Hong Kong had 17,235 licensed restaurants this morning, according to the Food and Environmental Hygiene Department's own licence file, which is republished every day at nine. Almost all of them serve customers who order in Cantonese, often mixed with English.
Not one of the major restaurant voice-AI platforms sold worldwide advertises Cantonese support. Check their pricing pages and the bilingual claim is Spanish.
That single gap explains most of what works and does not work when a Hong Kong restaurant buys AI. This guide separates the two, using published figures rather than vendor promises.
What can AI actually do in a restaurant today?
Five jobs have working evidence behind them: forecasting how much food to prepare, writing marketing and menu content, handling repetitive customer messages, drafting review replies, and helping with rosters. Everything else is either a pilot or a sales pitch. The tasks that work share one trait: a mistake is cheap, and a person sees the output before a customer does.
A survey of more than 340 Hong Kong SMEs commissioned by Dah Sing Bank in May 2026 found the same shape across all industries. Marketing and content creation was the top use at 56 per cent, customer service 42 per cent, sales support 35 per cent, data analysis 30 per cent, and process automation or workflow improvement only 14 per cent.
That last number is the interesting one. Four times as many small businesses use AI to write posts as to remove steps from a process, and in a restaurant it is the process work that frees up hours.
Does AI order-taking work in Cantonese?
Not reliably, and there is a number for it. On real conversational Cantonese, a leading general-purpose speech model gets roughly two words in five wrong. On clean, read-aloud Cantonese it gets about one in seven wrong. Taking an order across a noisy counter is the first kind of speech, not the second.
The figures come from the WenetSpeech-Yue research released in September 2025. Whisper-large-v3 recorded a mixed error rate of 45.09 per cent on conversational Cantonese against 15.46 per cent on clean read speech. After Cantonese-specific fine-tuning, the conversational figure fell to 18.69 per cent, which tells you the problem is solvable but not solved by default.
The reason is data. Cantonese has a fraction of the annotated speech that Mandarin has, and Hong Kong customers do not speak in clean sentences anyway. A separate Hong Kong University of Science and Technology dataset published in 2024 found that none of the standard Cantonese speech corpora contained mixed Cantonese-English speech at all, which is exactly how a customer asks for a set lunch.
Then there is the industry record. McDonald's ended its AI drive-through partnership with IBM in June 2024 after a trial in more than 100 restaurants, with two sources telling CNBC the system had trouble interpreting different accents and dialects. Presto Automation disclosed in a filing that 70 per cent of orders taken by its voice product required human intervention, after marketing had claimed 95 per cent needed none. In August 2025 Taco Bell's own technology chief told the Wall Street Journal, of the voice AI running in more than 500 locations, "sometimes it lets me down".
The practical conclusion for a cha chaan teng: voice ordering is the worst possible first AI project in Hong Kong, and the most heavily marketed one.
What is already working in Hong Kong?
Three named local deployments have public detail behind them, and each one is doing something narrower than "AI runs the restaurant". One handles ordering in a controlled setting, one forecasts ingredients, one weighs waste.
The airport case. On 7 August 2026 an AI ordering agent branded 「AI 店小二」 launched at a cha chaan teng in Hong Kong International Airport, developed by a local company after 18 months of work. Reported across HK01, Hong Kong Economic Journal and i-CABLE, it takes voice orders in Cantonese, Putonghua and English, and handles cha chaan teng shorthand such as 「A餐」 and 「凍檸茶少甜少冰」. It also reads a flight number and suggests dishes that fit the remaining wait. Note what makes this the easy case: a queue of one customer at a time, a fixed menu, and a controlled acoustic environment. The founder's stated aim was not cutting headcount but strengthening the brand.
The forecasting case. Maxim's has run demand and ingredient forecasting on IBM's platform across its Hong Kong outlets since around 2018. Its digital solutions lead told IT PRO Magazine in 2023 that AI helped prepare ingredients more precisely and let colleagues finish work half an hour earlier. It is a 2023 account with no newer public update, so treat it as evidence that the approach works at scale, not as this year's news.
The waste case. Mandarin Oriental, Hong Kong reported a 73 per cent reduction in food waste and annualised savings of HK$375,000 using Winnow's camera-and-scale system in the kitchen, in a case study published by the vendor and quoting the property's executive sous chef by name. A separate group-level figure for a four-hotel pilot is 36 per cent, so read the 73 per cent as one property's result rather than a guarantee.
Where does the money actually come from?
Two places, and neither is the one restaurants usually buy first. The savings sit in food you did not throw away and hours you did not pay for twice. Both require measuring something you currently guess at.
Start with waste, because there is independent evidence. Champions 12.3, working with the World Resources Institute, studied 114 restaurants across 12 countries and found kitchen food waste down 26 per cent on average within a year, with restaurants saving seven dollars for every dollar invested. Read the fine print though: that study covered measurement, staff training and menu redesign, not AI. Most of the win came from measuring at all. AI makes the measuring cheap and continuous, which is a real contribution and a smaller claim than the vendors make.
Then look at hours. Hong Kong's food and beverage sector employed about 218,900 people as at March 2025 with an average of only 12 staff per establishment, down from 16 two decades ago. Fewer people are doing the same shifts. The value of AI here is not replacing a cook, it is removing the forty minutes an owner spends every evening on ordering, rosters and social posts.
One correction worth making before you budget: the sector is not simply short of workers. Government figures for March 2026 record 3,630 vacancies in food and beverage services, down 25 per cent year on year. Total restaurant receipts in the second quarter of 2026 were HK$27.2 billion, up 0.4 per cent in value but down 0.4 per cent in volume. Prices are carrying the revenue, not covers. That is a margin problem, and margin problems are solved by cost per order rather than by a shinier ordering screen.
What does it cost, and can a grant pay for it?
Real Hong Kong prices are surprisingly hard to find because most vendors publish nothing. The ones that do publish sit between roughly HK$900 a month and HK$100,000 up front, and as of today there is a gap in government funding you need to know about.
Prices that are actually published:
--- iCHEF lists HK$888 per month on its Hong Kong site, with no service fee or deposit and an extra tablet at HK$198 per month.
--- Eats365 does not publish a standalone price, but its listing on the government's own solution portal shows a base plan at HK$48,888 and HK$99,888 with the AI sales and purchasing analysis module added, as 24-month bundles including hardware, installation and training.
The funding position matters more than any of those numbers. The Digital Transformation Support Pilot Programme funded up to 50 per cent of a project or HK$50,000, whichever was lower, and approved 8,799 local SMEs with an average subsidy of about HK$49,000. Food and beverage took 34 per cent of those approvals. Applications closed in May 2025.
An enhanced round with a HK$300 million budget, explicitly covering off-the-shelf AI and cybersecurity solutions, was confirmed in February 2026 with a target of launching in the second half of the year. As at 27 August 2026 the programme portal still displays a closed-for-applications notice. The money is announced and budgeted; the window is not open. Watch the portal rather than committing to a purchase on the assumption a subsidy will arrive.
One trap to avoid. The BUD Fund raised its Easy BUD ceiling from HK$100,000 to HK$150,000 on 15 June 2026 and now promises more targeted support for projects involving AI. But BUD exists to help Hong Kong enterprises develop markets outside Hong Kong. A single-outlet restaurant serving local customers is effectively outside its scope, whatever a vendor's brochure implies.
What should a restaurant owner do first?
Pick the task you repeat most often where nobody is waiting. That rules out the front counter and rules in the back office, which is the opposite of how this technology is sold.
A sensible first ninety days:
--- Weigh what you throw away for two weeks. No software needed. The number itself usually changes the ordering behaviour before any AI is installed.
--- Move one repeating written job to AI. Weekly social posts, menu translation, replies to reviews. Cheap to try, obvious when it fails, and a person always checks before it goes out.
--- Handle repeat enquiries in writing, not by voice. Opening hours, whether you take a booking for six, whether there is a vegetarian option. Text-based handling avoids the Cantonese speech problem entirely because the customer types. Leave forecasting until last, since it pays best but needs several months of clean sales history.
Three things owners get wrong
"An AI ordering system will let me run one less person on the floor." The two largest documented deployments say otherwise. Presto needed human intervention on 70 per cent of orders, McDonald's ended its trial, and the Hong Kong developer of the airport system explicitly framed it as brand-building rather than headcount reduction.
"If it handles Mandarin it will handle Cantonese." Language support is not one setting. The published error gap between Cantonese and better-resourced languages is large, and performance varies enormously depending on which model sits underneath the product. Ask which model, then test it with your own staff saying 「凍檸茶少甜」 at lunch volume.
"We should wait for the subsidy." Waiting has a cost. The technology voucher scheme closed to new applications at the end of 2024, the digital transformation round closed in May 2025, and the AI-inclusive round is still not open. Two of the three most useful first steps above cost nothing at all.
Frequently asked questions
Is any Hong Kong point-of-sale system genuinely AI-powered? Eats365 sells an AI sales and purchasing analysis module, which is bundled with inventory and procurement rather than built into the core system. iCHEF's AI push is documented as a Taiwan rollout, and its Hong Kong site does not mention AI at all. Ask for a Hong Kong reference customer, not a regional press release.
Can AI reply to my OpenRice or Google reviews automatically? Not as a shipped feature you can switch on here. Google was spotted testing AI review replies in March 2026 with selected accounts in the United States, Brazil and India, and no equivalent tool was found on OpenRice. Drafting a reply with a general AI tool and posting it yourself works today.
What about delivery platform commissions, can AI help there? Not directly, and this is where the real margin goes. After Deliveroo left Hong Kong in April 2025, merchants reported commission notices rising from 25 per cent to 35 per cent. No AI tool changes a commission rate. Knowing your true cost per delivered order does change which orders you chase.
The takeaway
The AI that helps a Hong Kong restaurant in 2026 is quiet and unglamorous. It counts what you waste, drafts what you have to write anyway, answers the same three questions in text, and eventually tells you how much fish to order on a rainy Tuesday. The AI that takes orders in Cantonese across a busy counter is coming, and the published error rates say it is not here yet.
Start with the two-week waste count. It costs nothing, it needs no vendor, and it tells you whether the rest is worth buying. We understand AI. UD stands with you.
Reviewed by the UD AI team, Hong Kong, 27 August 2026.
Related reading: What Is AI Workflow Automation?, Can the BUD Fund Pay for Your AI Project? and What Online Booking Costs in Hong Kong.
Not sure which job to hand over first?
Every restaurant has one task that repeats daily and pays nobody back. Finding it is the whole first decision, and it is easier with someone who has seen a few hundred of them. We will walk you through it step by step, from picking the task to measuring whether it actually saved you anything.
