AI in facilities management: what works today, what is hype

Updated: 19 October 2026 · 8 min read

Every FM software vendor now says "AI". Most of the articles that rank for the topic open with the same promise: sensors on every machine and failures predicted before they happen. For a Saudi FM company that still runs its jobs on Excel and WhatsApp, that is the wrong place to start. This guide separates what AI does reliably today from what is still hype, explains why the gap is mostly about data, and gives you eight questions to ask before you pay for anything.

What does AI do well in facilities management today?

The AI that works now is the kind that reads and writes text. You already have that text: residents' complaints, technicians' notes, work order histories, contracts and invoices. None of these needs a sensor.

Capability	Why it works now
Understanding and translating Arabic and English requests	Language models handle Arabic and English text well enough for daily use, and translation is already one of the most common work uses of generative AI among Saudi employees
Sorting tickets by trade, priority and location, with a person confirming	Works on the words in the ticket; research shows good but imperfect accuracy, so a supervisor confirms
Summarising a long ticket thread or an asset's job history	Text in, text out: generative AI's strongest skill, and easy for a person to check
Answering plain-language questions about your own data ("how many tickets are past SLA in Riyadh this week?")	Needs only the records your system already holds; low risk when the assistant is read-only
Suggesting the best technician for a new job, with a reason, for a supervisor to approve	Uses data FM software already holds: trades, documents, contracts, availability
Reading fields out of invoices, contracts and certificates	A mature task for language and vision models, and errors are easy to spot on review
Rule-based alerts (overdue PPM, repeat failures on one asset, warranty expiry)	No trained model needed: simple expert rules already prevent many failures

The last row matters. A Deloitte position paper on predicting equipment failures notes that simple rules built from expert knowledge "may prevent a large proportion of failures" before any modelling at all. For most FM companies, well-chosen rules are the first step, not a second-best.

What is still hype for most Saudi FM companies?

Claim	Why it is not ready for most firms
"Predict failures before they happen" across the whole portfolio	Needs sensors on the equipment and a history of failures logged with exact dates and times. Feasible for a few critical assets with a building management system; not across hundreds of buildings
Building-wide digital twins	Need an accurate asset register, live sensor feeds and integration between systems. Savings figures quoted in the region are mostly conference statements, not studies
Fully autonomous AI agents running operations	Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, and warns about "agent washing": chatbots relabelled as agents
Camera or drone inspections at scale	Need large sets of labelled images and consistent capture; they work in narrow pilots, not as a general feature
Instant return on investment	JLL's 2025 technology survey found 88–92% of real-estate firms piloting AI, but only 5% had achieved all their programme goals
AI that understands every dialect and abbreviation	Dialect handling is a documented gap, and mixed Arabic and English asset names still need a person to check

None of this means these ideas are wrong. It means they come later, once the basics are in place.

Why does predicting failures need data most firms don't have?

Predicting failures before they happen is the use case every vendor leads with, so it is worth being precise. The often-quoted benefits (10–20% more equipment uptime, 5–10% lower maintenance cost, 20–50% less planning time) come from Deloitte's analysis of industrial maintenance, not from FM portfolios. The same paper is blunt about the price of entry: data is "the limiting factor", failures recorded "with exact date and time are the very basic requirement", and businesses "need to add sensors to their machines" plus the IT and trained people to use them.

Ask yourself how many breakdowns last year were logged against a specific asset, with the date, time and cause. If the honest answer is "few", the first AI project is not prediction. It is getting every job into one system with the asset, the time and the result recorded. That is what makes every later project possible. A CAFM system is mostly about exactly this.

How well does AI handle Arabic and Saudi dialect?

Better than two years ago, but not perfectly. Arabic-first models now exist, including ALLaM-based services hosted in the Kingdom with Arabic dialect input and switching between Arabic and English. At the same time, a 2025 benchmark across five Arabic dialects, Saudi included, found "persistent gaps" in how models handle dialect compared with Modern Standard Arabic.

In practice, AI reads a resident's message in Gulf dialect well enough to pick the trade and the building most of the time. Where it struggles is site shorthand, mixed Arabic and English equipment names and voice notes recorded in a noisy plant room. That is why the rule for every AI step in maintenance is the same: the AI suggests, a person confirms.

What do Saudi rules expect from AI?

There is no FM-specific AI regulation, but national guidance is clear. SDAIA's AI Ethics Principles and its generative AI principles set out accountability, transparency, reliability and privacy, and warn against data leakage, misinformation and bias. The Personal Data Protection Law applies to residents' names, phone numbers and complaints as soon as they reach any system, AI or not.

For an FM company this comes down to three questions: who approves what the AI suggests, can you explain any decision to a client, and where is your data processed. We go into the approval question in detail in AI work order dispatch.

8 questions to ask before you buy AI for maintenance

Is every breakdown logged with the asset, the date and time, and the cause? If not, predicting failures is premature.
Do you have one asset register, or several spreadsheets?
Do requests reach you as text you can see, or only as calls and voice notes?
Do any assets have sensors or a building management system connection? Which ones?
Who approves what the AI suggests? Name the role.
Can the vendor show the reason behind each suggestion?
Where is the data processed and stored, and does that meet the PDPL and your clients' contracts?
Which result will you measure after three months: response times, repeat visits, maintenance cost or energy?

If a vendor cannot answer questions 1, 6 and 7 plainly, the "AI" is marketing.

How Mubaney uses AI

Mubaney uses AI where it works on data an FM company already has. An AI assistant answers questions about your own data in Arabic or English and is read-only. AI dispatch proposes the best qualified person for each new ticket with a reason, and a supervisor approves it; it is switched on for each company once the data behind it is in place. Mubaney's asset condition alerts are rule-based, and it does not predict failures from sensor data. To start with the basics, see what a CAFM system does.

FAQ

How is AI used in facilities management?

Today it is mostly used on text: understanding and translating requests, sorting tickets, summarising job histories, answering questions about your own records and suggesting technicians. Energy optimisation and failure prediction from sensors are also used, but mainly where buildings have sensors and a building management system.

Is predicting equipment failures realistic for Saudi FM companies?

For a few critical assets with sensors and a clean failure history, yes. Across a whole portfolio, usually not yet: it needs every failure logged with asset, date, time and cause, plus sensor data, which most firms do not have. Start by recording every job properly.

Can AI understand Arabic maintenance requests?

Mostly. Current models read Modern Standard Arabic and Gulf dialect well enough to identify the trade and location in most requests, but studies still find gaps in dialect handling, and site shorthand and voice notes cause errors. A person should confirm the AI's reading.

Should AI make maintenance decisions on its own?

No. Saudi national guidance from SDAIA expects human oversight and clear accountability, and research on maintenance records recommends a human in the loop because no model is fully accurate. Let AI suggest and a supervisor approve.

What should an FM company do first with AI?

Get every request and job into one system with the asset, time and outcome recorded, then use AI on that text: sorting tickets, summaries, questions about your data and dispatch suggestions. Leave sensors and digital twins for later, for the assets where they pay off.