AI Agents for Commercial CleaningOperations in 2026

Start with the demand signal, because it is the one thing in this business that has changed structurally since 2019 and the one thing most cleaning specifications have not caught up with.

Kastle's Back to Work Barometer reported a ten-city average of 55.6% for the week of July 24 to 30, 2026, down 0.9 points week over week, with a Tuesday peak of 65.2% and a lowest metro of 46.6% in San Jose. Its post-pandemic record week, the week of December 8, 2025, reached a 56.3% weekly average and a 66.0% Tuesday — and in that same week, Class A+ buildings ran 78.8% weekly and 95.5% on the Tuesday.

Read those two numbers together and you have the operating fact of the decade for a building service contractor: the work did not come back uniformly. A ten-city average near 55% and a Class A+ figure near 79% in the same week are not describing the same market. They are describing a portfolio that has split, where the trophy asset three blocks away is generating close to its old soil load and the mid-market tower you also service is generating a fraction of it.

The consequence for scope is precise. Cleaning tasks divide cleanly into two families. Traffic-driven work — restrooms, trash and recycling, break rooms, entry glass, high-touch surfaces, consumable restocking — scales with the number of humans who walked through. Time-driven work — floor care, vertical dusting, vents, window interiors, carpet extraction — scales with elapsed time and does not care whether anyone showed up. A spec written for 2019 prices both families against 2019 attendance. In a building running at half its old norm, you are performing traffic-driven work far below its priced frequency and time-driven work exactly as priced. In a Class A+ tower whose spec was trimmed during the vacancy years, you may be performing traffic-driven work well above what you are paid for.

Occupancy split and its consequences for scope, using Kastle Back to Work Barometer readings — Frenchy Digital analysis, August 2026.

The intra-week line in that table deserves its own emphasis. A 65.2% Tuesday against a 55.6% weekly average means midweek soil load is materially higher than Monday or Friday load, in the same building, at the same square footage. Almost every nightly crew is staffed flat across five nights. Variable-frequency scheduling by weekday is, in our experience, the most commonly unexploited margin lever in an office portfolio — and it is a scheduling problem, which is to say a software problem, not a labour problem.

A note on where else this data leads. Facilities directors on the client side are reading the same numbers and drawing conclusions about space, and the maintenance side of their budget is under its own pressure — GAO reported the governmentwide federal deferred maintenance and repair backlog rising from $171 billion in FY2017 to $370 billion in FY2024, and added federal building condition to its High-Risk List in 2025. That is a federal figure, not a commercial one, but the direction of travel is the same everywhere: buildings that are less full are also getting less capital, and the contractor is often the last vendor in the building who actually sees the condition of it. That observation is worth something commercially, and an inspection record is how you make it legible to a client.

Most articles about this industry quote a market size, a labour percentage and a turnover rate, none of which are sourced. What follows is assembled from BLS, the Census Bureau and an SEC filing — all free, all checkable — along with an explicit note wherever the number everyone repeats turns out to have no origin.

Commercial cleaning industry economics from primary federal and filed sources — assembled by Frenchy Digital, August 2026.

Four things in that table change how an operator should think, and they are worth taking one at a time.

The employment gap is the second thing. QCEW counts 1,069,345 people in the janitorial services industry. OEWS counts 2,209,760 janitors and cleaners across all industries. Both are correct and they measure different universes: one is an industry, the other an occupation. Roughly half the people doing this work are employed directly by schools, hospitals, governments and building owners rather than by a contractor — which means your competition for labour is not only other BSCs, and your total addressable market includes a very large pool of in-house operations that could be outsourced. QCEW also covers only UI-insured employment, so franchisees and 1099 workers are structurally invisible in it. The real contractor workforce is larger than 1,069,345.

Margins are genuinely thin, and there is exactly one audited public number. ABM Industries' FY2025 Form 10-K reports revenue of $8,745.9 million with a 12.3% gross margin, 3.6% operating margin and 1.9% net margin. ABM is diversified well beyond janitorial, so this is directional rather than a benchmark. But it is a filed, audited figure for the largest public company in the neighbourhood, and it tells you the shape of the business: at a 1.9% net margin, a bid that overruns by three percent of contract value does not reduce profit, it eliminates it and then some.

And the turnover statistic everyone quotes does not exist. BLS JOLTS publishes no janitorial series at all. The 200-percent-plus figure that appears in nearly every vendor deck in this industry is marketing, not data. The nearest published measure is Professional and Business Services at 55.3% total separations in 2025, against 39.6% for total nonfarm — high, and nowhere near the circulating claim. Turnover in this industry is real and it genuinely widens the gap between a theoretical production rate and a realized one. That is the operationally relevant fact, and it does not require an invented number to be true.

Structural cost analysis for a janitorial contract, anchored to QCEW, Census, OEWS and ABM's FY2025 filing.

A contractor who knows their own realized production rates to two decimal places does not need an industry benchmark. A contractor who does not will be misled by one — and in this industry most of the available benchmarks turn out to have no author.

