McKinsey’s 2024 analysis of AI adoption in commercial real estate found that only 5% of AI programs achieved their stated objectives. That number isn’t a technology indictment. It’s a deployment indictment.

The technology works in the use cases where it has been correctly deployed — with clean data infrastructure, defined action thresholds, and integration into operational workflows. It doesn’t work where it was bolted on top of a system that wasn’t ready for it, or where the goal was “use AI” rather than “solve this specific operational problem with AI.”

Understanding where AI genuinely helps FM operations today requires drawing a hard line between what the pitch promised and what the operational evidence shows.

→ How technology integrates with real FM operations: Predictive Maintenance for Commercial Buildings

The technology works in the use cases where it has been correctly deployed. It doesn’t work where the goal was “use AI” rather than “solve this specific operational problem with AI.”

AI Use Case A: Pattern Detection in Work Order History

What it does: AI applied to work order history identifies recurring issues on the same asset, the same vendor, or the same location — patterns that are invisible to the human eye when spread across hundreds of work orders over 24 months but statistically clear when processed at scale.

What this looks like in practice: A commercial portfolio with 40 locations generates thousands of work orders per year. A specific HVAC unit at location 12 has generated 7 corrective work orders in 18 months — but they’ve been logged by three different technicians under slightly different descriptions, so no single person ever saw the pattern. AI pattern detection surfaces it as an asset flag in the next morning’s dashboard review.

Why it works here: The data already exists in the work order system. AI doesn’t need to collect new data — it needs to read existing data at a scale and speed that humans can’t match. This is the use case where AI is genuinely replacing a task that was previously impossible, not just faster.

What it requires: Clean, consistent work order records with asset IDs linked to individual work orders. If assets aren’t properly tagged in the work order system, the pattern can’t be detected. Data quality is the prerequisite. Johnson Controls’ 2026 AI and FM Report identifies integration and data quality as the number one friction point in FM AI adoption — cited by one third of business leaders as the primary obstacle.

AI Use Case B: Predictive SLA Breach Detection

What it does: AI monitors open work orders against SLA windows in real time and predicts which work orders are at risk of breach before the breach occurs — based on vendor response time patterns, work order age, and historical breach frequency for that vendor/trade combination.

What this looks like in practice: A work order for HVAC PM at location 8 was dispatched to a vendor with a 24-hour response SLA. The AI knows this vendor’s historical response time for HVAC PM work orders at this location averages 31 hours. At hour 18, the system flags the work order as breach risk and generates an escalation prompt — before the SLA window closes, not after.

Why it works here: SLA breach prediction requires processing historical patterns across many variables simultaneously: vendor, trade, location, time of week, work order type. That’s exactly what ML models do efficiently. A human reviewer checking each open work order manually can’t process those variables at the same speed or consistency.

What it requires: SLA windows defined in the work order system and a sufficient history of vendor response data — typically 6 to 12 months of tracked work orders per vendor to build a reliable prediction model.

AI Use Case C: Automated Work Order Triage and Routing

What it does: AI classifies incoming work order requests by urgency and impact at intake, assigns a priority tier, and routes to the appropriate vendor or technician based on trade, certification, location, and availability — without requiring a human dispatcher to review each request individually.

What this looks like in practice: A tenant submits a maintenance request through the building portal. The AI reads the description, identifies it as an HVAC airflow issue (not an emergency), classifies it as Tier 2, matches it to the vendor with current EPA 608 certification who has available capacity at that location this week, and generates the dispatch — notifying the FM for review before sending, or sending autonomously if the work order type falls within pre-approved automated dispatch rules.

Why it works here: Triage and routing is a rule-based decision with defined inputs. AI executes rule-based decisions faster and more consistently than humans — and doesn’t have bandwidth constraints that create backlogs when volume spikes. Facilities Dive’s 2026 FM predictions identified agentic AI — systems that can query, detect, route, and close work orders semi-autonomously — as the most significant near-term AI development in FM operations.

What it requires: Defined dispatch rules, vendor certification data in the system, and FM-set approval thresholds that determine which work orders route automatically vs. require human review.

Where AI in FM Still Doesn’t Deliver

Predictive failure modeling without sensor data is the most common oversold AI capability in FM. The pitch: AI will predict equipment failures before they happen. The reality: AI predicts failures based on sensor data that shows the asset’s condition in real time. Without sensors on the equipment, there’s no signal to process. AI applied to maintenance logs alone — without condition data — can identify historical patterns but cannot predict real-time failure timing with useful accuracy.

AI dashboards without connected action produce reports that nobody has time to act on. An AI dashboard that shows 14 assets at elevated risk, presented to an FM team that has to manually create a work order for each one, follow up with vendors, and track completion separately — that’s a better report, not a better operation. The AI use cases that deliver in FM are the ones where the AI output connects directly to an automated action: a flag that creates a work order, a SLA risk that generates an escalation, a pattern that triggers a diagnostic PM.

Kent State University’s AI-monitored maintenance program — documented by Facilities Dive in 2026 — saved $470,000 annually. The conditions that made it work: clean historical data, sensor infrastructure on critical assets, and AI output connected to the work order system so alerts generated actions automatically, not additional review tasks.

The Question That Separates the 5% From the 95%

Before deploying AI in any FM use case, one question determines whether the deployment will succeed: what specific operational decision will this AI output change, and what data infrastructure does it require to produce that output reliably?

If the answer to the first part is “it will give us better visibility” — that’s a dashboard. Dashboards don’t change operations. Connected actions do. If the data infrastructure required doesn’t currently exist — clean asset IDs, consistent work order records, SLA windows defined in the system, sensor data on critical assets — the AI layer has nothing to process. Deploying AI on top of incomplete data produces incomplete AI.

STAGE 1 Operational Decision

Identify the specific workflow or decision the AI is meant to change, avoiding vague goals like “better visibility.”

STAGE 2 Data Infrastructure

Establish clean historical records, consistent asset IDs, and active sensor data before turning on the AI layer.

STAGE 3 Connected Action

Link the AI’s predictions and pattern detections directly into automated work order creation or SLA escalations.

The Deployment Reality

The 5% of FM AI programs that succeeded started with the operational decision, built the data infrastructure, deployed the AI, and connected the output to automated action. The 95% that didn’t reversed that order. Deploying AI on top of incomplete data produces incomplete AI.

Sources: