Back to featured insights

Generative AI in Custom Software Development: 8 Use Cases for Hong Kong Enterprises

Generative AI is no longer a pilot project in Hong Kong. The real question is which feature to build into your software first. Explore 8 proven use cases, from smarter e-shops to AI document processing, and find the right starting point for your business.

Table of Content

Share

Generative AI has moved from pilots to production in Hong Kong: 23% of companies in Hong Kong and the mainland already run AI with measurable financial impact (Deloitte-HKU AI Adoption Index 2026).

The real question is which feature to build first. Here are eight that Hong Kong enterprises are already building, and what to check before you start.

8 generative AI use cases Hong Kong enterprises are building

Each use case below is a feature built into software a company already runs. AI-assisted coding, where developers use AI to write the software itself, is a different subject; we cover it in AI-Assisted Software Development: A 5-Stage Workflow.

1. Customer service chat connected to orders, bookings or CRM

Unlike a scripted FAQ bot, it answers from live account data and hands over to an agent when needed. Cathay Pacific launched conversational AI with Hong Kong’s Fano Labs and expects chatbot training cycles to fall from 4-6 weeks to 1-2 weeks (Cathay).

Customer asks What the bot uses and its guardrail
“Where is my order?” Data: Order and delivery statusGuardrail: Logged-in customer’s records only
“Can I change my booking?” Data: Booking system, fare rulesGuardrail: Customer confirms before any change
A complaint or complex case Data: Conversation historyGuardrail: Handover to a human agent

The model connects via API to your order, booking or CRM systems, in Cantonese, English and Traditional Chinese.

On privacy, the bot sees only the logged-in customer’s data, and customers know it’s AI. Monitor answers after launch, as the HKMA advises for customer-facing chatbots (HKMA). Cost rises with chat volume.

2. AI site search and product Q&A

Shoppers type what they need, not your product names:

Shopper types Keyword search returns AI search returns
waterproof hiking shoes under HK$800 ▲Pages containing “hiking” or “shoes” ✓Waterproof hiking shoes, filtered by price
gift for a 5-year-old who likes science ▲Little or nothing ✓Age-appropriate science toys and kits
Does this kettle work on an induction hob? ▲No answer ✓A short answer from the product specs

HKTDC added AI product search to hktdc.com Sourcing (HKTDC). YOHO launched a YOHO AI assistant as a shopping concierge (HKUST Business School).

Products, specs and FAQs are indexed by meaning, and results flow into your existing product pages and cart. Answers must come only from catalogue data. Cost depends on catalogue size and search volume.

3. Bilingual product content for e-commerce

TVCMALL launched a GenAI translation and content solution in one month, cutting translation costs by 40% (AWS).

A model applies your brand rules and sends drafts for review before publishing. Privacy risk is low; the real risk is claims the product can’t back up. Cost is mostly one-off per catalogue. See our e-commerce web development and WooCommerce services.

What goes in What comes out, for review
NameSpecsMaterialsSize English and Traditional Chinese descriptions
CategoryTarget keywords SEO title and meta description
Product images Image alt text

4. Document processing for invoices, contracts and KYC forms

GenAI reads documents in English and Traditional Chinese and extracts the fields your system needs. In the HKMA’s GenAI Sandbox, one bank cut memo document processing from one day to about five minutes (HKMA).

  1. Capture from email, scanner or upload portal.
  2. Extract with OCR plus a model, via API or private for sensitive files.
  3. Validate against master data such as supplier lists and PO numbers.
  4. Review low-confidence fields with a person.
  5. Post approved data into the ERP.

Extract only the personal data you need, in line with the PCPD’s AI Model Personal Data Protection Framework. Test on Chinese and handwritten samples. Cost grows with page volume.

5. Internal knowledge assistant over policies, SOPs and project files

Staff ask in plain language and get an answer linked to the source document. It uses retrieval-augmented generation (RAG): search your files first, then answer only from what was found.

Hong Kong example What it covers
Citi AssistFor Citi’s Hong Kong staff · CDO Magazine HRComplianceFinanceRisk policies
EMSD ChatbotHKSAR Government · EMSD Authorised internal knowledgeKept up to date
  • Access control: single sign-on applies each user’s rights to every answer. This is the main PDPO risk.
  • Accuracy: test with real staff questions before launch.
  • Cost: driven by document volume, question volume and index updates.

It’s often the safest first project: users are internal and errors never reach customers.

6. Natural-language reporting on ERP and sales data

  • Retail operations manager: “Which stores missed their sales target last month, and by how much?”
  • System: a ranked table and bar chart, plus the query used to build it.

Managers get answers without waiting for a report.

The model should query a curated data model with defined metrics, not raw ERP tables, or a wrong join gives a confident but wrong number. Row-level security keeps each manager to their own data. Most of the cost is preparing the data model. See Arestós Generative BI and Business Intelligence.

7. Sales and proposal drafting inside the CRM

Prudential Hong Kong’s AI Underwriter, built with Alibaba Cloud, gives consultants preliminary underwriting indications at the point of sale (Prudential).

Task Today With GenAI in the CRM
Proposal Written from scratch Drafted from the customer record and approved content
Follow-up email Typed after each meeting Drafted from meeting notes
Call preparation Manual review of past activity One-page account summary

The model uses only approved product documents, and a person signs off every draft. Product claims must stay within approved wording, a regulatory point for insurers and banks. Cost grows with users and drafts.

8. Compliance and contract review support

In the HKMA’s GenAI Sandbox, banks cut preparation time for suspicious transaction reports by 30-80% (HKMA).

