1790, it is believed Spain lost upwards of 500 treasure ships, including the loss of five fleets. At least one of those was carrying the equivalent of the nation’s entire annual GDP. Imagine a country’s entire annual economy sinking because you didn’t know about the Atlantic hurricane season? I feel like somebody should have gotten fired for that.
But none of these maritime disasters were as costly as the San Jose, sunk near Baru Island, Colombia in a 1708 naval engagement with the British. Experts believe there were ~11 million gold coins on the ship. Initially discovered by a private treasure hunting company—wild business to be in, BTW—in the late-90’s, the announcement this week by the Colombian government provides a level of confirmation that the wreckage is indeed the San Jose. And it sounds like the government of Colombian President Gustavo Petro has designs on retrieving the booty and splitting it with the treasure hunters. Crazy story, that stuff like this actually happens.
Here’s a list of the most notable Spanish treasure fleet shipwrecks in the Atlantic/Caribbean:
1. San José (1708)
Location: Off Cartagena, Colombia
Treasure Lost: ~$17–20 billion
Notes: Possibly the richest shipwreck in history; ongoing legal battle
2. Nuestra Señora de las Maravillas (1656)
Location: Bahamas
Treasure Lost: ~$1 billion (estimated cargo)
Notes: Crashed on reef; rediscovered in 2022; much treasure still missing
3. Nuestra Señora de Atocha (1622)
Location: Off Key West, Florida
Treasure Lost: ~$450 million
Notes: Gold, silver, emeralds; found by treasure hunter Mel Fisher
4. 1715 Treasure Fleet
Location: Florida east coast
Treasure Lost: ~$400 million
Notes: 11 ships wrecked in hurricane; treasure still washes ashore. This is why the Treasure Coast of Florida is called the Treasure Coast.
5. Santa Margarita (1622)
Location: Florida Keys
Treasure Lost: ~$50+ million
Notes: Companion to Atocha; partially recovered
6. 1733 Treasure Fleet
Location: Florida Keys
Treasure Lost: ~$20 million
Notes: 20 ships lost in hurricane; some treasure salvaged shortly after
AI in BUILDING AUTOMATION
PART 1: EFFICIENCY
I feel like I’m not alone in patiently awaiting the AI tech wave that has been promised to the CRE world for years now. I have been paying attention to this space for awhile…probably back to 2018’ish. Lots of talk about ambitious projects related to BMS platforms portending to organize, deploy, and optimize themselves autonomously. Of continuously improving algorithms and large language models (LLMs) that will make installing and operating “smart building” technologies easier, more effective, and more beneficial to the end user. Of tech that will be a building operator’s greatest tool in the age of the Great Retirement, as Boomers ride off into the sunset and throw their keys over to the less mechanical/more tech-savvy younger generations.
Lot’s of talk, but I hadn’t seen anything particularly compelling. Platforms I have previously seen were problematic to deploy and have seemed like more marketing than material. I’ve just been kind of waiting and wondering when I’d see something eye opening.
Good news: I got just that last week when I started to dig into BrainBox AI’s tech stack. My opinion has shifted pretty quickly; AI for buildings is ready, and it is going to change things sooner than later.
The BrainBox platform isn’t brand new; they have been around since 2019 and were acquired by Trane in early 2025 after previously raising a fair amount of venture capital. But recently seeing their evolution up-close is what has me looking at this with keen interest.
The key to this tech is their neural network. According to IBM:
“a neural network is a machine learning program, or model, that makes decisions in a manner similar to the human brain, by using processes that mimic the way biological neurons work together to identify phenomena, weigh options and arrive at conclusions. Every neural network consists of layers of nodes or artificial neurons, an input layer, one or more hidden layers, and an output layer. Each node connects to others, and has its own associated weight and threshold. If the output of any individual node is above the specified threshold value, that node is activated, sending data to the next layer of the network. Otherwise, no data is passed along to the next layer of the network. Neural networks rely on training data to learn and improve their accuracy over time. Once they are fine-tuned for accuracy, they are powerful tools in computer science and artificial intelligence, allowing us to classify and cluster data at a high velocity.”
Here is a graphical representation of a basic neural network:
And here is a representation of a Deep Learning model, which is essentially a neural network with increased complexity (more hidden layers, enhanced quantity/quality of connections between nodes, etc).
The BrainBox AI tech stack utilizes a deep learning model of connected points (variables) pulled from a building automation platform. It’s algorithms are trained through machine learning on the thermal dynamics of a given space. By running “in the background” for a few weeks, the model will become trained on how a wide range variables impact the overall desired outcome: typically, space temperature. Once trained, the neural network will then make predictions with over 99% accuracy as to the future state in the space. Adjustments and modulations are then made on the fly to “optimize” the systems to achieve the outcome with minimal waste.
Here’s a visual representation. The black squiggles are the model runs for a morning warmup sequence. The yellow line is the sequence optimized for future looking load conditions so the space achieves setpoint at occupancy with maximum efficiency.
BrainBox AI’s platform states their system can be used to optimize the following:
HVAC runtime/scheduling
Zone-based setpoints/multi-zone systems
RTU/AHU system coordination
Ventilation/economizer systems
CHW/HW loops
Condenser water reset strategies
Equipment staging
Central plant performance
Demand response sequences
etc
Now…consider the neural network approach described above—a living, continuously learning, predictive sequence constantly optimizing towards a desired outcome—and compare that against a standard, static BMS sequence. Layer in the ASHRAE 36 compliant sequences, and you get this:
Even buildings with a modern BMS can see 10%-15% savings, while buildings without modern BMS can achieve up to 40%.
Noting there are human barriers to deploying this type of tech—operational change is never easy—it is nonetheless fascinating to start envisioning the impact this will have on the future of the industry. I am really starting to see it now, and it is hard to conceive a scenario where neural network-driven technologies don’t come to dominate. Seems like they will be a pre-requisite for designing and deploying truly smart buildings and dynamic electrical grids. They also are useful tools in driving occupant comfort and equipment longevity outcomes.
BrainBox AI is certainly not the only player in the space, they just happen to have the most compelling use cases I have seen to-date. To be sure, all of the Big Four controls manufacturers are working hard to get into this space, either by developing their own software stack or through M&A. It will be very interesting to watch this evolve over the next few years, including how this does/does not wind up getting specified into new projects. My bet is that the most useful tools will be developed by smaller firms nimble enough to develop software at the market’s pace, and that many of those firms will get gobbled up by the bigger players. Costs will likely need to continue their downward trend to accelerate adoption.
I’ve watched most of BrainBox AI’s YouTube videos so you don’t have to. This is a good place to start if you are interested:
AI IN BUILDING AUTOMATION
PART 2: OPERATIONS
I had intended to pull together a single write-up about “AI in BMS”, but I have come to realize it will be better to break it up into parts. Part 1 above focuses on effective technologies I am starting to see impact the efficiency of how building systems operate by embedding AI within the operating sequences of the mechanical equipment.
This section is dedicated to another use case for AI tech: improving the efficiency of operational processes that ultimately service those building systems.
I feel safe at this point making the assumption that most people who read this are going to be familiar with ChatGPT. The familiar chat bot user interface allows users to query against OpenAI’s proprietary Large Language Model (LLM) in a conversational way. The user makes a prompt into a text bar (the agent), the chat bot responds, and a conversation typically ensues. These tools can seem both all-knowing and grossly inaccurate at the same time, illustrating the paradox of AI. Generally speaking, active skepticism is a healthy framework when working around these tools.
But thinking of them as tools is a good thing, and I want to highlight a use case where AI is being deployed as an efficiency tool for delivering improved service to the HVAC industry.
BrainBox AI’s optimization capabilities were highlighted in the prior section. I’m going to build on that example, assuming a building with a modern BMS infrastructure and a BrainBox AI overlay. In this scenario, every point in the BMS has been integrated into the platform, and the model’s neural network is continuously analyzing the system’s data.
Now, let’s assume there is a cold call at the building. A service ticket is issued and a technician assigned. Using the BrainBox ARIA platform, the technician can engage a chatbot-style agent and query for information about the system. Instantly analyzing all system data, the tool will conversationally respond with useful information about the nature of the mechanical problem, potential diagnostic steps to follow, and can even go so far as recommending required parts or tools.
Fault Detection & Diagnostics (FDD) are not new, but I’ve always sensed a scale issue. The FD part has always seemed pretty straightforward me, and can be pretty easily programmed into a control. Even better if the OEM provides the logic. But that second D—the diagnostic piece—has always been a struggle. It requires a considerably bigger lift in terms of custom programming. In my experience, that leads to high price tags that have driven true FDD into the realm of well-capitalized institutional users and not many others.
It seems like ARIA solves for this by virtue of its real time data extraction capabilities, conversational user interface, and HVAC-first training. I think that UI piece is a really big deal. Having a platform that can progress beyond the detected fault and get you to a prescriptive repair is a major leap forward. Being able to then engage with that tool for added diagnostic insight will be a game changer. Accomplishing that with a familiar chat interface on a mobile app will allow the technology to get where it is most needed: in the hands of the technicians.
I am operating right now under the premise that labor efficiencies gained through AI-driven insights will dwarf energy savings from optimized sequences. Time will tell.
This ARIA tool provides a chatbot interface into all of the data within the BMS, with many use cases beyond mechanical diagnostics. I will be sure to highlight more of these into the future.
AI in BUILDING AUTOMATION
PART 3: WHAT DOES THE AI SAY ABOUT ITSELF?
Thought it would be interesting to ask AI where it sees the industry going. Straight copy/paste here from ChatGPT 4o:
🧠 Where AI Is Taking Building Automation (2025–2035)
1. Autonomous, Self-Optimizing Buildings (Next 5–10 Years)
The Holy Grail of AI in building automation is self-optimizing, self-healing buildings—facilities that automatically:
Adjust HVAC, lighting, plug loads, and shading based on real-time and predictive data
Diagnose and correct faults without human intervention
Continuously learn from occupants' behavior and environmental data
Optimize for carbon, cost, comfort, and grid interaction simultaneously
Example Technologies:
Neural ODEs (like those from BrainBox AI) that simulate future thermal states
Digital twins combined with reinforcement learning (RL) to test and deploy optimal strategies
AI-enhanced fault detection and diagnostics (AFDD) that prescribe and execute fixes
ETA: Already happening in pilot buildings today. Widespread adoption in high-performance and LEED Platinum buildings by 2030.
2. Context-Aware, Human-Centric Systems
AI will move beyond static schedules and setpoints. We’ll see:
Systems that learn individual occupant preferences (e.g. preferred temp, lighting level)
AI interpreting unstructured feedback (e.g., "too stuffy" = tweak CO₂ ventilation)
Seamless integration with wearable tech and smartphones for hyper-personalized control
ETA: Expect high-end commercial and healthcare spaces to start adopting by 2027–2029.
3. Grid-Interactive Buildings with AI-Powered Demand Flexibility
AI will enable buildings to dynamically respond to grid signals (real-time pricing, carbon intensity) without compromising occupant comfort.
Think HVAC systems that precool a building because tomorrow will be a peak carbon or pricing event
Load shaping through AI-controlled thermal storage, EV charging, battery, and HVAC orchestration
ETA: Pockets of this exist in places like California and NYC already. Full-scale demand-flexible campuses with AI orchestration will be mainstream by 2032–2035.
4. AI Co-Pilots for Facilities Teams
Generative AI and natural language models will evolve into assistants that:
Translate control logic to plain English (and vice versa)
Autonomously generate control sequences
Help field techs troubleshoot via chat + augmented reality (AR) overlays
ETA: Early forms in 2025. Expect widespread availability for techs and engineers by 2026–2028.
5. Predictive CapEx Planning & Asset Management
AI will forecast:
Equipment failure and replacement years in advance
ROI of retrofits based on local policy, weather, and utility rates
Budgeting needs for compliance (e.g., LL97, BEPS) tied into policy engines
ETA: Analytics platforms are just starting to integrate this. Full AI-powered portfolio-level planning tools likely by 2028–2030.
6. Federated Learning Across Portfolios
Rather than siloed machine learning models per building, portfolios will benefit from shared learning:
A failed AHU behavior in Boston helps predict a failure in Baltimore
Privacy-preserving algorithms (federated learning) ensure data never leaves site
ETA: Concepts proven in adjacent sectors (healthcare, mobile apps). Could hit building ops by 2030 with the right frameworks.
What’s Holding AI Back?
Data quality & silos (bad trend logs, no standard tagging)
Legacy BMS platforms with poor APIs
Lack of trust from facilities and risk-averse operators
Unclear ROI or value prop for mid-tier buildings
Expect most of the early AI disruption to land in:
Class A commercial real estate
University and healthcare campuses
Governments subject to emissions laws (e.g., NYC, Boston, DC)
ACOUSTIC JAMS
Peaches. Working on a version right now we’re calling Peaches^3 where we mash up this song with Peaches by Justin Bieber and Peaches by Jack Black from the Mario Movie. It’s a banger.