Many Hong Kong SMEs now see the same words in system proposals: AI-powered, AI agent, agentic workflow, intelligent automation, AI CRM, AI ERP.
The problem is that a vendor saying "we have AI" does not mean the buyer is getting a business system that can actually work in daily operations. Some products only add a chat box. Some wrap a fixed prompt in a user interface. Some rename traditional automation as AI. The demo may look smart, but after launch the team may still need to copy data, chase status, update records and repair broken handoffs manually.
So before buying an AI system, the better question is not "does it have AI?" The better question is: "What exactly does the AI do inside our workflow?"
1. Direct answer: judge the workflow, not only the demo
For Hong Kong SMEs, the practical way to tell whether an AI system is real or just AI packaging is not to judge how fluent the demo answer sounds. The more important test is whether the AI can read the right business data within controlled permissions, understand the current workflow state, prepare the next step, create tasks or draft content, hand sensitive decisions back to staff for approval, and leave a clear activity record.
If a system only adds a chat interface, cannot explain its data sources, has no role-based permissions, no approval flow, no audit trail, or cannot connect to real CRM, ERP, WhatsApp, forms or email workflows outside the demo, treat it carefully. It may be AI packaging rather than a practical AI-ready business system.
Real AI is not just "sounds human". A useful AI system should help people prepare work inside the right data, permissions and process boundaries.
2. What is fake AI or AI washing?
AI washing means a company claims its product or service uses AI to improve its capability, but either does not actually use AI or exaggerates the role AI plays.
Cornell Wex defines AI washing in simple terms: companies claim to use AI technology to enhance services, but in fact do not. The SEC and FTC have also warned companies against misleading AI-related claims. These are not meant to turn this article into a US legal guide, but they show that exaggerated AI claims are a real market issue.
In the context of Hong Kong SMEs buying software, fake AI does not always mean there is no AI model at all. The more common problem is a gap between the marketing claim and the real workflow capability:
- The product connects an AI model but has no real workflow design.
- The system adds a chat box but cannot read the company's real data.
- The AI only answers fixed FAQs but is marketed as an AI agent.
- The vendor says the system can automate business work, but cannot explain permissions, approvals or responsibility.
- The demo uses a perfect scripted scenario, but the system cannot handle real exceptions.
- The AI generates text, but staff still have to copy, paste, update records and chase follow-up manually.
The point is not to argue over labels. The point is to test whether the marketing claim matches the real workflow.
3. Does using ChatGPT, Gemini or Claude API make a system real AI?
Not necessarily.
A system can use a large language model and still not be a true AI-ready business system. The model is only one layer. What matters for implementation is whether the AI can read the correct data, understand the workflow state, know what it may or may not do, use tools or create tasks when needed, and leave a record after the work is done.
OpenAI's agent guide makes a useful distinction: simple chatbots, single-turn LLM apps or sentiment classifiers that do not control workflow execution are not agents. OpenAI's Agents SDK also describes agents as applications that can plan, call tools, keep state and complete multi-step work.
When a vendor says "we have an AI agent", you do not need to debate terminology first. Ask:
- Can it handle work in steps?
- Can it decide the next step based on data state?
- Can it use tools or trigger tasks?
- Does it know when to stop and hand control back to staff?
- Does it keep enough state and records for the team to trace the process?
If those answers are unclear, the system may be an AI chat feature rather than an AI agent system.
4. Seven signs of a practical AI-ready system
This is not a technical audit standard. It is a practical buyer checklist for owners and operations teams.
| Evaluation point | What a real AI-ready system should show | Common AI packaging problem |
| Data source | It can explain what data AI reads and how fresh that data is | It says AI learns company data but cannot define boundaries |
| Permissions | Roles and AI actions have clear permission limits | Everyone shares the same access and AI boundaries are unclear |
| Workflow state | It knows where an enquiry, customer, order or task currently stands | It only answers questions and does not know process state |
| Task support | It can summarize, draft, create tasks or prepare the next step | It only generates text and staff still move data manually |
| Human approval | Sensitive actions go back to staff for approval | High-risk actions are described as fully automatic |
| Audit trail | It records AI suggestions, human edits and final versions | When something goes wrong, nobody can trace what changed |
| Pilot validation | It can be tested on a real workflow | It only shows a preset demo and avoids real examples |
If a system cannot answer most of these points clearly, the buyer should not give it extra credit just because the proposal uses AI language.
