For Hong Kong's export desks: which replies may be fast without a human?
Let Google AI Mode, ChatGPT or an AI reply tool answer a buyer at once only with facts you have already written down: the receipt, the catalogue, the minimum order, the standard lead time, the certificates you hold. Anything that commits money or changes terms, such as a quote, a discount, payment terms, bank details or a complaint, waits for a person.
The buyer writes at night; the reply that comes first is the one read
It is three in the morning in Kwun Tong. A buyer in Hamburg asks Mr Chan's trading firm for the minimum order and the lead time on a carton of LED drivers, and sends the same email to two other suppliers.
Mr Chan's desk knows every product. By the time it opens the email, a competitor has already answered.
So the question is not "should we use AI?" A Hong Kong Business commentary put the real one in October: "which replies are allowed to be fast without a human in the loop?" 1. Its author sells such tools; the question still stands.
Hong Kong exported more than HK$5 trillion of goods in 2025, up 15% on 2024 3.
Fast is safe for facts you have written down, not for promises
Think of the AI as the junior who joined the export desk this week. They may send the catalogue and confirm an email arrived. They may not sign a quote. The commentary uses the same picture: "a junior colleague who prepares work, not a signatory" 1.
| The buyer asks | AI may answer at once? | Why |
|---|---|---|
| Did you get my enquiry? | Yes | A receipt commits you to nothing |
| What are the specifications? | Yes, from the published catalogue | The answer is already written and checked |
| What is your minimum order and standard lead time? | Yes, if both are written on your site or price list | A stated fact, the same for every buyer |
| Which certificates do you hold? | Yes, only those listed with their numbers | A certificate claimed in error is a compliance problem |
| What is your price for this volume? | No, a person | A quote is a commitment |
| Can you do better terms, or a discount? | No, a person | It changes money and Incoterms (who pays shipping and when risk passes) |
| Please send to our new bank account | No, a person, by phone | The classic opening of an email scam |
| The goods arrived damaged | No, a person | A complaint needs judgement and an apology someone means |
The commentary draws its gate in the same place: numbers, dates or exclusivity wait for a person 1. The AI may repeat; it may not decide.
Most sites never write down what a buyer asks first
An AI that answers "from your facts" can only use the facts it can find. Between 13 September 2026 and 10 October 2026 we read the text of 385 business websites in Singapore and Hong Kong. 94 of them, about 24%, state a price in words.
The limit, plainly: those were local service trades, not exporters; we cannot say how many trading firms state a lead time. Writing a number down is still the step most businesses skip.
A reply tool facing a blank page does one of two things. It asks a person, which is slow. Or it fills the gap, which is worse: Hong Kong's privacy regulator says confident wrong answers ("hallucinations") tend to be inherent in these models 4. A Canadian tribunal held an airline responsible for the wrong refund rule its website chatbot gave a customer 7.
The machine does not know your factory shuts for the holiday. Nobody told it.
Money, terms and bank details stay with a person, always
The bank-details row is the one not to bend. Email scams in Hong Kong are fewer than they were: 767 email fraud cases in 2020, 128 email scam cases in 2025, by police figures 56. But in 2020 alone they cost victims more than HK$2 billion 6, and the usual opening is an ordinary-looking email asking you to pay or be paid into a new account.
A polite, quick, eager AI is exactly the colleague a scammer hopes to reach. Any message about payment or a changed account goes to a person, who calls back on a number already on file.
The AI reply belongs inside your order process
First draw the path of an enquiry on one page, from the email arriving to the goods shipping, and mark where the AI reply sits. Bolted on outside it, a reply is fast and wrong just as fast. Three guardrails keep it accurate:
- It reads your order system. Asked for an urgent order, it checks stock and open orders first, so its draft says whether the date can be met. A person confirms it.
- It reads one documented product list, kept current by one named person, and answers from nothing else.
- It knows what it may say alone. For anything else it drafts, and a colleague confirms. One click still beats a morning's wait.
