Two facts about AI in building operations, and they appear to contradict each other. Kent State University documented roughly $470,000 in annual savings from AI-driven monitoring of its campus systems, as reported by Facilities Dive. McKinsey, surveying the broader market, found that only 5% of AI programs in commercial real estate achieved their stated objectives.

Same category of technology. Opposite outcomes. Which means the interesting question isn’t “does AI monitoring work” — both facts answer that — it’s under what conditions. And that’s a decision framework, because the conditions are choices made before any software is bought, and the cost of choosing wrong isn’t just a wasted license. It’s a burned organization that won’t try again for five years.

The interesting question isn’t “does AI monitoring work”—it’s under what conditions. The conditions are choices made before any software is bought.

Profile A: The conditions behind the $470K

Strip the Kent State story to its operating conditions and a pattern emerges that has little to do with algorithms.

  • The data existed before the AI did. Campus systems were instrumented — building automation, metering, equipment telemetry — and the asset inventory was real. The AI was pointed at signals that already flowed. It was the analysis layer on top of a data layer, not a substitute for one.
  • The scope was specific. Energy and equipment fault detection — drifting setpoints, simultaneous heating and cooling, equipment running off-schedule, early-stage mechanical faults. Problems with known signatures and known fixes. Not “transform our operations.”
  • Someone owned the output. Detections became work orders that someone was accountable for executing. This is the quiet hinge of the whole case: an alert that doesn’t become a work order is a notification, and notifications don’t save $470,000. Closed loops do.
  • The savings were measured against a baseline. The organization knew its costs before, which is the only way “saved $470K” can be a sentence anyone can defend.

Profile B: The conditions behind the 95%

Now the failure profile — assembled from the same market research and, frankly, from the stories operators told us in interviews about pilots that died quietly.

  • The AI was bought to avoid building the data layer. Asset registries incomplete or fictional, histories in spreadsheets, sensors absent — and a hope that the algorithm would somehow compensate. It can’t; a model pointed at missing data produces confident nonsense, and the spreadsheet era destroyed the history at creation.
  • The scope was a mission statement. “AI-driven operational excellence” — no defined fault classes, no target assets, no number to beat.
  • The output had no owner. A dashboard was stood up; viewing it was nobody’s job; alerts accumulated like unread email until everyone agreed, without a meeting, to ignore them.
  • No baseline existed. Even genuine wins were unprovable — and unprovable wins don’t survive budget season.

The variable that decides

Read the two profiles side by side and the deciding variable isolates cleanly: it was never the AI. The same class of software sits in both stories. What differs is whether the organization built the loop around it — instrumented assets feeding clean data in, and owned work orders carrying detections out. AI monitoring is the middle of a pipeline. The 5% built the pipeline; the 95% bought the middle and waited.

STAGE 1 Data Layer First

Ensure critical assets emit data today and real inventories exist before applying AI analysis.

STAGE 2 Detection Layer Second

Point the intelligence at a specific fault list rather than a vague vision statement, comparing against a known cost baseline.

STAGE 3 Owned Output

Guarantee that when the system detects an anomaly at 2 a.m., it seamlessly creates a work order with an accountable owner.

This reframes the investment decision into a readiness checklist any operator can run before signing anything: Is the asset inventory real? Do the critical assets emit data today? Is there a baseline cost to measure against? Is the scope a fault list or a vision statement? And the decisive one — when the system detects something at 2 a.m., whose work order is it?

In predictive maintenance for commercial buildings, the honest sequencing is data layer first, detection layer second, and that ordering is exactly why Sweven FM installs the work-order and sensor infrastructure before any intelligence sits on top — the loop is the product; the AI is a component.

The Readiness Audit

So before the next vendor demo, run your operation against the five questions. If the answers are mostly no, the good news is that the $470K story is still available to you — it just starts a layer lower than the brochure suggests. Which layer is your operation actually on?

Sources: