AI in facilities management: 10 use cases for the connected workplace

AI in facilities management has one genuinely well-documented anchor result: machine learning cut the energy Google used for cooling its data centers by up to 40 percent^[1]. The catch is that the result came from the most instrumented buildings on earth, run by the company that built the model. Everything a facilities leader should buy — and refuse to buy — follows from understanding that gap.

Buildings are attractive AI terrain for two structural reasons. First, they already generate operational data: building management system (BMS) points, utility meters, work-order history, badge swipes, and camera feeds exist whether or not anyone analyzes them. Second, the labor that operates buildings is scarce and getting scarcer, so anything that targets technician hours better has a real economic base. But facilities is also a category where vendor claims have far outrun public evidence, and where the honest, citeable record is thin. This guide works only from that record: what the primary sources actually establish, which of the ten use cases inherit that evidence directly, and which are extrapolations you should price as such.

By the numbers

The flagship result — and why it does not transfer for free

In July 2016, DeepMind reported that applying machine learning to Google's own data centers had reduced the energy used for cooling by up to 40 percent, which translated to a 15 percent reduction in overall PUE overhead after accounting for electrical losses and other non-cooling inefficiencies^[1]. In that first phase, the system produced recommendations that human operators implemented. By August 2018, DeepMind had closed the loop: the AI system was directly controlling data-center cooling on a five-minute cycle, with every proposed action vetted against an internal list of safety constraints and verified again by the local control system, and with operators able to exit AI control at any time. Running autonomously, it delivered consistent energy savings of around 30 percent on average^[2].

This remains the strongest public evidence that machine learning can materially cut building energy use, and it is a legitimate reason to take the category seriously. It is not a forecast for your portfolio. Hyperscale data centers are single-purpose buildings with dense, calibrated sensor coverage, a continuously tracked success metric, and one organization that owns the building, the data, and the model. A typical commercial portfolio has none of those properties, and the delta between the two environments — not the headline percentage — is the correct starting point for a facilities AI business case.

The ten use cases, grouped by the data they stand on

The ten use cases below are grouped into four families by their dominant data dependency, because data readiness — not vendor category — is what predicts whether a deployment produces value in a quarter or stalls in integration for a year. Within each family, the ordering runs from the use case with the strongest existing data foundation to the one that demands the most new instrumentation or organizational change.

Energy and comfort (use cases 1–3)

The first family runs on data almost every portfolio already has: meter feeds, BMS trend logs, weather, and occupancy schedules. Use case 1 is energy anomaly detection — models that learn a building's consumption baseline and flag deviations such as HVAC serving unoccupied zones, equipment idling outside operating hours, or simultaneous heating and cooling. It is the right first purchase because it needs no new control integration and its findings are individually verifiable by an engineer. Use case 2 is demand forecasting and load shifting: predicting consumption by the hour from weather and occupancy, then moving flexible loads away from peak-tariff windows. Use case 3 is HVAC control optimization, the direct descendant of the DeepMind work — models that continuously tune setpoints and equipment staging against comfort constraints.

The sequencing lesson from the primary record matters more than any single number: DeepMind ran two years in recommendation mode, with humans implementing the model's suggestions, before it let the system control equipment directly — and when it did, it wrapped the controller in a two-layer safety check and kept a human exit at all times^[2]. A facilities team should replicate that arc deliberately. Recommendation mode builds the trust, the baseline, and the evidence of model quality; closed-loop control is an earned second phase, not a day-one feature to buy.

Asset care (use cases 4–5)

Use case 4 is automated fault detection and diagnostics (FDD): rule-based and machine-learning hybrids that watch BMS streams and rank faults — stuck dampers, drifting sensors, short-cycling equipment — by energy and comfort impact, turning a reactive call queue into a prioritized work list. Use case 5 is predictive maintenance for building mechanical assets: chillers, air handlers, pumps, and elevators instrumented with vibration, temperature, and current sensors feeding failure-prediction models. The modeling mechanics — vibration analysis, anomaly detection, model lifecycle — are covered in depth at /use-cases/ai-manufacturing-guide and are not repeated here; the facilities-specific decision is different. In a plant, the asset owner usually employs the maintenance team. In a building, asset care is often contracted out, so the buyer must decide who receives predictions, whether the service contract rewards acting on them, and who is accountable when a predicted failure is ignored.

The economic backdrop for this family is labor scarcity, and here a verifiable number helps. The US Bureau of Labor Statistics counts 425,200 HVAC mechanic and installer jobs in 2024, projects 8 percent employment growth from 2024 to 2034 — faster than average — and expects about 40,100 openings per year over the decade, largely from replacement needs^[3]. Skilled building-trades hours are the binding constraint, and they are not getting cheaper. The honest framing for maintenance AI is therefore triage: routing scarce technician time to the equipment most likely to fail, rather than the headcount-reduction framing some business cases still lead with.

