Top 10 AI Agents forRestaurants in 2026

Every ranking of restaurant AI you will read this year is built on numbers that nobody outside the vendor has ever checked. Open five of them and you will find the same shape: a product claims 94% accuracy, or 96% order completion, or 86% autonomous handling, and the listicle prints it as though it were a measurement. It is not a measurement. It is a marketing assertion about the seller, published by the seller, with no methodology, no denominator, and no definition of the word being measured.

That would be a soft criticism in most industries. In this one it is not soft at all, because the claim has already been tested by a regulator with subpoena power, and it failed. In January 2025 the Securities and Exchange Commission entered an order finding that a publicly traded drive-thru voice AI company had made materially false and misleading statements about exactly this kind of number — while, on the SEC's findings, human order takers abroad processed the vast majority of orders placed through the product. That is not a hypothetical about vendor puffery. It is a federal administrative proceeding, with paragraph numbers, about a restaurant technology product.

So the claim under test in this article is narrow and specific: can an operator rely on any published performance figure in this category? Our answer, after checking every vendor site in the roster on 23 August 2026, is no — and the useful consequence is that you should stop trying. Rank on what you can verify from a browser and a contract, put the performance question into your RFP as a written commitment rather than a number you read on a homepage, and measure the thing yourself at your own sites, in your own lanes, with your own guests.

This piece sits inside a wider set of vertical rankings built on the same method; the cross-industry version, with the same refusals applied to general-purpose tools, is our guide to the top AI agents in 2026. What follows is the restaurant edition: what we scored, what we would not score, ten vendors in order, the constraint that decides whether any of it works, and the parts we could not verify.

We scored eight attributes, every one of which you can re-check from a browser in under an hour. That constraint is the whole method. If a buyer cannot independently confirm a cell in our table, the cell does not belong in a ranking — it belongs in a sales deck.

We refused to score accuracy, order completion, automation rate, deflection, containment, labour hours saved, revenue lift and ROI.Not because they do not matter — they are the only things that matter operationally — but because every published figure in this market is self-reported by the seller, and there is no neutral party checking any of them. Printing a vendor's accuracy claim beside a competitor's in a comparison table implies a common measurement standard. There is no common standard. There is not even a common definition: one vendor's “automation rate” may count an order as automated when restaurant staff did not intervene, while a remote agent did.

That last distinction is not a hypothetical we invented for rhetorical effect. It is drawn word for word from a federal enforcement order, and it is the subject of the next section.

On 14 January 2025 the SEC entered a settled order against Presto Automation Inc., in what is widely described as its first AI-washing enforcement action against a public company. The citation is In the Matter of Presto Automation Inc., Securities Act Release No. 11352, Exchange Act Release No. 102177, Administrative Proceeding File No. 3-22413. The findings concern statements about the company's flagship drive-thru voice product made in Commission filings and public statements from November 2021 through May 2023.

The marketing side of the story is familiar to anyone who has sat through a vendor demo. A January 2022 press release claimed the product delivered “over 94% accuracy even in noisy environments”. Investor materials claimed 95%–99% “automated order completion.” Those are precisely the numbers a competing listicle would have reprinted in a comparison table.

What the order found is the part that should change how you buy.

Presto Voice units powered by Presto's proprietary AI speech recognition technology lacked the capability to take orders on their own and required substantial human involvement.

— SEC, In the Matter of Presto Automation Inc., Securities Act Release No. 11352 (14 January 2025), ¶3

Presto hired, trained, and supervised human order takers located abroad (primarily in the Philippines and India), who processed the vast majority of drive-thru orders placed through Presto Voice.

— SEC, In the Matter of Presto Automation Inc., Securities Act Release No. 11352 (14 January 2025), ¶3

The order also records how the company's own metric was constructed. The completion rates it cited referred, in the order's words, to rates at which drive-thru orders were completed without restaurant staffinvolvement — but not without any human involvement at all. That is the definitional trapdoor. A number can be literally accurate and still describe the opposite of what a buyer hears.

And the internal record quoted by the SEC shows the company understood the gap. An internal message quoted in the order, dated January 2023, describes the practice of telling investors the AI was running 95%+ accuracy without disclosing that the AI was doing none of the work and all orders were processed by humans. An October 2022 message quoted in the same paragraph objects to publishing an automation rate to customers because it infers no supervision, which was not true. Messages from January 2022 quoted at paragraph 27 put the mechanism plainly: with humans in the loop, accuracy is not a major concern, and the figure can even reach 95% or more with humans. We attribute these to the SEC's order rather than to any named individual, because the order does not name them.

