Monday Live, April 2026 | AI across the smarter stack: the upper layers
This week’s Monday Live series has been working through the smarter stack from the bottom up, examining how AI fits into each layer of the building intelligence stack. The first two sessions established the foundation: knowledge graphs, semantic structure, and why connected data is not the same as understood data. This third session turned the view around and looked from the top down, asking a question that turns out to be harder than it sounds: what does AI actually do at the purpose and operations layers of a building?
Purpose is not changing. The ability to answer for it is.
The session opened with a framing observation that kept resurfacing in different forms throughout the hour. At the executive level, the questions being asked of buildings have not fundamentally changed. A CEO or CFO wants to know how the portfolio is performing, how space is being used, whether facilities are serving the mission. Those questions existed before AI, and they remain the same.
What is changing is the lag between the question and the answer. Today, that question gets delegated. Someone is assigned to gather data, compile a report, and return weeks later with information that is already partially outdated. The reason is not a lack of will. It is a structural problem: the data exists, scattered across disconnected systems, but no one can surface it on demand.
AI exposes that problem in a way that previous tools did not. When someone tries to prompt a system for a live answer, and the system has no accessible data, the prompt itself makes the failure visible. As one participant put it, it is not a new problem. Buildings have been opaque to their owners for decades. AI just makes the opacity impossible to ignore.
A related but distinct point emerged around the kinds of questions now being asked. AI is surfacing questions that were never asked before, not because curiosity was absent, but because the tools to pursue them were not there. The ceiling on what gets asked has historically been set by what seems answerable, and that ceiling is now higher.
The cost of tokens is forcing a useful distinction
One of the more practically grounded threads in the session concerned the economics of AI use. Early access to large language models felt free and frictionless. That has changed, and organizations are beginning to encounter the real cost of routing everything through an LLM, including queries that do not require inference.
The distinction matters in buildings specifically because a well-structured knowledge graph handles a large class of questions as simple queries. If the relationships between assets are already explicit, asking how one system connects to another is a lookup, not an inference. Routing that question through an LLM is wasteful and introduces unnecessary uncertainty. The building already knows the answer precisely. Asking the AI to guess it from context costs tokens and introduces risk.
The more useful architecture keeps the two things separate. Explicit, structured data answers exact questions directly. AI handles the layer above that: reasoning about incomplete data, identifying patterns, surfacing recommendations, and working with the many real-world situations where the connection is not yet documented, and inference is genuinely required. Getting that distinction right is partly a governance problem, not just a technical one. Organizations need to decide which questions warrant inference and which should simply be answered from the graph.
The fire scenario makes the case without abstraction
The most pointed discussion of the session came through the use case of fire response. It is a purpose that every building has, one that is not optional, and one where the consequences of information failure are immediate and measurable.
The observation from participants who have worked directly with fire departments was striking in its specificity. When a fire team arrives at a building, they reverse engineer it on the way in. They use experience, physical observation, and pattern recognition to understand a structure they often have no prior knowledge of. Yet the building they are entering may have thousands of pages of design documentation, decades of maintenance records, and an active sensor network. None of that is accessible to them in the moment.
This is not a gap that AI alone solves. It is a data preservation and access problem. The building’s knowledge has not been maintained in a connected, retrievable form. But AI can play a specific role once the data is in order: filtering the total mass of information down to exactly what the fire team needs at that moment, in that building, for that situation, and nothing else. The parallel to knowledge graphs is direct. A query against structured data about structural load limits, hazardous materials, and ventilation zones could surface life-critical information in seconds. Without the underlying structure, the AI hallucinates or returns generalities that are not useful under pressure.
The point was made that this use case is not exotic. Every user of a building, whether an occupant, a facilities manager, a maintenance technician, or an emergency responder, has a specific slice of information they need at a specific moment. The value of AI in buildings is not that it knows everything. It is that it can find the right slice for the right person at the right time and return nothing else.
Operations: from reactive to proactive
The second half of the session examined AI at the operations layer more directly. The consensus was that current deployment is largely reactive: a technician or facilities manager prompts a system, gets information, and acts on it. That is genuinely useful and already represents an improvement over the previous state, where getting an answer required assembling a report.
The more interesting question is when AI becomes proactive. Not predictive in a grand sense, but simply running as an agent across building data and flagging things that warrant attention before anyone thinks to ask. A concrete example from outside buildings illustrated how near that threshold already is. A team working on a set of documents updated one of them, and, without being prompted, the system identified nine specific changes to a related document. The reaction in the room was not excitement at the capability. It was a moment of recognition that something had shifted.
The analogy holds for buildings. A system that watches VAV performance data over a week, identifies trends suggesting an emerging inefficiency, and surfaces that observation in a morning summary is not science fiction. What it requires is the same as the fire scenario: data that is connected, structured, and accessible. The proactive capability is latent in the systems most buildings already have. It is waiting on the data infrastructure.
Robots, timelines, and the velocity problem
The session closed with a broader discussion of physical AI in buildings, autonomous cleaning systems, inspection robots, and the overall trajectory of AI-assisted facility management. The honest answer to the timeline question was that no one is confident, because the rate of change in the underlying capabilities has outpaced every prior estimate.
A pointed observation: the smarter stack framework that Monday Live uses as its analytical structure was created three or four years ago. When it was built, AI was not part of the conversation. That is how fast the landscape has changed. The tools that seem speculative today may be unremarkable within the same window.
The practical implication is that the decisions being made now about data infrastructure, openness, and semantic structure are not just relevant to current AI capabilities. They are setting the conditions for capabilities that do not yet exist. A building that cannot explain itself today will not be able to use the tools that emerge in the next three years either. The data problem does not solve itself when the tools improve. It has to be solved before the tools can be useful.
The session ended with a reflection that has run through all four Monday Live discussions this month: getting AI organized to do its job effectively may matter less to the AI than to the humans overseeing it. When the data is structured clearly enough for a machine to query with precision, it is also clear enough for people to manage effectively. The discipline that good data infrastructure requires turns out to be the same discipline that good operations requires. AI did not create that need. It is just making it harder to defer.
Monday Live meets every Monday. Join the discussion at mondaylive.org. The full session recording is posted below.