— Frenchy Digital estimating principle

One occupation code contains several different labour markets, and bidding as though it contains one is a reliable way to lose money on institutional work. These figures come from the BLS Occupational Employment and Wage Statistics release for May 2025, SOC 37-2011, pulled from the BLS API.

Wage distribution and sector gradient for janitors and cleaners — BLS OEWS, May 2025.

The first thing to notice is the gap between the mean of $18.64 and the median of $17.71. Nearly a dollar separates them, and the direction tells you the distribution is pulled upward by employers you are not competing with on price. If your estimating template carries the national mean as a default wage, it is carrying a number that describes the occupation rather than your segment.

The second is the sector gradient. Contract cleaning — the sector you operate in — pays a mean of $17.44. K-12 schools pay $20.07. Local government pays $21.42. That spread is not a curiosity; it is a bidding constraint and a recruitment constraint at once. Bid a school district at your commercial wage assumption and you have underpriced the labour you will actually have to hire, because the district down the road is paying fifteen percent more for the same skills. And in any city where local government is hiring, that $21.42 is the wall your retention runs into.

One reconciliation worth understanding, because it will otherwise confuse anyone comparing sources. QCEW reports average annual pay in the industry at $34,450, while $18.64 an hour at full-time hours would imply roughly $38,800. The difference is not an error. QCEW's figure is an industry-wide average across all roles including part-time work, of which this industry has a great deal — many cleaning positions are four- or five-hour shifts. Read the hourly figures for bidding and the annual figure for understanding what the job actually pays a household.

The ranking below is not about which workflow is most technically interesting. It is a function of three things: how often the workflow runs, how expensive an error in it is, and whether the data needed to do it well already exists somewhere in your business. Bidding wins on the second factor by a wide margin, and the margin data above is why. Shift fill wins on the first.

The five agent workflows for a building service contractor, ranked by expected payback — Frenchy Digital, 2026.

One framing note that shapes everything below. Every one of these is administrative automation under human review. None of them cleans anything. None of them makes a disciplinary decision, a pay decision, a safety decision, or a decision that concedes liability to a client. The agent assembles, drafts, ranks, monitors and evidences; a person decides. That is not caution for its own sake — in a business where the operational record is also the evidence in a wage claim or a contract dispute, the division of labour between agent and human is a design requirement.

This is the highest-value agent in commercial cleaning, and the economics section above is the argument for it. Most operational mistakes in this business are recoverable within days. A cleaner does not show; you send someone tomorrow. A client complains; you fix it that night. A mis-bid is different. It is a single decision, made in an afternoon, from a walkthrough and a set of assumed labour minutes, that then binds your cost structure on that building for the entire contract term. At the margins this industry runs — 12.3% gross and 1.9% net in ABM's FY2025 filing, with labour at a verified floor of 45.1% of revenue — you cannot fix a mis-bid by working harder. You can only absorb it.

The mechanics of a proper bid are well understood and rarely executed consistently. You measure cleanable area by surface type. You map the specification to a list of tasks, each with a frequency. You apply a production rate — labour minutes per unit of area, per task — to get labour hours. You load those hours with burden, supervision, equipment and supplies. You add margin. The failure is almost never in the arithmetic. It is in the area schedule and in the production rates.

Here is the workflow an agent should run, and — more importantly — where it must stop.

Bid generation workflow with human approval gates — Frenchy Digital reference design for building service contractors.

The second thing a bidding agent does, and the reason its value compounds, is that it keeps the assumptions as structured data. In most contractors, an estimate lives in a spreadsheet that gets copied for the next bid and overwritten. Three months into a losing contract, nobody can reconstruct which line was wrong. Was it the restroom fixture count? The stripping frequency on the lobby? The assumption that one person could cover two floors? If the estimate is a set of records rather than a document, you can compare realized hours to bid hours at the task level and find the specific error. That is how a contractor stops repeating the same estimating mistake across a portfolio, and it is worth more over three years than the time saved writing the proposal.

A third, smaller function pays for itself surprisingly fast: scope-creep detection. Every contractor has a version of the “while you're here, could you also…” problem, where a facilities coordinator adds a task by email and it quietly becomes part of the nightly routine. An agent that reads inbound client requests against the signed scope lines and flags anything outside them — then drafts the change order — is a margin protection tool disguised as an inbox assistant. It works because the comparison is mechanical and the human effort of doing it manually is exactly why nobody does it.

It is 4:47pm. Three people have called off across two buildings, one of them a site with a 6pm start and a client who notices. The supervisor starts phoning down a list. She calls fourteen people, reaches six, and two say yes — one of whom has already worked thirty-eight hours this week and one of whom has never been in that building and is not on the access list.

This is the daily operational emergency at every building service contractor in the country, and it is a near-perfect fit for an agent, because the task is fundamentally a constrained ranking and outreach problem executed under time pressure with information the supervisor cannot hold in her head.

A shift-fill agent builds the ranked list, sends timed offers in waves rather than blasting everyone at once, tracks acceptances against the number of open slots, escalates when a wave lapses, and hands the supervisor a decision when the list runs out. The signals it ranks on, and the constraints that apply to each, look like this.

Shift-fill ranking signals with their operational and compliance constraints — Frenchy Digital reference design.

Two of those rows are the whole design. The first is the distinction between hard gates and soft weights. Site training and building access are gates: if the record is missing, the person is not offered the shift, full stop, unless a supervisor overrides with a logged reason. Everything else — proximity, hours, consecutive days, reliability — is a weight. Products that treat training as a weight will eventually send an untrained person into a building because the optimiser found them convenient, and the incident report will be very hard to explain.

The second is hours-to-date, which is where operations meets wage-and-hour law. Under the FLSA, overtime is owed after forty hours in a workweek, so the marginal cost of a fill decision changes sharply at a threshold most dispatchers cannot see in the moment. An agent that surfaces “this person is at 37.5 hours; filling this shift costs you 4.5 hours at time-and-a-half” converts an invisible cost into a visible one. It must not, however, decide to incur it. Overtime authorization is a management decision that gets logged with a named approver.

Quality assurance in commercial cleaning has a structural problem: the work is performed at night, in an empty building, and evaluated in the morning by someone who was not there. The result is that service quality is argued about rather than demonstrated, and the argument is conducted almost entirely out of memory.

That asymmetry is why inspection is the third-ranked workflow here despite being the least glamorous. At renewal, an incumbent contractor is competing against the client's recollection — and recollection retains the three complaints, not the two hundred and sixty uneventful nights. A structured evidence record inverts that. It also disciplines your own operation, because a scope line that is never inspected is, reliably, a scope line that is not being performed.

The thing that makes an evidence record valuable is not the photograph. It is the linkage. A photo with no reference to a contracted obligation is an opinion with an image attached. A photo tied to a specific scope line, in a specific area, at a verified time, with a corrective action and a re-inspection that closed it, is a record of contract performance.

Inspection evidence record structure — the fields that make QA data defensible at renewal.

Where the agent adds value is in the loop rather than the capture. Reading the inspection stream, it can cluster recurring deficiencies by area and by shift, draft the corrective action with the specific task and standard cited, assign it to the responsible supervisor, chase it, and confirm closure against the re-inspection photo. It can also generate the client-facing summary — which is the artifact most contractors produce late, inconsistently, or not at all, and which is precisely what a facilities director needs in order to defend the contract internally.

One caution that operators consistently underweight: an evidence record cuts both ways. A complete, timestamped, scope-linked history of every deficiency you ever found in a building is also a complete history available to opposing counsel in a dispute, and to a client building a termination file. Given that California's client-liability provision has already produced a seven-figure settlement reaching a customer and its cleaning contractors together, this is not an abstract concern. It is not an argument against building the record — the asymmetry still favours the contractor who can prove performance — but it is an argument for deliberate retention policies, a considered default on what is pushed to a client portal in real time, and a conversation with your own counsel before the first record is written.

Nightly crews at fixed sites are a solved scheduling problem. The parts that are not solved, and that quietly destroy margin, are periodic work and shared resources.

Periodic work — strip and wax, burnishing, carpet extraction, high dusting, window interiors, upholstery, pressure washing — is contracted at an annual or quarterly frequency and then forgotten until either the client notices it has not happened or someone realises in month eleven that four quarters of work need to fit into six weeks. It is contractually owed, labour-intensive, and scheduled by whoever remembers. An agent that maintains a live obligation calendar against contract terms, available crews, equipment, and building access windows is unglamorous and disproportionately valuable, because the failure it prevents is both a margin loss and a renewal risk at the same time.

Shared equipment is the other constraint. You have three auto-scrubbers, two extractors and one propane burnisher across eleven sites, plus the vehicles to move them and the trained operators to run them. Equipment location, availability and the operator qualification to use it are a genuine constrained-resource scheduling problem, and it is exactly the kind of problem that is tedious for a human and tractable for software.

Day porters and route workadd travel as a real cost. Sequencing supply runs, day-porter coverage across a cluster of nearby buildings, and periodic-work crews so that drive time does not consume the job is straightforward routing with domain constraints layered on: building access hours, tenant restrictions on noisy work, elevator availability, and the client's own rules about when contractors may be on site.

Sensor-driven frequency is the newest input and the least proven. Dispenser and door-counter sensors can, in principle, tell you which restrooms actually need servicing tonight — and the mechanism is sound, since a fixed-frequency round in a building at half its old attendance is obviously mismatched. Tork's own research, with 600-plus respondents across North America and Europe, reports that 71% of cleaners say eight out of ten dispenser checks are unnecessary. That is a vendor figure and a survey of perception rather than a measurement of hours, but it describes the mechanism accurately and matches what every supervisor will tell you. The savings claims attached to those sensors are a different matter, and the next two sections deal with them directly.

Client communication in this industry arrives through every channel at once: an email to the account manager, a text to a supervisor's personal phone, a ticket in the client's own work-order portal, a note left at the desk. The result is that response quality depends entirely on which channel a request happened to land in and whether that person was on shift.

An agent normalises the intake. It reads across channels, classifies the request — service deficiency, extra work request, scope question, billing query, access issue, safety matter — resolves it to a site and a contract, drafts an acknowledgement with a realistic time commitment, routes it to the right supervisor, and then closes the loop with the inspection evidence when the work is done. The last step is the one that changes the client relationship, because “we handled it” and “we handled it, here is the re-inspection photo taken at 9:42pm” are different messages.

• Draft, do not send, anything that concedes fault: An apology that accepts responsibility is a document in a dispute. The agent may draft it; a human sends it.
• Never authorise a credit or a price change: Financial concessions are an owner or account-manager decision, always, regardless of how small. This is also a favourite target for anyone attempting prompt injection through an inbound message.
• Never accept new scope: Route every extra-work request to the change-order path. This is where the scope-creep detection from the bidding workflow earns its keep.
• Escalate safety and security matters immediately to a human: Biohazard, injury, a security incident, a suspected break-in, a chemical exposure — the agent's only job is to route these fast and loudly, never to triage them.
• Match the language of the sender: Both client-side and worker-side communication in this industry is genuinely multilingual, and a translated response that a human reviews is better than a slow one in the wrong language.

Robotic floor care is the most-discussed technology in this industry. For the first time there is defensible data about how many contractors intend to buy it — and, more usefully, there is now a clear divergence between the manufacturers about whether the category has a future at all.

Start with intent. The best available adoption read is the 2026 Report on the Building Service Contractor Market, published by Contracting Profits (Trade Press Media / CleanLink) and sponsored by BSCAI— the contractors' association, not an equipment manufacturer. That sponsorship matters: it makes this the least conflicted adoption data in the category, because the party funding it sells memberships to contractors rather than machines to them.

2026 Report on the Building Service Contractor Market — Contracting Profits (Trade Press Media / CleanLink), sponsored by BSCAI. 12,952 invited, N=181, fielded March 9 – April 6, 2026, margin of error ±7.2% at 95%.

Read that table carefully, because two things in it point in opposite directions. Roughly 32% of respondents plan to adopt robotic floor equipment within twelve months, double the 2025 figure — a genuine and fast-moving shift in intent — and about 38% expect client demand for automation in contracts to grow, up nearly thirteen points year over year. But at the same time, 56.83% expect their adoption of new technologies such as IoT or robotics to stay exactly the same, and 69.91% report no plans at all for robotic carpet equipment. The modal contractor in this panel is not automating.

Now the part that almost never appears in coverage of this category. The manufacturers are not moving in the same direction, and one of the largest has just left.

Autonomous cleaning equipment manufacturers, from company filings and investor communications — Frenchy Digital review, August 2026.

Hold those two rows side by side. Tennant is guiding to $130–145 million of robotics revenue in FY2026 against a $250 million target for 2028, and announced its 10,000th autonomous machine sold in June 2025. Nilfisk wrote its autonomous joint venture, Thoro LLC, down to zero, stating that the product portfolio “will not be part of Nilfisk's future business,” in a year that produced a €36.7 million loss and a December 2025 takeover offer.

Two serious manufacturers looked at the same category and reached opposite conclusions. That is not proof the technology fails — Tennant's revenue is real and growing, and a write-down can reflect execution rather than the market. But it is a material fact that no vendor presentation will show you, and it should be sitting in your head next to the adoption survey. An industry where a major participant has just exited is not an industry where the returns are settled.

One arithmetic caution, because this figure is widely mis-stated. Brain Corp's often-quoted “68% growth” describes new deployments within a period, not installed base. Its fleet went from 37,000 to more than 50,000 machines — roughly 35%. Its cumulative and half-year figures do not reconcile with each other, so do not stack them, and do not let a vendor present deployment growth as fleet growth.

Independent of all that, the structural role of the machine is worth stating clearly, because the structure is enough to make a buying decision even in the absence of ROI data.

• They cover open floor area, not detail work: An autonomous scrubber cleans large, contiguous, unobstructed hard floor. In a typical office building, open floor is a minority of the nightly labour. Restrooms, trash and recycling, break rooms, high-touch surfaces, glass, vertical surfaces, detail edges and corners, and everything requiring a decision are outside the machine's envelope entirely.
• They change the labour mix rather than eliminating it: The realistic deployment pattern is that the machine runs a floor while a person performs detail work in the same window. That is a different allocation of a shift, not a removed shift. If your business case assumes headcount reduction, examine it hard before signing.
• They generate their own labour: Charging and dock management, water filling and dumping, pad and squeegee changes, brush maintenance, route mapping and re-mapping when a floor layout changes, and exception handling when the machine stops in front of a fire door or gets confused by a pallet left in a corridor. Someone does all of that, on your payroll.
• They depend on site conditions you may not control: Contiguous open floor, low obstruction density, floors cleared before the shift, elevator access for multi-floor deployment, network coverage for telemetry, and secure storage. A building that fails two of those is a poor candidate regardless of the machine's specifications.
• They are a poor fix for a scheduling problem: If your realized hours exceed your bid hours because periodic work is unplanned or because call-offs go unfilled, a machine will not address either. Diagnose the cause of the overrun before buying capital equipment to treat a symptom.