By the time the company disclosed the true picture, the numbers had inverted. Its 17 November 2023 prospectus supplement stated that over 70% of orders taken by the product required human agent intervention. A 14 December 2023 Form 8-K disclosed human intervention on 100% of orders at the substantial majority of locations running the original version. Those are the same product, the same period, and a completely different story from the press releases.

One more precision point, because it is the error most likely to be repeated. Presto Automation Inc. is not Presto Phoenix Inc.The original company was delisted from Nasdaq in September 2024, defaulted on its loan, and its lender foreclosed; the assets were sold on 3 December 2024 to a venture investor group led by Remus Capital with $18 million of new capital, and the product now trades under a new entity whose site reads © 2026 Presto Phoenix Inc. The SEC order concerns the old public company. Attaching it to the current entity as though they were one continuous business would be exactly the kind of sloppiness this article exists to refuse — which is also why we do not rank Presto Phoenix in the ten. The entity that made the claims no longer exists, and the current entity publishes no more verifiable information than anyone else in the roster.

Use the order the way a regulator would want it used: as the one documented instance of what happens when someone actually audits a published AI performance metric in this industry. It found the metric did not mean what buyers thought it meant. That is why we score only what you can check.

No independent benchmark of these products exists — but restaurants are unusually lucky, because one genuine third-party measurement of the modality does. Intouch Insight published its 25th Annual Drive-Thru Study with QSR Magazine on 1 October 2025. The methodology is public: fieldwork in June and July 2025, across the United States at varying days and times, 165 mystery shops per brand across 13 quick-service brands, plus 120 AI-enabled orders evaluated across three chains testing voice-AI ordering.

Errors on AI-enabled orders were, in the study's own words, most frequently attributed to order customization. Hold onto that sentence — it reappears in the allergen section, because allergen substitutions are order customisations, and it reappears in the human-in-the-loop table, because it tells you exactly where to put the escalation trigger.

Now put the two independent findings side by side, because their convergence is the strongest evidence in this entire category. A federal regulator, looking at internal documents, found a vendor concealing that humans were doing the work. A mystery-shop study, looking only at outcomes in the lane, found that the good accuracy number appears when a human steps in. Neither knew about the other. They point at the same variable: the undisclosed human. And not one vendor in this roster publishes a human-in-the-loop rate.

The speed finding deserves a fair hearing too, because the honest read is not one-sided. Roughly 21 seconds per car is a real operational gain, and satisfaction scored higher on AI-enabled orders than the study average. If your binding constraint is throughput at peak and your order mix is simple, that is a legitimate reason to run a pilot. If your order mix is heavy on modifiers — build-your-own formats, allergy-sensitive menus, complex customisation — the measured error concentrates precisely in your traffic, and the same study is a reason for caution. Read it as a map of where the technology is strong and weak, not as a verdict.

Every cell below is either a citation you can open or the literal words “not publicly disclosed.”There is no accuracy column, no ROI column and no rating out of five, because none of those could be filled honestly. All entries were checked on 23 August 2026 against the vendor's own pages and the public corporate record.

Three patterns jump out of that table and they are more useful than any ordering we could impose on it. First, published compliance documentation is nearly absent from this category. Exactly one vendor publishes a real trust centre with named report types. Two display badges without a reachable report. Seven publish nothing we could verify. For an industry that processes card payments and, increasingly, voice recordings, that is a finding in itself.

Second, published pricing is rarer still. Two vendors publish amounts. In a market selling to operators who budget to the penny per cover, that asymmetry is a choice, and it is a choice worth pricing into your evaluation time: every gated quote is a sales cycle you are paying for in calendar days.

Third, and most importantly, not one vendor discloses a human-in-the-loop rate.Nobody publishes what share of orders is completed with no human — not restaurant staff, not a remote agent — touching them. After the SEC order and the Intouch data, that is the number a serious buyer should demand in writing. It is also the easiest thing in the world to ask for and the hardest thing to fake in a contract.

The ordering below reflects verifiability, not performance.A vendor that publishes its price, names its integrations and documents its compliance ranks above one that publishes nothing — not because its software is better, but because you can make an informed decision about it and hold it to what it said. That is the only ranking anyone can defend right now.

