AI for SME Back Office in Hong Kong: What Works and PDPO Rules
AI already saves a Hong Kong SME back office real time in four places: reading receipts and supplier invoices, suggesting matches between bank lines and open invoices, drafting routine emails such as payment reminders, and answering plain-language questions about your own figures. It does not replace judgement on tax, provisions or payroll. Under the Personal Data (Privacy) Ordinance, your company remains the data user when staff paste personal data into an AI tool. The PCPD has issued AI guidance in 2024, 2025 and August 2026.
Last reviewed: September 2026 · Sources: Office of the Privacy Commissioner for Personal Data (PCPD), Personal Data (Privacy) Ordinance (Cap. 486)
Where AI genuinely helps today
The useful test is simple. AI is good at turning unstructured material into structured suggestions that a person can check quickly. It is poor at being the final word on anything with legal, tax or financial consequences.
Reading receipts and supplier invoices
Optical character recognition combined with a language model can read a photographed receipt or a PDF invoice. It pulls out the supplier, date, amount, currency and invoice number and proposes an expense category. This is the most mature back-office use, and it removes the most tedious typing.
The failure modes are predictable: faded thermal receipts, handwritten amounts, mixed Chinese and English layouts, and totals that include a service charge or a deposit. A person should confirm the amount and the category before the entry posts. The receipt image stays attached as evidence, which matters because business records must be kept for at least seven years.
Matching bank lines to invoices
Bank reconciliation is pattern matching: this FPS receipt of HK$18,400 from "CHAN TAI MAN CO" is probably invoice 1042. Software can rank likely matches by amount, date and payer name, and a person accepts or corrects them in a matching workbench.
The hard cases stay hard. Examples include one transfer paying three invoices, a payment net of bank charges, and a payer name that differs from the customer record. Those need a person, and the system should make that exception queue short and visible. Our month-end close checklist shows where reconciliation sits in the monthly routine.
Drafting routine correspondence
A model can draft a polite first reminder, a firmer second one, a supplier query or a reply to a routine customer question in English or Chinese in seconds. The time saved is in the first draft, not the decision. Someone still chooses who gets chased, in what tone, and whether a long-standing customer deserves a phone call instead.
For the cadence behind those reminders, see our accounts receivable playbook.
Asking questions of your data
"Which customers are more than 60 days overdue?" or "How did gross margin move by product line this quarter?" are questions an owner can now ask in plain language, where previously they would have waited for a report. The model turns the question into a query over the ledger and presents the answer.
The answer is only as good as the books behind it and the query the model wrote. Treat the result as a starting point, and check any figure you intend to act on against the underlying report.
Where AI does not belong, or not yet
| Task | Use AI? | Why |
|---|---|---|
| Extracting data from receipts and invoices | Yes, with review | High volume, low judgement, errors easy to spot |
| Suggesting bank-line matches | Yes, with acceptance by a person | A human confirms each match |
| Drafting reminders and routine emails | Yes, edited before sending | Tone and relationship calls stay human |
| Answering questions about your figures | Yes, as a first pass | Check figures before acting |
| Payroll and MPF calculations | No, use rules-based calculation | Statutory arithmetic must be deterministic and repeatable |
| Profits tax positions and deductibility | No, as the final answer | Needs professional judgement and evidence |
| Bad-debt provisions and write-offs | No, as the decision | An accounting judgement the directors own |
| Anything that sends data or money externally | Only with human approval | Irreversible actions need a person in control |
Payroll deserves a specific word. MPF contributions, holiday pay and final payments follow fixed statutory formulas. A language model can explain the formula, but the calculation itself should come from a rules engine that gives the same answer every time.
Your PDPO obligations when you use AI
The Personal Data (Privacy) Ordinance does not have an AI chapter. It applies to AI the same way it applies to any other processing, through its six Data Protection Principles (DPPs). The PCPD has spelt out what that means in three documents.