Many companies use a private model here, since contracts are confidential, and keep an audit trail. Test how many known risky clauses it catches before relying on it.

AI handles People decide
Flagging clauses that differ from your playbook Whether a deviation is acceptable
Summarising changes between versions The negotiation position
Drafting first-pass compliance reports Final content and filing

Build custom or buy off-the-shelf?

In Hong Kong, the “buy” list is short. Hong Kong is not on OpenAI’s supported list (OpenAI), so ChatGPT Enterprise isn’t an official option. Microsoft serves eligible Hong Kong customers via Azure OpenAI (SCMP).

Factor Microsoft 365 Copilot ChatGPT Enterprise Custom GenAI feature
Available in Hong Kong Yes Not officially Yes, using models available to Hong Kong customers
Best for Staff productivity in email, documents and meetings Staff productivity A specific process in your website, e-shop or business system
Works with your systems Microsoft 365 data; others through connectors Through connectors Built around your e-shop, ERP, CRM or database
Customer-facing use Not designed for it Not designed for it Yes
Pricing US$30 per user per month, paid yearly (Microsoft); Copilot Business US$21 (Microsoft) Not applicable in Hong Kong Build cost plus usage-based model cost
Control over where data goes Microsoft’s regions and terms Not applicable in Hong Kong Your choice of API region, private cloud or on-premise model

Copilot speeds up general office work. A custom feature changes a specific process and can face customers. Many enterprises use both; see Build, Buy, or Staff? Choosing the Right AI Capability Strategy.

How to choose your first GenAI use case

The best first project is rarely the most ambitious one. It is the one most likely to work and to show results within a quarter. Score each idea against five checks:

  1. Data readiness. Do the documents, catalogue or records already exist, and are they reasonably clean?
  2. Volume. Is there enough repetitive work to matter, such as hundreds of documents, chats or products a week?
  3. Risk. Can you start with internal users or low-stakes output, with a person checking results?
  4. Owner. Is there one business owner who will measure results and drive adoption?
  5. Baseline. Do you know today’s number to beat, such as minutes per document or tickets per day?

An idea that passes all five is a strong candidate. For many Hong Kong companies, that means AI site search, product content, document processing or an internal knowledge assistant before a customer-facing chatbot.

Build timeline and cost drivers

A focused GenAI feature can go live in weeks; one inside a regulated workflow takes a few months. Two public examples:

Project Time from start to launch
TVCMALL generative AI translation and content solution 1 month
Prudential Hong Kong AI Underwriter, a regulated insurance workflow Under 3 months

Most projects move through the same five phases:

  1. Discovery and data review: confirm the use case, the baseline and whether the data is usable.
  2. Prototype on real data: test accuracy with your own documents, products or questions.
  3. Integration and security review: connect to the website, e-shop or business system and check access rights.
  4. Pilot with one team: measure results against the baseline.
  5. Rollout and monitoring: extend to more users and keep checking answer quality.

Cost is shaped by six drivers:

  • Model usage: the tokens each request uses, multiplied by the number of requests.
  • Hosting: a model called via API, or a private deployment on your own GPU infrastructure.
  • Data preparation: cleaning and indexing documents, catalogues or data models.
  • Integration: connecting to your website, e-shop, ERP or CRM.
  • Testing and monitoring: accuracy testing in English and Traditional Chinese, before and after launch.
  • Support: updates as your data, products and policies change.

How Arestós builds GenAI features into custom software

At Arestós, we build generative AI features as part of the websites, e-shops and business systems we deliver, so the AI works with your real data and workflows rather than sitting in a separate tool.

On every project, we choose models that are available to Hong Kong customers and fit your data rules, build a human review step into the workflow, and test in both English and Traditional Chinese before launch.

Frequently Asked Questions

1. Can Hong Kong companies use OpenAI models?

Not directly from OpenAI: Hong Kong is not on OpenAI’s list of supported locations (OpenAI). Microsoft serves eligible Hong Kong customers through Azure OpenAI, using models deployed in regions outside Hong Kong (SCMP). Open-weight models can also be deployed privately on your own infrastructure.

2. Is it allowed to send customer data to an AI model under the PDPO?

The PDPO applies to AI processing as it does to any other use of personal data. The PCPD’s AI Model Personal Data Protection Framework guides organisations that procure AI and customise it with personal data, including collecting only the data needed and assessing risk before deployment.

3. Is there official generative AI guidance in Hong Kong?

Yes, though it is not legally binding. The Digital Policy Office published the Hong Kong Generative AI Technical and Application Guideline in April 2025. It covers the scope and limits of GenAI applications and risks such as data leakage, model bias and errors.

4. What is RAG, and why does it matter?

Retrieval-augmented generation (RAG) makes the model answer from your own documents or catalogue instead of its general training data. That reduces made-up answers and lets every answer link back to its source, which is essential for staff assistants, product Q&A and anything customer-facing.

5. Which use case should we start with?

Start with one that has clean data, high volume, low risk and a clear owner. For many Hong Kong companies, that is AI site search, bilingual product content, document processing or an internal knowledge assistant.

Conclusion

The GenAI projects that pay off first sit inside a system you already run, such as your website, e-shop, ERP or CRM. They use data you already have and keep a person in the loop until the results are proven.

In Hong Kong, also plan for bilingual content, the PDPO and local model availability from day one. Getting those right early is what turns a pilot into a feature your team and customers actually use.

Not sure which use case to start with? Book a free consultation and we’ll help you shortlist the one with the clearest return for your business.

Start Innovating Your Business Now! Want to get FREE Consultancy and Quick Estimation?
Reach us at +852 3796 0101 or email us