AI Computer Vision for Facility Management: Where It Works Today

Facility management runs on a lot of imagery that mostly goes unread. Tenant complaint photos, technician completion photos, statutory inspection photos, contractor job-sheet photos, roof and façade survey drone footage, CCTV streams from lobbies and plant rooms. Every image is a data point that a competent computer vision system could interpret. The technology has moved to production over the last two years, but adoption in FM remains uneven. This post covers the FM-specific use cases that actually work today, the ones that are still fragile, and how facility leaders should think about buying and deploying vision into their operations.

Scope note: this article is the FM-specific variant of the broader Computer Vision for FM and CMMS hub. The utilities-CMMS variant covering SCADA-adjacent contexts, industrial gauge reading, corrosion detection on pipe networks and autonomous drone inspection of transmission lines is a separate post linked at the bottom.

Tenant complaint photo triage

Most FM helpdesks now accept photos with tenant complaints, and most of those photos land in a queue that a helpdesk operator sifts manually. Computer vision changes the economics. A model trained on the past year of complaint imagery can classify most incoming photos automatically: leak in a meeting room ceiling, blocked toilet, faulty lighting fixture, damaged furniture, cleaning issue in a shared area. The classification pre-populates the work order type, priority and likely trade, and the operator confirms rather than composes.

The cycle-time impact is significant. Where an operator today spends five to ten minutes classifying and routing each complaint, a vision-assisted flow compresses that to seconds, with human confirmation catching the small percentage the model gets wrong. For a mid-size FM operation handling several thousand complaints a month, the productivity gain is measured in reclaimed operator time, faster tenant response, and more consistent classification (which the reliability analyst downstream will care about a year later).

Statutory PPM evidence capture

Statutory PPM lives on photographic evidence. The LOLER inspection on a passenger lift, the fire door survey, the emergency lighting test, the fire risk assessment, the water hygiene audit all produce photographs that end up attached to the compliance record. Traditional workflow: engineer photographs, engineer describes, engineer files. Vision-assisted workflow: engineer photographs, model tags the equipment, model auto-drafts the descriptive note, engineer confirms.

The bigger opportunity is cross-inspection consistency. Over years of statutory records, similar equipment gets described inconsistently by different engineers, which makes cross-portfolio trend analysis unreliable. A vision layer standardises the description language across all engineers and all portfolios, so ten years of records become a queryable data asset rather than a filing-cabinet archive.

The honest caveat: statutory records have regulatory weight. AI-generated descriptions must be reviewed and confirmed by the qualified inspector, not accepted blindly. The vision layer speeds capture; it does not replace the qualified sign-off. See the AI governance pillar for the broader framing.

Façade, roof and structural inspections

Building envelope inspections are a growing computer vision use case. Drone flights around a building façade produce hundreds of photographs. Vision models flag defective sealant, missing pointing, cracked panels, discoloured or damaged rendering, and displaced roof elements. What used to be a specialist survey engineer walking a scaffold or reviewing every drone photo manually is now the engineer reviewing only the model-flagged exceptions.

Roof inspections have similar economics. Vision models trained on roof imagery can flag ponding water, damaged waterproofing, blocked drainage, moss growth, missing tiles or panels, and general deterioration. Combining vision with periodic drone flights creates a longitudinal record of the roof condition that trending against previous flights reveals patterns invisible in single snapshots.

For portfolio-scale FM operations (hundreds of buildings), this changes the frequency and cost of envelope inspection. What was previously an annual or biennial survey per building becomes quarterly drone flights with vision-assisted review, at a fraction of the historical cost per inspection.

PPE compliance and security monitoring

FM sites (particularly mixed-use developments, tenant fit-out projects, and buildings undergoing major maintenance) increasingly need visible PPE compliance monitoring. Fixed cameras at site entrances and inside contractor work areas can detect whether personnel have appropriate PPE (hard hats, high-visibility vests, safety footwear) for the zone. Real-time alerts on non-compliance shift the safety function from post-incident review to preventive intervention.

