AI for
Hospitality
Wage bills are up. The sector lost roughly 100,000 jobs since the autumn 2024 Budget. Two venues a day are closing. We build the AI loops that actually move the needle behind a front desk, a pass and a phone line, without quoting rates you can't honour or allergen advice that ends up in tribunal.
~100k
JOBS LOST SINCE OCT 2024 BUDGET (UKH)
£1.4bn
EXTRA WAGE COST FROM APR 2026 (UKH)
2/day
VENUES CLOSING H1 2025 (CGA/ALIXPARTNERS)
Most hospitality AI is a chatbot bolted onto a booking widget.
It answers ten questions, gets the eleventh wrong, and quotes a rate the PMS won't honour. In February 2024, the BC Civil Resolution Tribunal made Air Canada pay out for exactly that. Your chatbot is your contract.
The interesting AI in hospitality isn't on the home page. It's the voice agent taking the booking when the host's on the floor, the model that knows which Tuesday in November will sell out, the loop that drafts a reply to a one-star review before the GM finishes their coffee.
That's the actual work. Quieter than a chatbot, and it pays back.
WHAT YOU'VE BEEN SOLD
- Chatbot that hallucinates the cancellation policy
- AI review replies thanking the wrong guest
- Allergen answers from a model with no menu source of truth
- Voice agent that loops and never escalates
- No DPIA, guest data piped into someone's prompt
WHAT ACTUALLY WORKS
- Voice agent wired into the PMS, knows real availability
- Review drafts a human signs off, sent inside an hour
- Allergen answers from your recipe spec, not a guess
- Demand model the revenue manager actually trusts
- Logs, consent and a DPIA your ICO answer points at
Five loops worth building. The rest is theatre.
Hotels, restaurants, pubs. Different shapes of the same problem: limited covers, limited rooms, limited people, demand that won't sit still. The loops that pay back close those gaps.
Voice booking
Inbound calls answered, modified, cancelled. Wired to OpenTable, SevenRooms, Mews, Cloudbeds or your own PMS, so it only offers what's actually available. Hands the call to a human the moment it should.
Revenue & demand
Forecasts that pull in weather, events, comp set, local school holidays. Feed into IDeaS, Duetto or Lighthouse if you have them, or stand on their own if you don't. Surfaces rate moves the GM can actually defend.
No-show reduction
Zonal's GO Technology survey put UK no-shows at around 12% of bookings. We score every reservation, send the right reminder at the right hour, and only ask for a deposit on the ones that need one.
Reviews & guest comms
Google, TripAdvisor, Booking.com pulled in, clustered by theme, drafted into replies that name the guest, the room or the dish. Human signs off. No hallucinated apologies, no replying to the wrong booking.
Back of house
Rota forecasting on top of Fourth, Harri or Deputy. Allergen lookup from your recipe spec, not from a model's memory. Waste tracking that points the chef at the bin, not at a dashboard.
Hospitality has rules. AI doesn't get a pass.
It's tempting to point ChatGPT at the menu and let it answer guest questions. It's also how you end up with an allergen answer that doesn't match the recipe spec, sent on your behalf, on your domain.
We build to the rules that already apply to your kitchen and your front desk. The AI doesn't get to invent them.
Allergens come from one source.
Since October 2021 the FSA requires a full ingredient list with the 14 regulated allergens emphasised on prepacked-for-direct-sale food. Any AI that answers "is this gluten free?" reads from your recipe management system, never from training data, and escalates if the data is missing.
Guest data is special category fast.
Face check-in, voiceprints and CCTV analytics are biometric data under the ICO's 2024 guidance. That needs Article 9 grounds, explicit consent, a real DPIA and a non-biometric alternative. We design the consent flow with you, not after launch.
Cards stay out of the prompt.
Voice and chat agents handle deposits via tokenised links from Stripe, Adyen or your existing payment provider. No card number ever enters a model prompt. PCI scope stays where it already is.
Air Canada lost. So will you.
Moffatt v. Air Canada (2024 BCCRT 149) ruled the airline liable for what its chatbot promised. Every booking, rate quote and cancellation answer is grounded in PMS data, with the verbatim source attached to the message. If the system can't be sure, it says so and gets a human.
Where we come in.
We're not a hospitality SaaS. We build the AI layer that sits on top of the systems you already pay for, pick the one or two loops that pay back fastest, and launch them.
First loop live at a single site, then rolled across the estate. No twelve-month transformation programme, no rip-and-replace of your PMS.
BOOK A CALLOne-site audit
We pick a single property or restaurant. Two days on site, a look at your PMS, POS, booking flow, phone logs and last quarter of reviews. Out the other end: the three loops with the biggest payback for you, named and priced.
Launch the first loop
Usually voice booking or review response, because both pay back inside a quarter and neither touches the kitchen. Live at one site. The team uses it. You see the call volume and the recovered covers.
Add the guardrails
DPIA, allergen source of truth, escalation paths, the per-property tone of voice, the logs the ICO would ask for. The bit most vendors skip and the bit that lets you sleep before a roll-out.
Roll across the estate
Same loop, every site, with the local quirks honoured. New properties onboarded in days, not months. Optional retainer if you want us building the next loops alongside your team.
It's already running. Look closely.
The big operators have been quietly running these loops for two years. Three you can verify, not a vendor brochure.
White Castle drive-thru.
SoundHound's voice AI is live across 100-plus White Castle drive-thru lanes, completing around 90% of orders without a human, per the expansion announced by both companies in 2023 and rolled out through 2024. Also live at Chipotle, Jersey Mike's and Church's Texas Chicken.
Hilton, Marriott, Accor, IHG kitchens.
Winnow's computer-vision waste tracker runs in 3,500+ commercial kitchens across 94 countries. Globetrender reported in January 2026 that clients have collectively saved about $100m a year and 28,000 tonnes of food.
And how not to do it.
McDonald's ended its IBM drive-thru voice AI trial at 100-plus US restaurants in mid-2024 after accent and accuracy failures, per CNBC. Voice AI works. Hand-rolling it without a hard fallback to a human doesn't.
Sources: SoundHound 2024 press release; Globetrender, January 2026; CNBC, June 2024; Moffatt v. Air Canada 2024 BCCRT 149; UKHospitality; CGA / AlixPartners Hospitality Market Monitor; FSA; ICO biometric guidance 2024.
The ones operators ask first.
Will a voice agent put my reservations team out of work?
No. It clears the easy calls so they can take the ones that actually need a human: amendments, group bookings, complaints, the regulars. Operators we work with redeploy reservations agents to revenue and guest experience, not the door.
We already use IDeaS / Duetto / Lighthouse. Do you replace them?
No. They're good at what they do. We feed them better signals, like local events, weather, comp-set anomalies and demand from your own channels, and we turn their outputs into something a GM can defend on a Monday call.
What about Natasha's Law? I'm not letting AI answer allergen questions.
Good instinct. Nor are we. Where AI handles allergen queries at all, it reads from a single source of truth, the dish or recipe record in your kitchen system, and quotes it verbatim. If the data isn't there, it doesn't guess. It escalates to a named person.
Our PMS is older than Wikipedia. Can you still integrate?
Usually, yes. We've talked to PMS, POS and EPoS systems with no public API, weekly CSV exports, and even nightly emailed reports. If the data exists somewhere, we'll get it. If it doesn't, we'll tell you up front rather than after the invoice.
How do you stop the chatbot promising things we can't deliver?
Every answer that could become a contract, rates, availability, cancellation terms, deposits, comes from your system of record. The model rephrases. It doesn't invent. The Air Canada tribunal made that question expensive to get wrong, so we treat it as a hard rule.
We're a single restaurant, not a group. Is this for us?
Yes, with caveats. A single-site voice booking agent and a no-show loop pay back inside a quarter on most independent restaurants. The bigger forecasting and rota work only makes sense above four or five sites.
How much does it cost?
One-site audit is fixed-fee. First loop live for a single site is priced per phase before we start. Estate roll-out is priced per property and the unit cost falls quickly after the first three. No retainers you can't cancel.
Pick the loop. We'll build it.
Thirty minutes. Tell us the site, the systems and what hurts most. You'll come away with a clear answer on which loop to build first and whether you need us at all.