Go Slow to Go Fast: Against AI in Building Operations

Interest in AI has exploded with the recent release of ChatGPT, a system which provides a conversational interface to a Large Language Model (LLM) called GPT-X (3.5, 4, etc.). Much has been written about the potential for AI more broadly to optimize processes, reduce costs, enhance sustainability, and improve user experience in commercial office buildings:

• Integrate multiple software tools behind a conversational interface
• Model predicted building performance using Neural Networks
• Control the building using Reinforcement-Learning Model Predictive Control (RL-MPC)
• Understand the facility's performance and identity areas for improvement
• Enhance tenant experience through personalized services and control
• Optimize the performance of HVAC and lighting to enhance efficiency and comfort
• Analyze sensor data and historical maintenance records to predict equipment failures

While these are all valid areas for improvement, I believe that involving AI in commercial building operations would in most cases be a mistake, and I will attempt to explain why in a way that will remain true as the field continues to evolve. I do think that there is a place for conversational interfaces and some statistical modeling, but that the majority of benefits will come from basic technical and organizational improvements.

What is an LLM like ChatGPT? These models are currently 'trained' on vast amounts of text data from the internet, and through this process 'learn' to understand the relationships between words and, according to some, concepts. You can think of them as a form of compression that does not necessarily retain the data they were trained on, but rather a process for generating what they would generally expect to see in response to your input. This gives them amazing capabilities that appear to approach rational thought, but they also exhibit strange 'hallucinations' and an inability to process basic information accurately. These models are complex and currently not 'explainable', meaning there is not and may not ever be a way to understand why a particular response was given.

Building operations relies on contextualizing accurate information and making decisions to address specific problems. Because LLMs do not store data or reproduce it faithfully, the field appears to be moving toward using them as a conversational interface, or middleware, for interacting with other databases and models. This clarifies that the value proposition of AI is in fact heavily dependent on your underlying data and models. Applying AI to building operations is in my opinion an attempt to deliver a silver bullet to underlying problems with data availability, data quality, user experience, and organizational challenges. In the words of Seth Godin:

In the case of a system replacing a previous system, [design thinking] questions often get replaced with: What’s the easiest way to polish what we compromised on last time?

Our industry has a very difficult time collecting, validating, and centralizing critical building records, let alone organizing them into a format which can be queried by an AI. Operational data is even more difficult to maintain, requiring physical verifications and sensor calibrations on systems that may not have ever been functioning correctly in the first place. If these issues are not addressed, involving AI threatens to make achieving high performance even more difficult.

Our opportunity is that not many buildings are operating at their maximum design potential, and in my opinion the most common deficiencies are not caused by a lack of advanced technology. Building records are typically fragmented, out of date, or inaccessible, and building systems are not well commissioned or regularly retro-commissioned. Operations staff may not have the training or information they need to operate the building in line with the design intent while responding to occupant requests. Before trying to solve operational problems with more technology, we should be trying to understand why operators face so many challenges to performance.

Rather than being the main instigators of an accident, operators tend to be the inheritors of system defects created by poor design, incorrect installation, faulty maintenance and bad management decisions. Their part is usually that of adding the final garnish to a lethal brew whose ingredients have already been long in the cooking. - The Field Guide to Understanding Human Error by Sidney Dekker

It seems to me that many in the industry believe that more data and simulation is the answer to the challenges we face. We transfer our data into a digital twin, only to be defeated by something as simple as a manual override, a change in space use, or an uncalibrated sensor. Buildings and their users are complex systems and the data we input is never enough to capture all the edge cases, creating a spiral of increasing effort to maintain a model that never truly captures reality. People with access to timely, quality information can continuously evolve their practices in a manner which will forever be difficult to quantify and automate. Replacing ourselves with technology is a dream that can never be achieved, but which consumes massive amounts of resources in the attempt.

Computational knowing requires surveillance, because it can only produce its truth from the data available to it directly. In turn, all knowing is reduced to that which is computationally knowable, so all knowing becomes a form of surveillance. Thus computational logic denies our ability to think the situation, and to act rationally in the absence of certainty. - New Dark Age: Technology and the End of the Future by James Bridle

For the foreseeable future, humans will (hopefully) be occupying our buildings and humans will be maintaining them, and this introduces either a massive challenge to simulate and control all of their activities or an opportunity to build something that helps us support each other as users of the building. I believe that the biggest opportunities for improving the performance of our buildings lie in mastering the fundamentals, and in becoming an organization where technology is implemented not to solve problems but to amplify our human ability to solve them. When we support the learning capacity of our organization, we build systems that can naturally manage complexity.

Counter to this, AI tools create a "black box" effect where the chain of human understanding is broken, creating a dependency on the technology that deprives us of our ability to learn and adapt. As an example, Model Predictive Control (a form of ML) was recently found to be 3% more efficient in most scenarios than ASHRAE's Guideline-36 on a VAV system. However, an MPC system cannot be understood by operators and in my experience will be overridden as soon as a few hot/cold calls are received. In contrast, Guideline-36 is fully documented, trainable, and exposes many control points for the operator to customize. In addition to the difficulty of development and implementation, is MPC really worth that extra savings?

Rather than rushing to implement AI, I would suggest taking a step back and evaluating the underlying foundation of your technology and organization.

• Have my goals been defined for the building and how can I measure progress?
• Have all my building systems been properly commissioned or retro-commissioned?
• Have the operations staff been fully trained on all systems and practices?
• Have my software systems been set up to get the most out of their capabilities?
• Are all of my building records digitized, validated, and in the right location?
• Do I have incentives in place to keep those records updated? How can I reduce friction?
• Have alarm and fault systems been validated and associated with workflows?
• How is institutional knowledge preserved and taught to new hires?
• How am I using feedback from operators and occupants on what isn't working?

If an organization has addressed all of these questions, I don't believe they would feel the need to implement an AI system. However, from this foundation I do believe it is possible to implement advanced technology in a manner which supports users of the building without becoming a black box. In my experience, once systems are functioning optimally, data is trusted, and user experience and workflows have been addressed it becomes possible to layer in expert systems, automations, and modeling. Because the underlying data has been validated and is maintained over time by people who understand the systems, advancements will have the greatest chance of offering value rather than becoming a new challenge to performance.

• Can I include a conversational interface for occupants to interact with the temperature/lighting controls or submit a ticket? Does it have to be AI or could it be simpler and achieve the same effect?
• Could I implement a model within a Fault Detection system to predict energy consumption and flag variations for review? Does it have to be machine-learning or could it be simpler?
• Within a Fault Detection system, could I implement a model to predict equipment failures before other systems detect them? Does it have to be machine-learning or could it be simpler?
• If these models are producing valid issues that operations trusts, could I automate work orders based on those issues? How can I make diagnoses more descriptive of potential root causes?
• If my models are consistently and reliably suggesting changes, could I automate them with a control override? Is there a simpler way to implement this in the underlying system?

More overlays, single panes of glass, and now conversational interfaces will not solve underlying problems with poorly functioning equipment, inaccurate data, missing records, and lack of workflow design or training. When we take these issues and feed them into automated systems, we get both poor results and a decreasing ability to understand what the variables for improvement are. I believe that if we focus instead on providing timely, trusted information to users of the building we can inform the very human process of learning and growing to achieve higher performance.

#artificialintelligence #buildingautomation #ai #realestatedevelopment #buildingoperations #machinelearning #hvac #smartbuildings #sustainability #teamhuman