The related use case is broader security and site-integrity monitoring. Unauthorised access to plant rooms, blocked emergency exits, tampering with fire safety equipment, obstructed circulation routes. All of these are visible-anomaly patterns that a computer vision layer atop existing CCTV can detect and alert on, without adding cameras.

Honest caveats: PPE detection deployed without deliberate policy design creates friction with contractors and workers. Focus on aggregate compliance trends and pattern detection rather than individual-worker enforcement. Security monitoring must respect data-protection regulations (GDPR in Europe, similar frameworks elsewhere); pattern detection without stored personal imagery is usually the right implementation choice.

Asset identification during walkarounds

A common FM data-quality gap: the CAFM asset register drifts out of sync with reality because assets get moved, replaced, or fall out of tracking. Vision-assisted walkarounds close this gap. A technician walking a floor points a phone camera at equipment; the vision model identifies each asset, cross-references against the CAFM register, and flags mismatches for review.

Combined with nameplate OCR, this makes bulk asset re-baselining a task of weeks rather than months. For FM operations that have accumulated a decade of asset-register drift, this is one of the fastest-payback vision use cases. (Broader context in the CAFM data migration pillar and the asset hierarchy design pillar.)

Where vision is still fragile in FM contexts

Four limitations specifically relevant to FM operations:

• Indoor lighting varies wildly: office, retail, restaurant, plant room, tenant kitchen. Model accuracy degrades in mixed-lighting environments unless the training data covers the range. Plan for site-specific tuning if imagery quality is uneven.
• Furniture, occupants and clutter obscure assets: office and tenant environments contain a lot of visual noise that plant rooms and industrial contexts do not have. Vision accuracy for identifying wall-mounted sensors, ceiling-mounted diffusers, and similar assets is meaningfully lower when the environment is cluttered.
• Multi-tenant privacy is a real constraint: monitoring cameras in tenant-occupied areas raise privacy issues that plant-room monitoring does not. Design vision deployments to respect tenant boundaries deliberately, not as an afterthought.
• Some FM failure modes are invisible visually: BMS sensor drift, water-loop chemistry, indoor air quality, structural fatigue beneath finishes. Vision helps with what is visible; pair with sensor-based monitoring for what is not.

Where FM leaders should start

Three practical steps:

1. Pilot tenant complaint photo triage first. High-volume, low-risk, immediately visible operational benefit. If it works at one building, it works at ten. If it fails at one building, you learn cheaply.
2. Then extend to statutory PPM evidence capture. Higher stakes but well-scoped. The qualified inspector remains accountable; the vision layer just makes their capture and description work faster and more consistent.
3. Consider drone-based envelope inspection for portfolio-scale operations. If you manage 20+ buildings, the economics of vision-assisted drone surveys are meaningfully better than traditional surveying. Below that portfolio size, traditional inspection typically still wins.

For the broader FM AI frame, see the AI copilot for facility management pillar and the NLP for facility management pillar. For the utilities-specific vision variant covering pipeline corrosion, gauge reading and transmission-line drone inspection, see the utilities CMMS computer vision pillar.

Final thoughts

Computer vision in FM is genuinely useful today for a handful of scoped workflows: tenant complaint triage, statutory evidence capture, envelope inspection, asset identification, PPE compliance. Where FM leaders go wrong is trying to deploy vision across every possible use case at once, or expecting vendor claims about generic "AI-powered inspection" to translate into their specific building environment without site-specific tuning.

The FM operations that get real value from vision are the ones that pick one workflow, do it well, prove the operational pattern, and expand deliberately from there. That is unglamorous compared with vendor marketing but it is what actually works.

Related reading: Computer Vision hub (FM + CMMS), AI copilot for FM, NLP for FM, asset hierarchy design, AI governance for enterprise operators.