Space and services (use cases 6–8)

Use case 6 is occupancy analytics and space optimization: people-counting sensors, Wi-Fi or badge signals, and booking-system logs reconciled to show which floors, rooms, and desks are actually used versus reserved. This is the use case with the most direct line to a CFO-legible outcome — lease consolidation — and it is best timed against upcoming lease events, when the analysis can change a real decision. Use case 7 is condition-based cleaning and service routing, which converts the same occupancy signal into dynamic janitorial schedules: high-traffic zones serviced more often, empty ones skipped. The labor pool here is large — 2,447,700 janitor and building-cleaner jobs in 2024, with about 351,300 projected openings each year^[4] — so even modest routing gains compound across a portfolio. Use case 8 is service-request automation: language-model assistants in front of the ticketing system that classify, answer, and route routine requests, escalating the rest.

This family carries the sharpest governance edge in the list. Occupancy sensing is workforce surveillance if it is deployed carelessly, and a facilities team that quietly installs desk sensors will spend its credibility on the wrong argument. Aggregate, anonymized counts are sufficient for every space decision that matters — portfolio sizing, floor consolidation, cleaning frequency — and are far easier to defend to employees, works councils, and counsel than anything resolving to an individual. Decide the data-minimization posture before the purchase, not after the first internal complaint, and put the service-desk assistant under the same output-quality review you would apply to any customer-facing model.

Security and the integration layer (use cases 9–10)

Use case 9 is physical-security analytics: machine learning over access-control logs and camera feeds to flag anomalies — after-hours access to sensitive zones, credential sharing, tailgating — so security teams review prioritized events instead of raw footage. It inherits both the strengths and the burdens of the space family: the data largely exists, and the privacy obligations are stricter still. Use case 10 is the digital twin: a live, queryable model of the building that joins the other nine data streams into one addressable graph.

The twin is the most hyped item on this list, so it pays to be precise about what the major platforms actually are. NIST's economic study of the technology defines a digital twin plainly: a digital computer model of a physical system — such as a machine or building — with high accuracy, precision, and flexibility to model various aspects of the system^[5]. Azure Digital Twins is a platform service for building twin graphs of entire environments — buildings, factories, farms, energy networks, stadiums — in which you define models such as a Building, Floor, and Elevator type and wire them to live IoT data^[6]. AWS IoT TwinMaker takes an entity-component approach aimed at keeping track of a physical factory, building, or industrial plant, with built-in connectors for sensor and video data and scenes organized, for example, one per floor of a facility^[7]. Both are integration substrates: they make the other nine use cases easier to compose, and deliver nothing by themselves.

On the economics, the best public analysis is instructive for what it does and does not cover. NIST estimates the potential impact of digital twins at $37.9 billion annually if fully adopted across the US manufacturing industry, with a deliberately downward-biased Monte Carlo simulation putting the 90 percent confidence interval between $16.1 billion and $38.6 billion^[5]. That is a manufacturing-scoped estimate; no equivalent public, methodologically transparent figure exists for commercial buildings. So treat the twin as a cost center until it is attached to a named question it will answer — which assets to instrument first, where the next retrofit dollar goes, how a floor consolidation changes HVAC load — and let the value of answering that question, not the platform, carry the business case.

Honest objections

The strongest objection to this entire category is that the public evidence base is one operator's own buildings. The DeepMind results were achieved and reported by the same organization that owned the facilities, and were never subjected to third-party audit or published as peer-reviewed research with replicable methodology. Nearly every other number in circulation traces to vendor marketing or to analyst reports whose methodology is not public. A skeptical CFO who says 'show me an audited result from a building like ours' is asking a question the industry mostly cannot answer yet — which argues for small, measured, self-baselined pilots rather than portfolio-wide platform commitments.

A second objection is economic: integration cost dominates, and it is systematically underestimated. The table above is really a list of integration liabilities — mislabeled BMS points, undocumented control sequences, protocol translation, split incentives with service contractors — and each one is paid for before the first model produces value. The NIST digital-twin analysis is candid that cost-effectiveness depends on circumstances and treats the investment decision, not the technology, as the core problem^[5]; buyers should import that discipline even for the modest use cases. A third objection is the labor narrative: BLS projects growth, not decline, in the trades that run buildings^[3], so a business case built on replacing operating labor contradicts the official employment outlook. The defensible case is throughput per scarce technician hour — and if a proposal cannot state its benefit in those terms or in metered energy, it is not ready for funding.

The read

For a facilities, real-estate, or IT leader deciding where to start: buy in the order of data readiness, not ambition. Energy anomaly detection and FDD come first — they run on data you already have, and every finding is independently checkable. Space analytics comes next, timed to lease events and gated by a privacy posture you have already settled. Maintenance AI follows where assets are instrumented and the service-contract incentives are aligned. Closed-loop energy optimization is an earned phase two, adopted with the safety architecture the primary record describes. The digital twin comes last, and only against a named question. And every one of these ships a production model whose alerts decay: drift detection, alert-quality review, and retraining cadence are the same operating discipline detailed at /guides/model-monitoring-production-guide, and an alert stream without an owner and a response SLA is a liability, not an asset.