Write the answers first, then switch the AI on
"My buyers want a person." They do, for the price. For the specification sheet at three in the morning, they want an answer. The order of work:
- Read your sent folder for a week. List the questions buyers ask most, in their words.
- Write the fixed answers on your website, where machines can read them: minimum order, standard lead time, sample policy, usual Incoterms, certificates held.
- Let the AI answer only from those, and from your order system. For everything else it replies that a colleague will answer within one working day, and routes the email with its draft.
- Read every reply it sends for a week and note each one you would have changed 1. Widen its list only when that note stays short.
Not yet worth paying for: a custom-built agent, before the facts it should repeat exist in writing.
A two-minute test before any tool is bought
Ask ChatGPT, or Google with AI Mode, "what is the minimum order and lead time of [your company] for [your main product]?" Does it find your answer, give someone else's, or invent one? One answer is an observation, not a diagnosis.
Getting a business found and described correctly where its buyers now ask is our work at AIqSight: what a machine can read on your site, which facts are missing, which fix is worth most. Nobody can promise that an AI will name or quote a firm, and we make no such promise. The free scan of your website is the place to begin.
Questions people ask
Can an AI answer export enquiries without a human?
Yes, for replies that repeat facts you have written down: a receipt, specifications from your catalogue, your stated minimum order and standard lead time, certificates you hold. Anything that sets a price, changes terms or mentions payment waits for a person.
Which buyer replies should always go to a person?
Quotes and volume prices, discounts, payment terms and Incoterms, any change of bank details, complaints about damaged or late goods, and enquiries about unusual destinations or controlled items. Each commits money, changes terms or needs a check an AI cannot make.
What should an AI reply tool be connected to before it answers buyers?
Your order system, so it can check stock and open orders before replying to an urgent order, and one documented product list it answers from. Add a written list of replies it may send alone; it drafts the rest for a person to confirm.
What happens if an AI reply gives a buyer a wrong price or term?
You may be held to it. A Canadian tribunal held an airline responsible for the wrong answer its website chatbot gave. The safer design: an AI never sends a number it did not copy from your own published facts or your order system.
Does Hong Kong law require a human to check AI replies?
Not as a general rule for trade emails. The Privacy Commissioner's AI framework is guidance: it advises more human control as the risk of an AI's output rises, up to a person keeping the decision. Its examples concern individuals; the scale suits quotes too.
Next: How does AI read your website, and what can it see?
Sources
- Hong Kong Business (commentary by Liu Wei, Corpable Marketing), How Hong Kong SMEs should set boundaries for agentic AI in B2B selling, 12 October 2026. Accessed 23 October 2026.
- Trade and Industry Department, HKSAR Government, What are SMEs?, undated. Accessed 23 October 2026.
- Census and Statistics Department, HKSAR Government, External merchandise trade statistics for December 2025, 27 January 2026. Accessed 23 October 2026.
- Office of the Privacy Commissioner for Personal Data, Hong Kong, Artificial Intelligence: Model Personal Data Protection Framework, 11 June 2024. Accessed 23 October 2026.
- Legislative Council Panel on Information Technology and Broadcasting (Government paper, police figures), Work Related to Safeguarding Information Security, LC Paper No. CB(2)1000/2026(03), Annex II, 13 July 2026. Accessed 23 October 2026.
- HKSAR Government (Hong Kong Police Force figures), LCQ14: Combating online and telephone frauds, Annex 1, 15 November 2023. Accessed 23 October 2026.
- Civil Resolution Tribunal of British Columbia, Moffatt v. Air Canada, 2024 BCCRT 149, 14 February 2024. Accessed 23 October 2026.
Our own counts come from the AIqSight Q4 2026 data brief: business websites in Singapore and Hong Kong scanned 13 September to 10 October 2026, rubric 0.2.0, 385 sites, 385 sites whose text we read (10 local trades); not exporters, scored with the published rubric. Not a market survey.