Where machines plausibly do pay is narrow and identifiable: large contiguous hard-floor areas, cleaned nightly, on a single floor or with reliable elevator access, at a site where you already know your baseline hours and can therefore prove the delta. That is a real profile and some contractors have it. It is not most buildings, and it is not a portfolio-wide strategy.

This is the finding that should change how you buy, and it is worth stating without hedging.

That absence is unusual enough to be worth dwelling on. Building technologies of comparable capital cost routinely get measured. Automated fault detection and diagnostics for HVAC, for instance, has a peer-reviewed field result: a study across 26 organizations, 550 buildings and roughly 97 million square feet found median whole-building energy savings of 8%, and DOE and LBNL separately report average savings around 9% with two-year paybacks — against circulating vendor claims several times higher. That is a different technology in a different part of the building. It is included here for calibration: it shows both that independent measurement in buildings is entirely possible, and how far vendor numbers typically sit from measured ones once someone finally checks.

Below is the provenance trace on the claims a contractor is most likely to encounter — including the two industry statistics that appear in nearly every deck. Every row was checked to its origin.

Provenance trace on the efficiency and industry claims circulating in commercial cleaning — Frenchy Digital, August 2026.

The Quinnipiac row deserves a moment on its own, because it is the specific trap in this category. A widely-shared “case study” of scrubber performance at Quinnipiac University is published by Tennant, the equipment manufacturer. The university is the deployment site, not the researcher. There is no university study. A campus setting confers no academic status on a document, and this pattern — real institution, vendor authorship, academic-sounding headline — is how marketing acquires the appearance of evidence. Check who published the PDF, not whose logo is at the top of it.

The restroom-sensor claims follow the same shape. Tork Vision Cleaning's widely-quoted “up to 20% cleaning time saved” rests, per Essity's own footnote, on the documented results of three customers measured before and after. A companion “24% fewer cleaning rounds” comes from two customers over 158 days. A “+30% visitor satisfaction” figure was measured across two sensor-equipped and two traditional washrooms at a trade-show exhibition in May 2016 — four washrooms, at a trade show, a decade ago. Notably, Tork's current product page has dropped the 20% headline entirely, replacing it with a survey of 69 self-selected existing users across 18 countries. That is a vendor quietly improving its own disclosure, and it deserves credit; it is also a signal about what the original number was worth. Four separate academic database sweeps found zero peer-reviewed field studies measuring labour hours or cost savings from sensor-driven cleaning in real buildings — the closest published work is prototype and algorithmic, not deployed ROI.

When the entire published evidence base for a category comes from the companies selling into it, the correct response is not scepticism about the technology. It is insistence on your own measurement — and a refusal to pay for a payback period you cannot reproduce.

— Frenchy Digital buyer's principle

If you are evaluating whether to buy AI capability from your existing platform or build it, it helps to know what the category has actually shipped rather than what it has announced. As of August 2026, the honest picture is thinner than the marketing conversation suggests.

Shipped AI capability across janitorial operations platforms — Frenchy Digital vendor review, August 2026. Status changes quickly; verify before you buy.

Two observations for a buyer. First, Otuvy's choice of first feature is instructive. Of everything a cleaning platform could have built with a language model, they shipped real-time translation — which is the correct read of where the friction actually is in a multilingual, distributed, nocturnal workforce. That is a product decision made by people who have spoken to supervisors.

Second, the gap between conference rhetoric and shipped features is wide. Aspire, now owned by ServiceTitan, has been vocal about AI investment while showing no named feature and no ship date; its March 2025 release covered submittals, item substitutions, rebill invoices and bulk scheduling — all useful, none of it AI. That is not a criticism of the product. It is a reason to ask a specific question during evaluation: what shipped, on what date, and can I see it running in a customer account today?

Every article in this cluster arrives at the same conclusion from a different industry, and commercial cleaning is no exception: the hard problem is not whether a model can do the reasoning. It is whether your agent can read and write the systems where the work actually lives. In this industry those systems are a mix of timekeeping telephony, workforce management, accounting, sensor and robotics clouds, and — critically — the client's systems, which you do not own and cannot change.

Integration surfaces for a building service contractor and the realistic write path for each — Frenchy Digital, 2026.

The generalisable pattern is this: read broadly, write narrowly, and own your own state. The agent maintains its own record for the workflow it runs — the bid with its assumptions, the fill attempt with its offers and responses, the inspection with its evidence chain. It reads from surrounding systems wherever an export or an integration path exists. It writes back only through paths that are either fully controlled by you or gated behind a human confirmation. And it never writes to a timekeeping system at all, for reasons covered in the next section.

Client-side systems deserve particular discipline in the sales process. It is common for a contractor to promise, during a tender, that their system will integrate with the client's work-order platform. Do not make that commitment before you have seen the API documentation and been issued a credential. In a meaningful share of accounts the honest answer is that the integration is an email address and a person, and it is far better to say so in the tender than to discover it in implementation.

An agent that touches scheduling and pay sits directly adjacent to the highest-frequency legal exposure in this industry, and the enforcement is not hypothetical. DOL Wage and Hour brought 299 janitorial-specific compliance actions in FY2025, recovering $4,715,195 in back wages for 2,231 employees — roughly $2,100 per worker and about $15,800 per action. That does not make an agent a bad idea. It makes the design constraints non-negotiable.

Compliance surfaces for scheduling and dispatch agents in commercial cleaning. Not legal advice — the items flagged as unverified require counsel.

The wage-and-hour rules themselves are stable and worth reading directly rather than through a summary. DOL Fact Sheet #22 covers what counts as hours worked, Fact Sheet #21 covers recordkeeping obligations, and Fact Sheet #13 covers the employment-relationship analysis. The practical translation for software is short: the agent may schedule, but the timekeeping system stays authoritative, human-approved, and beyond the agent's write path. An agent that can adjust a punch is a liability with a user interface, no matter how convenient the feature looks in a demo.

The biometric row in the table is new for 2026 and easy to miss. TEAM Software added biometric check-in verification in June 2026 and does not market it as AI, which means it can arrive in your stack through a routine release note rather than through a procurement decision. Several states regulate biometric identifiers, some with private rights of action, and we did not verify the current statutory landscape for this article. Do not enable a biometric feature because it appeared in a changelog. Consent, retention and deletion obligations are a legal review, not a settings toggle.

Safety training is the one compliance area where an agent is unambiguously a net positive, because the control is simple and mechanical. OSHA's Hazard Communication Standard requires labelling, safety data sheet availability, and worker training on the chemicals in use, and the bloodborne pathogens standard applies where exposure is reasonably anticipated. A dispatch agent that will not assign a task to a worker whose training record for that task is missing or expired is a real, buildable, testable control — and it is the same gating mechanism used for site training in the shift-fill workflow, applied to a different record. If you are choosing one compliance feature to build first, build this one.

One adjacent note for accounts that care: EPA's Safer Choiceprogram certifies products meeting its safer-chemistry criteria, and green-cleaning requirements appear routinely in institutional and public-sector tenders. An agent that maintains your product list against a tender's certification requirements, and flags a substitution that would break compliance, is a small feature with disproportionate value on the specific accounts where it matters.

This section exists because an agent that dispatches work is, whatever else it is, a machine for recording direction and control. In an industry built substantially on franchising and subcontracting, that record is a legal artifact before it is an operational one.

Worker classification, franchising and client-liability landscape for commercial cleaning — Frenchy Digital, August 2026. Not legal advice.

Three points deserve emphasis, because published commentary on all three is frequently out of date.

The Jan-Pro misclassification litigation is over. It resolved with a $30 million non-reversionary fund, final approval in May 2024, zero opt-outs and zero objections, and the case terminated on January 6, 2026. Awuah v. Coverall concluded in 2013. Any article describing either as ongoing is stale — and a zero-objection settlement of that size is a stronger signal about the theory's strength than a pending case would be. New filings exist, including a Jan-Pro matter in Michigan in 2025, a Coverall matter in Texas in 2026, and five Jani-King suits, but their merits are unverified here. Read the dockets on CourtListener rather than relying on secondary summaries.

And your client can be pulled in with you. California Labor Code §2810.3 makes a client jointly liable for a labor contractor's wage violations, and in January 2024 a $1,000,000 settlement covering 589 janitorswas announced against Cheesecake Factory and its cleaning contractors under that provision. Read commercially rather than defensively, this changes the sales conversation: your client's procurement team now has a direct financial interest in your wage-and-hour records being clean and exportable. A contractor who can produce them on request is selling something a competitor cannot.

On the federal picture, two things are in motion and neither has landed. DOL's January 2024 independent-contractor rule is still operative; a rescission NPRM published in February 2026 with comments closing in April, but no final rule has issued — so do not plan as though the rule is gone. And a joint-employer NPRM published in April 2026, which is directly material to tiered janitorial subcontracting. Track both; do not build assumptions about either into a system you will have to unwind.

Have counsel review the agent's control surface before you automate it, not after the record exists. Dispatch, sequencing, completion monitoring and quality scoring are all forms of direction — and unlike a supervisor's phone call, an agent timestamps every one of them.

— Frenchy Digital design principle

The moment an agent reads inbound client email, tender documents, portal submissions, or scanned specifications, it is processing content written by someone outside your organisation who may not have your interests in mind. Prompt injection — text crafted to be interpreted as an instruction rather than as data — is not a solved problem, and no vendor claim to the contrary should be believed. What you can do is reduce blast radius, which is a design activity, not a product feature.

The concrete scenarios in this industry are mundane, which is what makes them plausible. A tender PDF containing embedded text that instructs an estimating agent to apply a particular production rate or to omit a scope line. A complaint email that includes a passage directing a customer-service agent to authorise a credit. A supplier quote engineered to alter a purchasing decision. None of these require a sophisticated attacker; they require an agent with tools and no gates.

• Retrieved content is data, never instruction: Content extracted from a document or an email enters the context clearly delimited as untrusted material to be analysed. It never occupies the same trust level as your system instructions or a verified human request.
• Tools are allowlisted per workflow: The estimating agent has no tool that can send a customer email. The customer-service agent has no tool that can change a price. Capability separation is the strongest control available and it costs nothing to design in from the start.
• Nothing consequential executes without human confirmation: Sending a customer-facing message, changing a price, altering a committed schedule, authorising overtime, issuing a credit, or releasing a purchase order — all of these require a human action, every time, with the agent's proposal and its reasoning shown.
• Retrieval is scoped to the contract already in context: An agent working a ticket for one building has no path to another client's contract, pricing, or evidence records. Scope at query time, not by filtering results afterwards.
• Every tool call is logged with its arguments: You need to be able to reconstruct what the agent did and on what input, months later. A log that records only the final output cannot answer the question that matters after an incident — or in a wage-and-hour inquiry.
• Injection cases run in CI on every prompt change: Maintain a suite of adversarial documents and messages drawn from real inbound mail, and run them on every change to a prompt, a tool definition, or a model version. Treat a regression the same way you would treat a failing unit test.

The OWASP Top 10 for LLM Applications is the right checklist to map a design against, and the NIST AI Risk Management Frameworkis the right vocabulary if an enterprise client's security team asks how you govern the system. Neither is long, and being able to answer those questions in a tender is increasingly a commercial advantage in institutional accounts.

The sequencing below reflects how these engagements actually go rather than how a project plan would prefer them to. The first phase is not automation. It is finding out whether your data can support automation, and in most contractors the honest answer at the start is no.

1. 1.Weeks 1–4: instrument before you automate: Assemble realized labour hours by site and task, area schedules for representative sites, bid-versus-actual for the last twenty jobs, and a roster with training state, access authorization and language per employee. Expect gaps. Closing them is the deliverable of this phase, and skipping it is why most tools in this category fail to produce a measurable result.
2. 2.Weeks 5–9: one workflow, one segment: Build a single agent — usually bid generation, sometimes shift fill if call-offs are the acute pain — and run it on a defined slice of the business. Not the largest account. A segment where a failure is recoverable within a night and where you have enough volume to see a pattern within a month.
3. 3.Weeks 10–13: measure against the baseline you captured: Compare bid variance, fill time, unplanned overtime, or complaint closure time against the numbers from phase one. If you cannot state the improvement in the units your business is run in, the pilot did not succeed regardless of how good the demo looked.
4. 4.After 90 days: add the second workflow, reusing the substrate: The first agent carries the data model, the identity and permission layer, the logging, the human-approval pattern and the messaging infrastructure. The second inherits all of it and costs a fraction. Contractors who sequence get materially better economics than contractors who pilot three vendors in parallel.

The pattern to avoid, which we see often, is the parallel pilot: three vendors, three data exports, three logins for supervisors who already carry two phones, and no baseline against which to evaluate any of them. It feels like moving fast. It generally produces a year of activity and no decision.

If a robot or a sensor deployment is anywhere on your roadmap, phase one is doing double duty — the baseline hours you capture for the estimating agent are exactly the baseline you will need to evaluate the hardware. Given that no independent measurement of that hardware exists anywhere in the public record, and no manufacturer publishes a price to divide into it, your own baseline is not a nice-to-have. It is the only evidence you will ever have.

Every item below is a filter you can apply during a demo, before a pilot, at no cost. Most of them are a single question.

Vendor red flags for AI and automation products sold to building service contractors — Frenchy Digital, 2026.

Ask for one artifact rather than a document set: an anonymised export of your own historical hours, produced from the system, in a format you can open. A vendor who cannot produce it during evaluation will not produce it after termination either.

— Frenchy Digital buyer's principle

These are the bands Frenchy Digital uses to scope operations-AI engagements in 2026. They assume the data work described in phase one is in scope rather than assumed away, because assuming it away is what makes these projects fail.

Frenchy Digital cost bands for commercial cleaning and building services 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, evaluation expansion, incident response, and a quarterly technical 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.

On sequencing: build the estimating agent first if you have the historical data to support it, because the payback is the largest and it funds everything after it. Build shift fill first if your unplanned overtime is visibly out of control, because that number moves fastest. Build inspection first only if you have a renewal at risk in the next two quarters and need the evidence record to exist before the conversation.

The economics in this article are sourced. Much of what surrounds them is not, and it is worth being precise about which is which.

• The labour-share figure is a floor, not a measurement of labour cost: 45.1% is QCEW W-2 payroll against Census receipts for 2022. It excludes payroll taxes, benefits, workers' compensation, and all franchise and 1099 labour — every one of which is a real labour cost. Your fully loaded number is higher, and the gap between 45.1% and your own figure is your burden plus your inefficiency. The circulating 50–70% range remains unsourced; we are not endorsing it, only noting that the truth sits above the floor.
• The revenue denominator is three years old: Census reports $72.795 billion for 2022 and has published no 2023 or 2024 figure. Any more recent revenue number you encounter is an estimate. The derived labour share carries that vintage with it.
• QCEW structurally undercounts this industry: It covers only UI-insured employment, so franchisee-operated units and 1099 arrangements — both common here — are invisible in the establishment count, the employment count and the wage total. The real workforce is larger than 1,069,345 and the real payroll is larger than $36.84 billion.
• ABM is not a janitorial benchmark: Its FY2025 margins of 12.3% gross, 3.6% operating and 1.9% net are audited and real, but ABM is diversified well beyond cleaning. Use it as directional evidence that this is a thin-margin sector, not as the number to compare your own janitorial P&L against.
• There is no janitorial turnover statistic, at all: BLS JOLTS publishes no series for it. The nearest is Professional and Business Services at 55.3% total separations in 2025 against 39.6% total nonfarm. The 200%-plus figure repeated across this industry has no source, and we could not construct one.
• There is no independent measurement of robot or sensor ROI: Sweeps of OpenAlex, arXiv and the GSA Green Proving Ground found no peer-reviewed, academic or government studies of autonomous scrubber productivity, labour displacement or payback, and four separate academic database sweeps found no peer-reviewed field studies of labour or cost savings from sensor-driven cleaning in real buildings. No manufacturer publishes a price or a RaaS rate. Every percentage in circulation is vendor-supplied, with disclosed sample sizes of two to three sites where they are disclosed at all.
• The adoption data is self-selected and measures intent only: The 2026 BSC market report drew 181 responses from 12,952 invitations — roughly a 1.4% response rate, ±7.2% at 95%, with heavy self-selection toward contractors already interested in technology. It measures what contractors plan to buy, not what any of it returned.
• Kastle is an attendance proxy, not a utilization measure: It covers only buildings where Kastle is the access-control vendor across ten metros — a convenience sample skewed toward larger multi-tenant towers — and counts first badge-in per day rather than dwell time. The denominator is the early-2020 norm, not physical capacity. Use it directionally; use your own site data for site decisions.
• Several legal questions here were deliberately left open: Current state janitorial statutes outside California, the biometric-privacy landscape, and the merits of the pending Jan-Pro, Coverall and Jani-King filings were not verified for this article. The federal independent-contractor and joint-employer rules are both mid-rulemaking. None of this is legal advice and all of it should go to counsel.
• An estimating agent is only as good as your realized-hours data: This is the most common reason these projects underdeliver. Contractors routinely believe their hours data is cleaner than it is — hours recorded against a site but not a task, sites with no area schedule, jobs where the crew size changed without a record. If phase one finds that gap, close it before building, not after.
• A fill-rate optimiser will happily create overtime: Any objective function that maximises coverage without a cost constraint and a human approval gate will spend money you did not authorise. This is not a hypothetical failure mode; it is the default behaviour of an unconstrained optimiser.
• Adoption fails at the phone, not at the model: This workforce is distributed, nocturnal, frequently solo, multilingual, and reachable on personal devices. Any interface assuming a desk, a corporate login, an app install, or English fluency will show a promising pilot and a collapsed month three. Build SMS and voice first, in the languages your crews actually speak.
• The vendor landscape moves faster than any article: The shipped-AI review here reflects August 2026 and will age. Verify current capability directly, and apply the same question to every claim: what shipped, on what date, and can I see it running in a customer account?
• None of this fixes underpricing: If you are winning work by bidding below cost to hold volume, better estimating will tell you so precisely and change nothing else. That is a commercial decision about which accounts to keep, and no agent will make it for you.

A closing note on the trade literature, because it is where most contractors will look next. ISSAis the industry's principal association and its 612 Cleaning Times is the standard estimating reference — but ISSA states plainly that the times were submitted from thousands of different sources and are not a time-motion study, and the association's membership includes the equipment manufacturers whose claims those times are sometimes used to validate. It is a bidding convention behind a paywall, and it is genuinely useful as one. It is not measurement.

Which brings the article back to where it started. The industry-level economics are now knowable: QCEW, Census, OEWS, JOLTS and an SEC filing are free and they answer most of the questions a contractor actually has about the shape of this business. What remains unmeasured is everything the technology vendors are selling. In that gap, the contractor who instruments their own operation is not merely better informed than their competitors — they hold the only reliable data that exists about the thing they are being asked to buy. That is a genuinely defensible position, and it costs a few weeks of disciplined record-keeping to acquire.