The 2024 Model Personal Data Protection Framework
The PCPD published Artificial Intelligence: Model Personal Data Protection Framework on 11 June 2024, for organisations that procure, implement and use AI systems. It is organised around four areas:
- AI strategy and governance, including a governance committee and clear roles.
- Risk assessment and human oversight, choosing between "human-in-the-loop", "human-in-command" and "human-out-of-the-loop" according to the risk.
- Customisation of AI models and implementation and management of AI systems, including using the minimum personal data needed.
- Communication and engagement with stakeholders, telling people how their data is used.
For an SME, "governance committee" can simply mean a named owner and a short written policy. What matters is that someone is accountable.
The 2025 checklist for employee use of generative AI
On 31 March 2025, the PCPD issued a Checklist on Guidelines for the Use of Generative AI by Employees. It suggests an internal policy covering five points:
- Scope of permissible use: which tools staff may use, and for what purposes.
- Protection of personal data privacy: what kinds and amounts of information may be entered.
- Lawful and ethical use and bias prevention: including checking outputs before relying on them.
- Data security: which devices and accounts may be used, and how to report incidents.
- Violations: the consequences of breaching the policy.
The checklist also recommends training, a support contact and a channel for feedback.
The August 2026 guidance on agentic AI
On 25 August 2026, the PCPD published Protecting Personal Data Privacy in the Use of Agentic AI. Agentic AI means systems that take actions, not just produce text. The guidance makes one point that every owner should absorb: an AI agent is not a legal person, so the organisation that deploys it remains responsible under the PDPO.
It makes nine recommendations:
- Avoid excessive collection. This applies especially to sensitive data such as HKID numbers, bank account numbers and passwords.
- Be transparent. Mention agentic AI in your Personal Information Collection Statement and Privacy Policy, and consider stating where data is stored or processed.
- Ensure accuracy. Use techniques such as retrieval-augmented generation, plus human review.
- Set retention periods for conversation histories, caches and memory.
- Do not use personal data for a new purpose without consent.
- Secure the system. Use official versions, separate runtimes and cautious plugin use. Grant the minimum access rights and never administrator rights. Add guardrails and logging for traceability.
- Uphold access and correction rights by choosing systems built for privacy by design.
- Conduct continuous risk assessments, keeping a person in the loop for high-stakes or irreversible actions such as sending personal data to third parties.
- Assign clear responsibilities and provide training, and bind external vendors contractually on retention and security.
The PCPD's artificial intelligence resource page collects all three documents.
Your AI provider is a data processor
When you send customer or employee data to an AI service, the provider is usually processing it on your behalf. DPP 2(3) and DPP 4(2) require you to use contractual or other means to stop a data processor keeping the data longer than necessary and to protect it against unauthorised access. The PCPD's leaflet on outsourcing to data processors explains what those contracts should cover.
Where JetFlow fits: JetFlow routes every AI feature, from Hot Folder receipt reading to ask-your-data questions in its BI module, through one gateway. That gateway logs each call's model, provider and the categories of personal data detected in the request, and redacts error messages. See how JetFlow handles AI in the back office.
An SME governance checklist
Most SMEs do not need a committee. They need these ten things written down and followed.
- Name an owner for AI use, usually the finance lead or operations manager.
- List approved tools and ban personal accounts on free consumer chatbots for company data.
- Classify what may go in. Customer names and invoice amounts may be acceptable in an approved tool. HKID numbers, bank account numbers, medical details and passwords should not go in.
- Strip what you do not need. A reminder draft needs the amount and the due date, not the customer's home address.
- Require human review before any AI output is posted to the books, sent to a customer or used in a decision about an employee.
- Keep the evidence. Retain source documents and a log of which AI outputs were accepted and by whom.
- Sign data processing terms with every AI vendor, covering retention, security, sub-processors and deletion.
- Update your privacy notices to say that you use AI tools and for what.
- Train staff once a year and when tools change.
- Review quarterly: what went wrong, what was corrected and which tools are no longer needed.
How to evaluate an AI vendor
Ask these questions in writing and keep the answers.
| Question | What a good answer looks like |
|---|---|
| Is our data used to train your models? | No, by default, and in the contract |
| Where is data stored and processed? | Named regions, with named sub-processors |
| How long do you keep prompts and outputs? | A defined period, with deletion on request |
| Can we see a log of what was sent? | Yes: per call, with user, time and data categories |
| What happens when the AI is unavailable? | Work continues manually; nothing is silently posted |
| What security standards do you follow? | Recognised standards and an incident-notification commitment |
| Can a person approve before anything posts or sends? | Yes, and it is the default |
The PCPD's 2024 framework specifically suggests checking that AI suppliers follow international technical and governance standards. Its cloud computing guidance covers similar ground for hosted services generally. If you are comparing accounting systems more broadly, our guide to accounting software for Hong Kong SMEs sets out the other criteria.
Common mistakes
- Letting staff use personal chatbot accounts with customer lists or payroll files. The company is still the data user, with no contract and no log.
- Auto-posting AI extractions. A misread "8" as "3" goes straight into the ledger and surfaces at audit.
- Asking a chatbot to calculate MPF or final pay. Use the statutory formula in a payroll system, then have a person check it.
- Pasting full HKID or bank account numbers into prompts when the task does not need them.
- No retention rule. Chat histories quietly become a second, unmanaged copy of personal data.
- Treating a confident answer as a correct one. Check any figure you will act on against the ledger.
Frequently asked questions
Is it legal to use ChatGPT or other AI tools with customer data in Hong Kong?
The PDPO does not ban it, but the company remains responsible as the data user. You need a lawful purpose, only the data necessary, adequate security and contractual controls over the provider as a data processor. Consumer accounts with no business terms rarely meet that standard.
What did the PCPD say about AI in 2026?
On 25 August 2026, the PCPD issued guidance on protecting personal data in the use of agentic AI. It sets out nine recommendations, from data minimisation and retention limits to least-privilege access, logging and human approval of high-stakes actions. It confirms the organisation stays accountable for what its AI agents do.
Can AI do my bookkeeping for me?
AI can do the reading, sorting and suggesting: extracting receipts, proposing bank matches and categorising expenses. A person still needs to review entries, make judgements on provisions and tax, and sign off the period. The directors remain responsible for the accounts.
Do I need an AI policy for a small company?
The PCPD recommends one, and it can be short: approved tools, what data may be entered, review requirements, security rules and consequences for breaches. A one-page policy that staff have read is better than a long one nobody opens.
Should employee data ever go into an AI tool?
Only where the purpose is clear, the data is minimal and the tool is approved under contract. Payroll, health and disciplinary data carry the highest risk, and decisions that significantly affect an employee should always have a person in control.
Key takeaways
- AI earns its keep in four back-office jobs: receipt and invoice reading, bank-match suggestions, drafting correspondence and asking questions of your data.
- Keep statutory calculations, tax positions and provisions out of AI's hands as final answers.
- Under the PDPO, your company remains the data user, and AI vendors are data processors that need contracts on retention and security.
- The PCPD's 2024 framework, 2025 employee checklist and August 2026 agentic AI guidance give a practical blueprint.
- A ten-point written policy, human review and a call log make AI use defensible.
Used carefully, AI removes the typing from the back office without removing the judgement. Start free with JetFlow to see receipt reading, reconciliation and ask-your-data questions working on your own books, with every AI call logged.
This article is general information, not legal advice. For your organisation's PDPO obligations, consult the PCPD's guidance or a Hong Kong lawyer.
Run your HK business in one place
JetFlow puts finance, payroll (MPF & IR56), expenses, funding applications and project management in one login — built for Hong Kong SMEs. Start free, import your data, and see it on your own numbers.