The binding constraint for restaurant AI is not accuracy. It is that a voice agent creates a recording, and in several states a recording of a voice is a regulated biometric identifier or a two-party-consent problem. The drive-thru speaker box is plausibly the highest-volume consent surface in American retail, and almost nobody treats it as one.

Start with Illinois, because the statute is unusually literal. Under 740 ILCS 14/10, a “biometric identifier” means “a retina or iris scan, fingerprint, voiceprint, or scan of hand or face geometry.” Voiceprint is right there in the enumerated list. Section 15(b) requires three things beforecollection: written notice that a biometric identifier is being collected or stored, written notice of the specific purpose and length of term, and a written release executed by the subject. Section 15(a) requires a published retention schedule and destruction guidelines — destruction when the initial purpose is satisfied or within three years of the individual's last interaction, whichever comes first. Section 20 provides liquidated damages of $1,000 or actual damages for negligent violations, $5,000 or actual damages for intentional or reckless ones, plus attorneys' fees, expert witness fees and litigation expenses.

One status point people get backwards. The 2024 amendment — Public Act 103-0769, effective 2 August 2024 — added that repeated collection of the same biometric identifier from the same person by the same method is a single violation for which the person is entitled to “at most, one recovery,” with a parallel limit for repeated disclosures. That legislatively overrides the per-scan accrual theory. It caps the number of recoveries. It does not reduce the $1,000 and $5,000 figures, which stand exactly as written. Secondary reporting suggests the Seventh Circuit held in April 2026 that the damages amendment is remedial rather than substantive and therefore applies retroactively to pending cases; we did not read that opinion ourselves and flag it as secondary.

There is a newer development operators should track without overstating. Nine coordinated BIPA voiceprint class actions were filed in the Northern District of Illinois in May 2026 against a group of large technology companies, alleging that voiceprints were collected for AI voice-model training without BIPA disclosures, consent or retention safeguards. That account comes from the plaintiffs' own law firm and consists of allegations, not findings, and we attribute it as such. Its relevance here is structural rather than factual: it establishes that the training-data useof captured voice is now itself a theory of liability — and that is exactly what a restaurant voice vendor's standard contract usually permits.

Now California, where the intuition of most operators is precisely inverted. Penal Code § 632 bars intentionally recording a “confidential communication” without the consent of all parties — but the statute's own definition excludes a communication made “in any other circumstance in which the parties to the communication may reasonably expect that the communication may be overheard or recorded.” An outdoor drive-thru speaker, with cars queued behind and a menu board in the open air, is a strong candidate for that exclusion. An inbound reservation call to a restaurant is not. So the phone line carries more CIPA risk than the lane, which is the opposite of how most operators rank the two. We present that as analysis, not settled law, and you should have counsel apply it to your states.

Section 631 is the provision that pulls the vendor into your problem. It reaches anyone who reads or attempts to read or learn the contents of a message in transit without authorisation, and equally anyone who “aids, agrees with, employs, or conspires with” another to do so, with penalties up to $2,500 for a first offence and up to $10,000 with a prior conviction under the related sections. In Ambriz v. Google, LLC, No. 3:23-cv-05437-RFL (N.D. Cal., 10 February 2025), the court denied a motion to dismiss a § 631(a) claim on a theory commentators have called a capability test: a cloud contact-centre AI provider could qualify as a third party rather than a mere extension of the business because it had the capability to use call data for its own purposes, including model training and product improvement, regardless of whether it actually did so. A denied motion to dismiss is not a holding of liability, and the theory faces standing pressure; present both sides in your own risk memo.

Finally, outbound calling. The FCC's Declaratory Ruling FCC 24-17 in CG Docket No. 23-362, adopted 2 February 2024 and released 8 February 2024, confirms that the TCPA's restrictions on an “artificial or prerecorded voice” encompass current AI technologies that generate human voices, so such calls require the prior express consent of the called party absent an emergency purpose or exemption — and prior express written consent for marketing. The ruling expressly declines to carve out technologies that purport to provide the equivalent of a live agent. Separately, 47 U.S.C. § 227(d)(3) requires artificial or prerecorded messages to identify the business at the start and provide a telephone number or address.

Do not let anyone tell you the TCPA got easier. Insurance Marketing Coalition v. FCC(11th Cir., 24 January 2025) vacated only Part III.D of a 2023 FCC order — the one-to-one consent provision — and the court's own footnote states that the 2012 Order is not at issue. One narrow consent provision was vacated. The written-consent regime and the artificial-voice classification both stand, and an AI voice agent placing outbound calls sits squarely inside them. Inverting that would be advice an operator could act on and get sued for. Note the scope discipline as well: this governs your waitlist callbacks, reservation confirmations and marketing calls. An inbound call the guest chose to place is a different analysis.

Three constraints follow the consent problem, and each one has a status detail that is routinely reported wrong.

The 80/20 rule is vacated — not stayed, not enjoined, not merely unenforced. In Restaurant Law Center; Texas Restaurant Association v. U.S. Department of Labor, No. 23-50562, the Fifth Circuit issued its opinion on 23 August 2024, holding that the 2021 Dual Jobs Final Rule “fails under the Administrative Procedure Act twice over. Because the Final Rule is contrary to the Fair Labor Standards Act's clear statutory text, it is not in accordance with law. And because it imposes a line-drawing regime that Congress did not countenance, it is arbitrary and capricious.” The disposition, from the body of the opinion, reversed the district court, rendered summary judgment for the associations, and vacated the Final Rule; the panel declined remand without vacatur because the rule suffered from a fundamental substantive defect the Department could not rectify.

Then keep the dates straight, because merging them is the common error. The opinion issued 23 August 2024. The Department of Labor's subsequent Federal Register notice refers to the court's order vacating the regulatory text as issuing on 29 October 2024. DOL then removed the corresponding text from the Code of Federal Regulations and reinstated the pre-2021 regulation through a final rule that was published and effective on 17 December 2024, described as a technical amendment accounting for changes in the law that had already occurred. Three dates, three different events. What is gone: the 20-percent-of-workweek limit on directly supporting work and the 30-minute continuous limit. What governs: the restored pre-2021 dual-jobs regulation.

Allergen accuracy is the un-delegatable risk, and the data tells you exactly where it lives. The Intouch study found that AI order errors were most frequently attributed to order customization. Allergen substitutions areorder customisations. That is the empirical link between the accuracy data and the safety risk, and it needs no additional statistic. Meanwhile, at least one vendor in this roster advertises on its own product page that its AI phone agent answers questions about hours, location, parking, allergens and more — putting an allergen answer in the mouth of an automated system whose contract will disclaim responsibility for it.

The defensible boundary, stated as analysis: an agent may read a documented menu attribute back to a guest; a human must own any answer about whether a dish is safe for a specific person.The liability is not delegable to a vendor whose terms exclude it, and the escalation trigger should be a hard-coded phrase list rather than a model judgement call. We were unable to verify current FDA Food Code text this session — every URL we tried returned 404 — so we state no edition year, adoption count or section number, and you should confirm the specifics with your food-safety lead and counsel.

Accessibility is the constraint nobody budgets for. The Department of Justice's web accessibility guidance of 18 March 2022 states that the Department has consistently taken the position that the ADA's requirements apply to all the goods, services, privileges or activities offered by public accommodations, including those offered on the web, and that businesses must ensure the programs, services and goods they provide to the public are accessible to people with disabilities. Status precision: that page carries a notice that it does not reflect requirements published for state and local governments in April 2024 — that 2024 rule is a Title II rule. There is no Title III technical web or kiosk standard, so do not claim a private restaurant must meet a specific technical standard by regulation; Title III obligations run from the statute and case law.

The operational point is simpler than the legal one. A voice-only drive-thru or phone channel with no equivalent alternative is the accessibility failure mode for Deaf and hard-of-hearing guests, and the mitigation is an equivalent non-voice path — a kiosk, an app, a text channel, a staffed lane — not a better voice model. Build the alternative path into the pilot scope, not into the remediation budget.

This is an illustrative scenario, not a client result. We have deliberately not attached a real customer outcome to it, because an article that refuses unsourceable industry statistics and then asserts an unverifiable client number has destroyed its own standing. Every figure below is either something you can measure at your own sites in a fortnight, or a published price from the table above.

Consider a 14-unit fast-casual group: eight units with drive-thru lanes, fourteen taking phone orders, one central kitchen, roughly 65% of orders carrying at least one modifier. The operator has two candidate channels and a board that wants an answer about AI this quarter.

The pattern generalises beyond ordering. Once the guest-facing channel is instrumented, the same discipline applies to demand-side systems — the trade-offs of letting software move prices and inventory in a hospitality setting are worked through in our piece on AI agents in hospitality revenue and dynamic pricing, and the boundary logic is the same: automate the reversible, escalate the consequential.

One note on our own portfolio, offered precisely rather than generously. Our published Pique Beverly Hills build is a Beverly Hills aesthetic clinic, not a restaurant — the case study describes a luxury website with a VIP concierge booking system, a private consultation request flow, encrypted messaging and after-hours VIP scheduling. We cite it here for one narrow reason: the booking-and-concierge pattern, including the discretion requirements around who may see what, is structurally close to a reservation-led restaurant's phone and booking stack. We are not claiming a restaurant outcome from it, and we would rather say that plainly than let an adjacent case study imply something it does not support.

Restaurant agents rarely fail dramatically; they degrade quietly while a number on a dashboard stays green.The table below lists the failure modes we would instrument on day one, the signal that reveals each, and the rollback that contains it. Note that the detection signal is almost never “accuracy” — a single site-wide accuracy figure is the metric most likely to hide the problem.

Two of those rows deserve elaboration. Silent human-in-the-loop dependency is the failure the SEC order documents at corporate scale and the Intouch data implies at lane scale. It is also the one that is easiest to detect and hardest to want to detect, because the fix is uncomfortable: if staff intervention is doing the work, the business case was a staffing case all along. Instrument intervention events explicitly, from day one, before anyone has a reason to prefer not knowing.

Prompt injection remains unsolved.Any agent that reads untrusted text — guest reviews, delivery-platform notes, supplier emails, inbound messages — can be instructed by that text, and no vendor filter reliably prevents it. Treat every detection percentage you are quoted with the same scepticism you now apply to accuracy figures. The workable controls are structural: keep the reading agent and the writing credential in separate sessions, scope credentials to the narrowest possible operation, log every tool call, and make sure the kill switch revokes the credential rather than pausing the code. That is blast-radius reduction, not prevention, and anyone selling you prevention is selling you something that does not exist yet.

Write the boundary down before you deploy, because it is much harder to argue for after an incident. The table below is the version we would put in front of a franchisee group: what an agent may do alone, what it may prepare for a human to commit, and what it must never touch regardless of how well it performs in testing.

The three “never” rows are not about capability. An agent could plausibly generate an allergen answer, run a pay run and draft a termination letter today. They are about where accountability has to sit when the answer is wrong: with a person who can be trained, supervised, and held responsible, working inside a process an auditor or a plaintiff's lawyer can reconstruct. Every one of those three has a real-world consequence that no vendor indemnity meaningfully covers.

The rows in the middle are where the actual value sits. An agent that captures a clean order, books a reservation, drafts a purchase order for a manager to approve and prepares a schedule for a wage-hour-trained human to sign off is doing substantial work with a defensible audit trail. That is a smaller claim than “autonomous restaurant operations,” and it is the claim that survives contact with a lane at Friday dinner.

If a published-price product covers your workflow, buy it — we will tell you so on the call.A $399-per-location phone agent that answers your reservation line is not a build, and pretending otherwise would be bad advice. Custom work earns its keep when the workflow crosses systems you already own, when per-site configuration matters, or when the governance around consent and escalation has to be yours rather than a vendor's default.

Senior-led work runs $150–$225 per hour. Ongoing retainers run $2,500–$9,500 per month, which for restaurant agents typically covers menu-drift monitoring, the regression suite, consent configuration as you open sites, and vendor change management. Every build carries a 30-day post-launch warranty, and full source-code and IP ownership transfers to you at delivery. We return a fixed-price phased proposal within 5 business days of a discovery call. Frenchy Digital is a senior-led, Black-owned agency based in Los Angeles; how we scope and stage this kind of work is described on our AI agent creation service page.

One sequencing note that saves money more often than any technical decision. Do the consent and contract work in the discovery phase, not after the pilot succeeds. Redlining a model-training clause during negotiation costs one email; discovering after rollout that your vendor has been training on eighteen months of guest audio across three states is a problem with no cheap ending.

Each of these is something we observed while building this ranking, not a generic warning list. Treat any two of them appearing together as a reason to slow the process down.

• An accuracy or automation figure with no published definition. Ask what the denominator is and whether remote human agents count as automated. The SEC order turned on exactly that distinction.
• Refusal to state a human-in-the-loop rate in writing. No vendor in this roster publishes one voluntarily; the ones that will commit to a number contractually are telling you something the others are not.
• A model-training clause the vendor will not narrow. Under the CIPA capability theory this is the term that can convert a service provider into a third party. Silence on deletion schedules is the same red flag wearing a suit.
• Compliance badges with no reachable report. A badge is a graphic. A trust centre with named report types and dates is evidence. In this roster, one vendor cleared that bar.
• A product page that has not been updated in over a year while the sales pitch is about the newest capability. One vendor's AI page states its own last-updated date as December 2024.
• A flagship product you cannot find on the vendor's own site. If the URL 404s and the only references are secondary marketing coverage, do not put it in your evaluation matrix.
• A revenue calculator in place of a case study. Missed-call and phone-abandonment revenue estimates have no methodology, and we refuse all of them regardless of which vendor is publishing.
• Case-study figures with no method — a headline dollar amount attributed to a restaurant with no baseline, no period and no counterfactual.
• Stale corporate facts in the vendor's own materials: a competitor described as publicly traded when it was taken private, or an acquired product still marketed as independent.
• No published pricing combined with no published compliance and no named integrations. Any one of those is common. All three together means every claim in the sales cycle is unverifiable by construction.
• Reluctance to run a four-site pilot with your instrumentation rather than their dashboard. The measurement should be yours.
• A promise that prompt injection is handled. It is not handled anywhere by anyone. The right answer is a description of blast-radius controls.

This section is part of the product, not a disclaimer at the end of it. Everything below is something we tried to check on 23 August 2026 and could not, or could only check through a secondary source. Naming it is the difference between a ranking and a sales page.

• No independent benchmark of these commercial products exists. The Intouch Insight study measures restaurant brands rather than vendors, never identifies which vendor powered which lane, and rests on 120 AI orders across three chains. It is the best third-party evidence available and it is not a product benchmark.
• No vendor in this roster discloses whether remote human agents complete or correct orders today. The only established instance is the SEC's finding about Presto Automation for the 2021–2023 period. This is the single most important undisclosed variable in the category.
• SOC 2, PCI DSS and ISO status could not be verified for most of the roster. Trust pages were absent, returned 404, or were gated. Only one vendor publishes full detail; two display badges without a reachable report.
• DPA availability, data residency and retention terms are not publicly disclosed by any vendor here. One publishes them behind an access-request gate. No vendor in this category publishes a DPA openly.
• Bikky could not be confirmed to be operating — both domains failed to resolve for us. We do not assert it shut down; we assert only that we could not confirm it trades, which is why it is not ranked.
• The Toast product referred to in secondary coverage as Sous Chef could not be confirmed on Toast's own site; the product URL returns 404. We therefore print neither the product as fact nor the average-order-volume figure attributed to it.
• The McDonald's 2026 drive-thru retry could not be confirmed from a credible source; the trade article we tried returned 403. We say only that McDonald's ended the previous IBM test, off in all restaurants by 26 July 2024, and is reported to be running a small new one. We do not print the order-volume or accuracy figures circulating about it.
• OpenTable's ownership was not verified this session, so we assert no parent company for it.
• Current FDA Food Code text could not be retrieved; every URL we tried returned 404. We state no edition year, adoption count or section number, and the allergen guidance here is reasoning, not regulation.
• Current federal tip-taxation provisions could not be verified this session, so no deduction amount, tax year, phase-out or occupation list appears anywhere in this article.
• Whether an AI agent changes tip-credit treatment under the restored dual-jobs regulation is unanswered by any source we located. Our framing is analysis and is labelled as such.
• The Seventh Circuit's April 2026 BIPA retroactivity decision is reported here from a secondary bar-association summary; we did not read the opinion.
• The Ambriz order itself was not fetched; we describe it through law-firm analyses and do not quote the court directly.
• The BIPA voiceprint class actions filed in May 2026 are described from the plaintiffs' own law firm's account. They are allegations, not findings.
• Turnover arithmetic is ours. BLS did not return an annual value, so we summed the twelve published monthly total separations rates for accommodation and food services — a series that includes hotels, because no restaurant-only series exists.
• Vendor customer lists, location counts and funding figures are as published by the vendor or reported in trade press, and we did not independently audit any of them.