Smart buildings have always had an operating cost. The batteries in sensors need replacing, networks need maintaining, software licences renew, and systems integration rarely remains finished for long. Artificial intelligence introduces another cost category, one that is (potentially) much harder to control - the cost of computation.
For most owners, the economics of building technology have been reasonably legible. A platform came with a licence fee, a service contract and perhaps a charge per square metre, building or user. Large language models and agentic systems introduce a more variable model. Every request, interpretation, retrieval and automated action can consume tokens, cloud capacity and third-party services. The building may now incur a charge each time it is asked to reason. At portfolio scale, that could be meaningful.
A single building can generate thousands of alarms, sensor readings, work orders, occupant requests and control events each day. An AI system that reviews every change in temperature, repeatedly reinterprets familiar faults or loads an entire operations manual into every query can create substantial expenditure without improving performance. Add several agents that consult one another, revisit earlier conclusions and continuously monitor the same systems, and the apparent efficiency of automation can quickly acquire the economics of a very enthusiastic junior consultant.
The problem is architectural as much as commercial. Many tasks within a building do not require generative AI. A threshold can identify that a temperature is too high. A control sequence can respond to occupancy. Conventional analytics can detect a drifting sensor or an inefficient operating pattern. Language models become valuable where the task involves ambiguity, interpretation, synthesis or interaction across fragmented information.
The sensible design principle is therefore selective intelligence. Existing control logic should continue to manage predictable and safety-critical processes. Established analytical tools should handle structured data and repeatable diagnostics. Language models should be reserved for questions that genuinely require contextual reasoning, and agents should be deployed where the value of autonomous action exceeds the additional computational and governance cost.
This distinction will become increasingly important as smart building platforms incorporate copilots and agents as standard features. Owners/operators will need to understand what triggers a model call, which model is used, how much context is processed and whether the system can route simpler tasks to lower-cost models. They should also ask whether recurring queries are cached, whether agents have defined limits and what happens when consumption exceeds the expected budget.
Procurement processes are not yet well equipped for these questions. A fixed subscription may conceal underlying usage until renewal. A usage-based contract may transfer all the risk to the customer. A demonstration that performs impressively across a handful of use cases may become financially unattractive when deployed across millions of square metres. The cost model must therefore be tested against real operational volume, including alarms, false positives, repeated requests and exceptional events.
This is where real estate may need its own version of AI FinOps - the disciplined management of model consumption, infrastructure cost and business value. Tokens alone will be a poor measure. More computation can be entirely justified when it prevents plant failure, reduces energy consumption or saves hours of engineering effort. The better metrics will relate expenditure to outcomes, maybe the cost per fault correctly diagnosed, per work order avoided, per occupant request resolved, or per unit of energy saved?
Those measures would also expose weak use cases. An AI agent that produces polished summaries of information already available on a dashboard may generate activity without creating meaningful value. One that identifies a failing chiller weeks before breakdown may be worth considerably more than its computing bill.
The next generation of smart building strategies will need to consider computational demand alongside energy demand. Owners already ask how much power a building consumes and what operational benefit it produces. They should begin asking the same of its intelligence.
A building does not become smarter because it calls an AI model more frequently. Its intelligence will be judged by the quality of the decisions it makes, the value those decisions create and the cost of asking it to think.
In Dr Marson’s monthly column, he’ll be chronicling his thoughts and opinions on the latest developments, trends, and challenges in the Smart Buildings industry and the wider world of construction. Whether you're a seasoned pro or just starting out, you're sure to find something of interest here.
Something to share? Contact the author: column@matthewmarson.com
About the author:
Matthew Marson is an experienced leader, working at the intersection of technology, sustainability, and the built environment. He was recognised by the Royal Academy of Engineering as Young Engineer of the Year for his contributions to the global Smart Buildings industry. Having worked on some of the world’s leading smart buildings and cities projects, Matthew is a keynote speaker at international industry events related to emerging technology, net zero design and lessons from projects. He is author of The Smart Building Advantage and is published in a variety of journals, earning a doctorate in smart buildings.