5. What should a vendor demo show?
Many AI demos are too clean.
The customer data is complete. The question is clear. The process is straight. There are no exceptions, complaints, discounts or missing fields. That kind of demo can show the interface, but it cannot prove that the system can handle your daily operations.
Ask the vendor to demonstrate one realistic, low-risk workflow:
1. A new enquiry comes from a website form or WhatsApp. 2. The system creates or updates a CRM record. 3. The AI summarizes the customer need and missing information. 4. The AI drafts the next reply or creates a follow-up task. 5. Staff review, edit and approve. 6. The system records the final version and next owner.
This workflow does not need to be complex on day one. It needs to be real. You should see how the AI reads data, handles missing information, marks uncertainty, hands control back to people and records the final action.
If the vendor can only show a polished chatbot answer, but cannot show data, tasks, approvals and records, the system may not be ready for business operations.
6. Twelve questions to ask when a vendor says it has an AI agent
Use these questions in the vendor meeting:
1. What data can the AI read, and can access be limited by module, field or role? 2. What data is the AI not allowed to read? 3. What data can the AI write or update? 4. Which actions can only be drafted, not sent automatically? 5. Which actions always require human approval? 6. Does the system record AI suggestions, human edits and final versions? 7. If the AI gives a wrong answer, how can we trace the source and correct it? 8. Does the AI know workflow state, such as enquiry, quotation, order, class makeup or payment status? 9. Can the demo use one real workflow from our company? 10. After launch, who maintains prompts, data, permissions and workflow rules? 11. If we change vendors later, can we export data and activity records? 12. What is the success criterion for the first pilot?
A good vendor may not have a perfect answer to every question immediately, but should be willing to define scope, risk, responsibility and acceptance criteria. If the conversation keeps returning to "our AI is very powerful" without specifics, be careful.
7. AI system claims that need extra caution
First, be careful when a vendor promises full automation from the beginning. For SMEs, complaints, refunds, discounts, payments, deleting records, changing contracts, formal commitments and high-value new leads should not be handed to AI for automatic decisions at the start. A safer design is: AI prepares, staff review, the system records, then the workflow continues.
Second, be careful when a vendor says AI can learn all company data but cannot explain boundaries. What AI can and cannot read should be definable. For customer records, pricing, payments, staff data and operational records, permission boundaries matter more than how smart the AI sounds.
Third, be careful when the demo cannot handle exceptions. Real business is not clean. Customers ask incomplete questions. Data is missing. Stock can be wrong. Quotations have exceptions. Parents or customers may complain. A useful AI system should mark uncertainty and hand exceptions back to staff.
Fourth, be careful when there is no audit trail. If the system does not record what AI suggested, what staff changed and what was finally sent or executed, the company cannot learn from errors or manage responsibility.
Fifth, be careful when the price is low but the scope is vague. AI system cost is not only model cost or monthly subscription. Data preparation, permissions, workflow design, training, UAT, post-launch changes and maintenance all affect the real cost.
8. How should a Hong Kong SME validate the first AI pilot?
The most practical way to separate real AI from AI packaging is to run a small, real pilot.
The first workflow should be repetitive, relatively low-risk, based on clear data, reviewable by staff and measurable.
Good first-pilot examples include:
- Enquiry classification and summary.
- Quotation draft preparation.
- Customer follow-up reminders.
- Internal knowledge search.
- Education centre leave or makeup-class data organization.
- Inventory, order or payment exception alerts.
Do not judge the pilot only by whether the AI answer looks polished. Check:
- Did the AI read the correct data?
- Can staff see the data source?
- Do sensitive actions require approval?
- