AI Agents forFacilities Management 2026

If you read one paragraph of this article, read this one. Vendors selling predictive maintenance and fault detection into commercial buildings commonly claim 20 to 40 percent HVAC energy savings. The largest peer-reviewed field study of automated fault detection and diagnostics — Lin, Kramer and Granderson in Building and Environment vol. 168 (2019), covering 26 organizations, 550 buildings and roughly 97 million square feet — found median whole-building energy savings of 8 percent.

That is roughly a quarter of the top of the marketing range. It is also, and this is the part that usually gets lost, a good number. DOE and Lawrence Berkeley National Laboratory separately put average FDD savings at around 9 percent with two-year paybacks, and estimate that HVAC system and control faults account for about 29 percent of commercial building energy use — roughly 4 to 5 percent of national energy use. Two independent sources landing within a point of each other, on a measure with a two-year payback, is about as solid as evidence gets in building operations.

Vendor claims about fault detection and diagnostics against the peer-reviewed and DOE/LBNL evidence — compiled by Frenchy Digital, 2026.

A median is a median, and the spread underneath it is wide. Buildings that were running badly before — stuck dampers, simultaneous heating and cooling, schedules that never got updated after a tenant left — save far more than 8 percent, because there was more to recover. Buildings that were already well commissioned save less. If your portfolio has not been retro-commissioned in a decade, your expected value is above the median. If you commissioned last year, it is below. That is a portfolio-specific judgement, and it is the judgement the 20-to-40-percent claim is designed to prevent you from making.

Build the case on the median and let the outliers be upside. A facilities program that beats a conservative forecast gets funded again. A program that misses an aggressive one gets audited.

— Frenchy Digital engineering principle

The savings side of the ratio gets all the attention. The cost side is better documented and easier to verify, and it is where most business cases quietly go wrong — because they price software and forget labor.

The same study reports cost data from 27 users. The published medians are $8 per point for base software, $2.70 per point per year in recurring software, and $8 per point per year in labor, against a median deployment of about 1,300 points. Those are the four numbers to take into a budget conversation.

Published median FDD costs, and the arithmetic at a median-sized deployment. The right-hand column is multiplication, not a survey finding.

Two things fall out of that table immediately. First, labor is roughly three times recurring software. The expensive part of fault detection is not the license; it is the person who reads what the system produces and turns it into work someone actually does. A deployment budgeted for software and staffed with nobody is the most common way an FDD program fails, and it fails silently — the system keeps working, the alarms keep arriving, and nothing changes.

Second, point count is the unit of everything. Cost scales with it, integration effort scales with it, and alarm volume scales with it. Any vendor pricing per building rather than per point is hiding the variable that determines whether the deployment is affordable. Get the point count from your controls contractor before you get the quote.

The other number every facilities plan is built on is office occupancy, and it is almost universally misread. Kastle's Back to Work Barometer is the most widely cited series in US commercial real estate. It is a genuinely useful dataset. It does not measure what most people planning against it think it measures.

Start with the readings. For the week of Thursday July 24 to Wednesday July 30, 2026, the ten-city average was 55.6 percent, down 0.9 points week over week, with a Tuesday peak of 65.2 percent and San Jose lowest at 46.6 percent. The post-pandemic record came in the week of December 8, 2025: a 56.3 percent ten-city weekly average and a 66.0 percent Tuesday single-day figure.

Kastle Back to Work Barometer readings. Methodology per Kastle: 10 metros, 3,400 buildings, 300,000+ users, anonymized keycard, fob and app data, counting only each person's first entry per day, against an early-2020 baseline of 100%.

Now the part that changes how you use it. The limitations below are not a critic's objections — they come from Kastle's own published methodology, and Kastle is straightforward about all of them. The problem is downstream: the number gets lifted into board decks and lease decisions with the caveats stripped off.

Documented limitations of badge-based occupancy measurement, and their operational consequences.

A badge event is a binary: this person entered the building today. That is genuinely useful for trend, comparison and staffing. It is not sufficient for any decision about how much space you need, because the two things that determine space demand — how long people stay and where they sit — are exactly what the metric does not capture.

Two employees badge in at nine. One leaves at ten-thirty after a single meeting. One stays until six. The barometer counts them identically. Multiply that across a tower and you can have a stable attendance figure sitting on top of a collapsing dwell-time distribution, or the reverse, with no way to tell which from the access-control system. And because only the first entry per person per day is counted, the series cannot see intra-day shape at all — which is precisely what cleaning schedules, catering volumes and HVAC setpoint scheduling depend on.

• Dwell time, not attendance: Time between first badge-in and last badge-out, or a sensor-derived equivalent. This is the input to space sizing. Attendance is the input to whether the building opens.
• Desk-level or zone-level sensing: Passive infrared, thermal counting, or seat-level sensors. This is what answers whether a neighbourhood is genuinely full or holding twelve people in eighty seats. Choose the sensing modality with your privacy counsel before the vendor demo, not after.
• Meeting-room utilization against booking: Booked-versus-occupied is usually the largest single recoverable inefficiency in an office portfolio, and it is measurable with equipment you may already own.
• Intra-day curve: Arrival and departure distribution by day of week. This drives cleaning shifts, catering, elevator dispatch and HVAC scheduling far more directly than any weekly average.
• Badge data as a control, not a conclusion: Use the barometer to sanity-check your own access-control data against a market trend. That is what it is good for. It is a benchmark, not a plan.

There is one figure in the Kastle series that is directly actionable, and it is not the headline. In the record week of December 8, 2025, Class A+ buildings recorded 78.8 percent weekly occupancy and 95.5 percent on that Tuesday, against a ten-city weekly average of 56.3 percent and a 66.0 percent single-day peak. Austin hit 92.9 percent on the Wednesday; Washington DC ran 64.3 percent on the Tuesday; New York averaged 59.5 percent for the week.

For an FM service provider the same figure reads differently and more urgently: the buildings still filling up are the ones where service standards are visible, and they are also the ones with budget. The competitive pressure in this market is concentrated at the top of the stock.

Energy savings are the pitch. Deferred maintenance is the reason anyone is actually buying.

The federal portfolio is the only large building estate whose backlog is published in a form you can audit, which makes it the best available proxy for the direction of the whole sector. GAO reported in April 2025 that the governmentwide deferred maintenance and repair backlog grew from $171 billion in FY2017 to $370 billion in FY2024 — more than double in seven years. A December 2025 GAO report restates the same trajectory as $170 billion to $370 billion.

• $370 billion: Governmentwide deferred maintenance and repair backlog in FY2024, up from $171 billion in FY2017.
• Over $17 billion: GSA's own repair and alteration backlog as of March 2025.
• 277,000: Buildings and structures in the federal real property portfolio.
• Over $10.3 billion: Annual federal operating and maintenance costs in FY2023.
• High-Risk List, 2025: GAO added federal building condition to its High-Risk List in 2025 — the designation it reserves for areas needing transformation.

Read those together and the operating reality is clear enough. When a backlog doubles in seven years while the annual O&M budget grows nothing like as fast, the binding problem stops being efficiency and becomes triage. Nobody is going to fix $370 billion of deferred work. Somebody has to decide, every week, which of the thousand things that need attention get attention — and to be able to explain the decision afterwards.

A note on private-sector benchmarks. The standard non-federal reference for facility condition indices and backlog is Gordian's State of Facilities in Higher Education, and we could not obtain it — it is gated, and we do not cite figures we have not read. Take the federal numbers as directional evidence about a sector-wide condition problem, not as a benchmark for your own portfolio. Your FCI is a number you have to produce yourself.

Nothing below is a new capability. All of it is work your team already does, in the gaps between systems that were never designed to talk to each other — the CMMS, the BAS, the warranty spreadsheet, the vendor COI folder, the parts system, the email inbox where half the work actually originates. That gap is where the hours go, and it is the only place an agent reliably pays for itself.

Facilities workflows suited to agent automation, with the human decision boundary for each — Frenchy Digital, 2026.

Three patterns run through that table and they are worth naming, because they are what separates a workflow that works from one that gets switched off.

• Every one is a lookup-and-route problem: Not a judgement problem. The agent is finding facts scattered across systems and putting them in front of a person with the routing already proposed. It is not deciding what is wrong with the air handler.
• Every one has an explicit human boundary: Priority overrides, purchase approvals, life-safety calls, inspection sign-off, payment decisions. The boundary is written into the workflow rather than assumed, and it is where the agent stops rather than where it asks nicely.
• Every one has a measurable before-state: Time from request to correct trade assignment. First-time fix rate. Percentage of work performed in-house that was under warranty. COI lapse incidents. Missed inspections. If you cannot measure it before, you will not be able to defend it after.

The third pattern is the one people skip, so it is worth being concrete about it. Before any agent is built, capture the before-state for the workflow you are automating. Not an industry benchmark — your own number, from your own systems, for a defined period. This is a week of unglamorous work and it is the difference between a program you can renew and a program you can only defend with anecdotes.

Before-state measurements to capture prior to building — Frenchy Digital facilities engagement checklist, 2026.

A caution that applies to the whole table: measure against your own history, not against a published figure. The sector-wide benchmarks a facilities director would most want here are either gated or absent — as the fault-detection section and the CMMS section both show, the figures circulating freely tend to be the ones nobody checked. Your own last four quarters are a better comparator than any number you can find on a vendor site, and they are the comparator your CFO will actually accept.

Here is how FDD deployments actually die, and it is not the technology failing. The system is installed, configured and commissioned. It works. It detects faults — correctly. It produces two hundred alerts a week across a portfolio, every one of them technically valid. By month three the daily digest goes unopened. By month six nobody can say what the system is currently reporting. The license renews for another year because cancelling it requires admitting something, and the 8 percent never materialises because no fault was ever actually fixed.

Detection was never the hard part. Prioritisation is the hard part, and it is the part an agent is genuinely well suited to, because it is a ranking problem over structured data with a clear objective function.

Turning fault-detection output into prioritized, costed work rather than more alarms — the pipeline Frenchy Digital builds for FDD deployments.

The design target is a number, and it is a small one. A facilities director should open one item per building per week that says: this fault, on this asset, is costing approximately this much per month, the fix is estimated at this much, here is the evidence, here is the proposed work order. Not two hundred alerts. Not a dashboard. One decision, with its arithmetic attached.

If the agent increases the number of things on someone's list, it has failed. The only defensible output of an alarm-triage agent is a shorter list with better reasons.

— Frenchy Digital operations principle

If you are evaluating facilities software or facilities AI right now, you are being sold against two statistics that fall apart the moment anyone traces them. It is the same lesson as the 8 percent, on a different number — and knowing it changes how you read every deck you are shown this quarter.

Two widely repeated facilities-software statistics traced to source, plus a common analyst-citation error.

On the spreadsheet figure. The multi-select problem is the one that matters most. In the 2019 edition, 58 percent of respondents reported having a CMMS while 45 percent also reported using spreadsheets — those are overlapping answers from the same people, not two halves of an industry. A team that runs a CMMS and also keeps a spreadsheet for the PM calendar is a normal, functional team, and it is not evidence that half the sector is on paper. Add that the respondents are manufacturing plant staff rather than commercial-buildings FM teams, and the number has no bearing on your portfolio at all. It is also worth knowing that Verdantix's 2025 commercial CMMS research contains no percentage-on-spreadsheets statistic whatsoever. The leading independent analyst in this space simply does not publish that number.

On the failure rates. Every page asserting 40, 70, 80 or 60-to-80 percent that we could trace is a CMMS vendor selling the alternative. The single attributed version credits Gartner with no report title, no document ID, no date and no link, and no matching document could be found. This is not a claim that CMMS implementations never fail — they plainly do, usually for reasons that have nothing to do with software. It is a claim that nobody has published a credible rate, and that repeating one in a board paper is a checkable error.

One more citation error worth catching. Gartner publishes a Market Guide for Field Service Management (31 March 2025), not a Magic Quadrant — and Market Guides name Representative Vendors rather than Leaders. There is no Gartner FSM Leader to be, and we found no current Gartner Magic Quadrant for enterprise asset management or CMMS. A vendor claiming Leader status in either category is describing something that does not exist.

Model capability is not what determines whether a facilities AI project succeeds. Integration with your existing CMMS is. We have watched more of these projects stall on a write API than on anything to do with the agent itself, and the failure always arrives late — after the pilot, after the budget, when someone finally asks how the work order gets created and the answer is a nightly CSV drop.

Facilities platforms are mostly closed or semi-closed. Read access is common. Write access is not. And an agent that can read your CMMS but cannot write to it is a reporting tool with extra steps — it can tell you which technician should take the job, and then a human retypes it.

The integration due-diligence question set Frenchy Digital runs before scoping any facilities AI engagement, 2026.

Now the cost question, which is unusually hard to answer in this category because most vendors do not answer it publicly. Published list pricing in facilities software is rare. Where it exists, it is worth having in front of you before a renewal conversation.

Published CMMS list pricing as of August 2026, taken from each vendor's own pricing pages. Per user per month unless stated otherwise.

The split in that table is itself the finding.The mid-market tools publish transparent per-seat pricing and put their AI features inside the paid tier rather than behind a separate SKU. The enterprise platforms — Limble, eMaint, Planon, Eptura, ServiceChannel, Corrigo — publish nothing and quote only on request. That is exactly the pattern to expect when you are trying to compare the cost of an AI capability against an incumbent's renewal: on one side of the market you can do the arithmetic yourself, and on the other you cannot begin until you have been through a sales cycle.

Two practical consequences. If you are on a mid-market platform, price the AI you are being sold against the delta between your current tier and the vendor's own paid tier — sometimes the capability you are shopping for is already in a plan you could upgrade to. If you are on an enterprise platform, budget the procurement time as part of the project, and note that IBM's AppPoints model prices capacity rather than seats, so headcount is the wrong variable to forecast with.

One caution that applies to every figure above. List price is a starting point, not a forecast. Seat counts, asset counts, module bundling and API tiering move the real number by multiples. Get quotes for your actual point count, asset count and user count, and ask specifically whether API access sits inside the quoted tier or outside it. That line item has ended more automation projects than any technical constraint.

The moment an agent reads inbound tenant email, vendor invoices, scanned service reports, or portal submissions, it is processing content written by people outside your organization. That content reaches a model that can call tools. Text inside a document can be shaped to read as an instruction, and the model has no reliable way to tell a request from a command.

Prompt injection is not solved. There is no filter, no system prompt and no model version that closes it, and any vendor telling you otherwise is describing something that does not exist. What can be done — and what we build for — is blast-radius reduction: assuming a malicious or malformed input will eventually get through, and constraining what it can cause. Both the OWASP Top 10 for LLM Applications and the NIST AI Risk Management Framework are the right references to map controls against.

Untrusted content surfaces in facilities workflows and the architectural controls that limit their consequences. These reduce blast radius; they do not eliminate the risk.

The building automation system deserves its own line, because it is the one place where the consequences stop being financial. NIST SP 800-82 exists because operational technology has a different risk profile than IT, and building controls are operational technology. Setpoints, schedules, ventilation rates and equipment protection have occupant-safety and equipment-damage consequences.

Ten things that should slow a purchase down. Each of them is checkable in a single question, and each has shown up in evaluations we have run.

The Frenchy Digital red-flag list for facilities AI and fault-detection procurement, 2026.

If you only have time for one question, make it the baseline question. Ask how energy savings are measured, against what baseline, weather-normalised by which method, and verified by whom. A vendor with a real answer will have it ready. A vendor without one will move to case studies, and case studies are not measurement.

A vendor who will not put the baseline methodology, the API scope and the data-export path into the contract before you sign will not produce them after you sign either. Everything that matters in this category is checkable before money moves.

— Frenchy Digital buyer's principle

These are the bands Frenchy Digital uses to scope facilities AI engagements in 2026. They assume integration work is in scope from the start, because integration is the project — treating it as a phase-two concern is what turns a nine-week build into a nine-month one.

Frenchy Digital cost bands for facilities management AI engagements, 2026.

Senior-led delivery runs $150 to $225 per hour, and ongoing retainers run $2,500 to $9,500 per month, covering model and dependency upgrades, integration maintenance as your CMMS vendor ships changes, expansion of the evaluation set, incident response and a quarterly review. Every engagement carries a 30-day post-launch warranty, and you receive a written scope with a fixed-price phased proposal within 5 business days of the discovery call.

One budgeting observation. The integration layer is largely a fixed cost, paid once and reused. The first workflow carries the CMMS connection, the identity model, the audit pipeline and the review queue; the fourth workflow inherits all of it and costs a fraction of the first. Facilities teams that sequence their automation get considerably better economics than teams that run four disconnected vendor pilots in parallel and end up with four integrations nobody owns.

Everything above assumes the technology works as described. It mostly does. The failure modes in facilities AI are almost never the model — they are data, integration, and people, in that order. If you are writing the business case, write these into it.

• Your asset data is worse than you think: Ranking work by asset criticality requires an asset register that is accurate, complete and current. Most are none of the three. The first real cost of a facilities AI project is usually data remediation nobody budgeted for, and it is unglamorous work that cannot be skipped.
• 8% is a median, and yours may be lower: A recently commissioned, well-run portfolio has less to recover than the study's median building. Savings estimates are not portable between portfolios, and a vendor case study from someone else's estate tells you very little about yours.
• Alarm fatigue defeats good systems: The most common outcome of an FDD deployment is not wrong alarms; it is correct alarms nobody reads. Ranking and suppression are not nice-to-haves — without them the deployment produces no savings at all, however accurate it is.
• The write API decides the project: An agent that cannot create or update a work order in your CMMS is a recommendation engine that generates retyping. Confirm write access in writing before scoping, not during build.
• Technicians have to want it: An agent that reassigns work without explanation gets routed around within a fortnight. One that saves a technician a wasted trip gets adopted. Sequence the workflows so the first one visibly helps the people who use it.
• Prompt injection is unsolved: Any agent reading tenant email, vendor invoices or portal submissions is processing untrusted content. Controls reduce blast radius; nothing eliminates the risk. Design as though a bad input will eventually get through, because it will.
• Benchmarks you will want do not exist publicly: We could not obtain the Gordian higher-education facilities benchmark, so we cite no FCI or backlog figure for the non-federal sector. Pricing is only half-knowable: the mid-market CMMS vendors publish list rates, the enterprise platforms publish nothing. We have not filled either gap with estimates. Where you need a benchmark, you will usually have to generate it from your own portfolio.
• Occupancy data cannot answer space questions alone: Badge-based series are attendance proxies against an early-2020 baseline. Any space decision built on them without dwell time and desk-level sensing is built on a metric that was never measuring the thing being decided.

None of this argues against building. It argues for building the measurement alongside the agent — one workflow, a defensible baseline, an honest forecast at the published median rather than the marketed one. Facilities teams that instrument the before-state are the ones who can still justify the program in year three.

And the boundary holds throughout: these are administrative and work-management systems operating under human review. They rank, route, look up and draft. Engineering judgement, control changes, purchase approvals, life-safety calls and inspection sign-off stay with the qualified humans who are accountable for them.