A single order in a Hong Kong import and distribution firm of 10–15 people passes through nineteen manual steps, re-typing the same twelve fields into four documents.A single order in a Hong Kong import and distribution firm of 10–15 people passes through nineteen manual steps, re-typing the same twelve fields into four documents.
It is three o'clock on a Friday in Sai Ying Pun. A staff member at a food import company is typing a product code, a unit weight and a case count into a packing list. She typed the same three fields into a purchase order on Tuesday, into a delivery note an hour ago, and into an invoice ten minutes before that.
Twelve fields. Four documents. One order. She has nine more to close before the end of the day.
In short
- AI automation for a small business means removing the re-entry and the chasing, not the judgement.
- A single order in a Hong Kong import and distribution firm of 10–15 people touches nineteen manual steps from enquiry to invoice. All nineteen are listed below.
- The same twelve data fields get re-typed into at least four documents — not because anyone is slow, but because the documents were never built from the same record.
- What stays human: the supplier relationship, and any price a customer will argue about.
Where the hours actually go: one order, nineteen manual steps
Follow one order through a Hong Kong food importer— the sequence anyone in this business will recognise.
Before you can quote
- Read the enquiry through a WhatsApp message: "Can you do 200 cases of the frozen dumplings at the usual?"
- Look up the buyer's pricing tier, in a spreadsheet on a desktop. Not the cloud but the desktop, because that is where it lives.
- Check stock in the Sheung Wan office.
- Check stock at the Dongguan warehouse through WeChat.
- Check stock at the third-party cold store by phone.
- Build the landed cost: freight from one source, duty from another, FX from a third.
Turning it into an order
- Assemble the quote by hand, applying the discount from memory or an old quote.
- Send it on WhatsApp. Re-price if the buyer negotiates.
- Type the accepted order into the order book-- Excel, sometimes backed by a paper notebook.
- Re-type the same details into a purchase order.
- Send the PO by email or WeChat.
Chasing
- Chase order acknowledgement.
- Chase the shipping schedule.
- Chase again by phone when the email goes unanswered. In peak season — before Chinese New Year, before Mid-Autumn — this alone can take a day to a week.
Documents, goods and billing
- Re-enter codes, descriptions, declared values and HS tariff codes into the customs paperwork.
- Arrange local delivery from port to the right warehouse, usually by phone.
- Count physical stock against the shipping documents. Short counts and substituted grades are common enough that nobody skips this.
- Update three separate stock records: office, warehouse, cold store.
- Produce the delivery note, the invoice and the packing list — same fields, three times.
Nineteen steps. Twelve fields recur throughout: product code, description, unit of measure, quantity, price, total, customer, delivery address, order reference, date, weight, case count.
At fifty orders a month across four documents each, that is those twelve fields entered roughly two hundred times.
The steps exist because the documents are not connected — and they are not connected because nobody had a reason to connect them until the volume made it painful.
What AI automation can do for an import and distribution business
| Workflow | What it costs now | What an agent can do | What stays human |
|---|---|---|---|
| Landed-cost calculation | Rebuilt per order. Errors compound into margin loss | Pull freight, duty and FX automatically; produce a costed line in seconds, every input shown | The final pricing decision — especially when a customer will argue the number |
| Quotation assembly | Manual lookup, manual discount, manual formatting | Generate the quote from the enquiry, applying the buyer's tier and current landed cost | Negotiation and relationship pricing — where the number is a conversation, not a formula |
| Supplier chasing | The first thing dropped when the office is busy, which is when it matters | Track every open PO, follow up on schedule, escalate when overdue, log the status | The call when something is actually wrong — a quality issue, a substitution, a missed shipment |
| Document generation | Four documents, twelve overlapping fields, typed separately | Produce all four from one confirmed order record | Review before sending. A misquoted weight is read closely by people outside your building |
| Stock reconciliation | Three locations, three records, reconciled from memory | Maintain one position, updated from each location, flagging disagreements | The judgement call when counts disagree and someone decides which number is real |
The right-hand column is never empty. Any proposal where it is empty is a proposal to remove the judgement your margin depends on.
What should a small business automate first?
Start with the workflow that has real volume, rules you could write down for a new hire, a tolerable cost of error, and one named person who will actually use it. We call these four questions the Ledger Test. A workflow that fails any one of them is not ready, however appealing it looks.
Volume. More than twenty times a month? Document generation in a firm handling fifty orders clears this easily. The supplier relationship call does not.
Rules. Could you write the decision down for a new hire? Landed cost follows a formula — freight, duty, FX, margin. It is long, not complex. Pricing negotiation follows a relationship.
Cost of error. What breaks, and who notices? An internal stock error is caught on Monday by the person who always catches it. An error on a customs document is caught outside your building, later, at cost — a reason to design sign-off in, not a reason to avoid it.
Ownership. In a 10–15 person firm there is usually one obvious candidate: the operations person or the buyer who handles order flow end to end. If nobody owns it, nobody uses it, and the build is wasted.
For most import and distribution firms, document generation is the first thing worth building — not because it is impressive, but because it touches the most steps with the least change to how people work. She still reviews the invoice. She still checks the packing list against the count. She is no longer typing the same product code for the fourth time.
What we would tell you not to automate
The supplier relationship.
The call to a factory manager in Dongguan when a shipment is late is not a script. It is a negotiation, conducted in Cantonese, informed by years of dealings and a sense of when to push and when to wait. An agent that sends a polite follow-up is useful. An agent that tries to replace that phone call is a liability.
The same applies to any price your customer will argue about. The moment a number becomes a conversation, automation is the wrong answer.
Does the BUD Fund cover AI automation?
Yes. In July 2026 the Government confirmed that the BUD Fund covers AI adoption, and that the Digital Transformation Support Pilot Programme received a further HK$300 million on top of its original HK$500 million, having supported around 8,800 SMEs.
For most owner-operated importers that turns an unbudgeted cost into a possibly co-funded one. Eligibility and caps vary by scheme and change over time, so check the current terms — the Government's July 2026 statement is here — or ask your funding advisor. We build operational systems; we are not funding or compliance advisors.
Common questions
How long does the mapping take?
Four to six weeks from first conversation to a written report. The mapping itself is one or two working sessions — with the people who actually handle the orders, not with management.
Do we need to change our existing tools?
No. Most firms in this position run on Excel, WhatsApp and email, and that is the normal starting condition. The agent reads from and writes to those tools rather than replacing them.
Where does our data go?
Your files stay where they are. For each task, only the fields that task needs are processed by a third-party model provider, under terms that disable training on your data and minimise retention. We show you the configuration and write down what touches what.
What does AI automation cost for a small business?
The audit is free — a live readout and a written report, yours to keep whether or not you proceed. Build costs are bespoke, scoped from the audit and quoted per project. See the BUD Fund section above before assuming the number is all yours to carry.
What if the audit finds nothing worth automating?
We tell you. The audit answers whether this is worth doing in your firm, not whether we can sell you something. If nothing passes the Ledger Test, the honest answer is to wait.
Can we start with one workflow and expand later?
Yes, and we recommend it. One workflow, end to end, four to six weeks to first meaningful test. Scale only after the first one works.
Want to know whether this is worth doing for your company? We rank three AI opportunities based on your company's internal processes.
Begin a correspondence → studio@agenticmaison.com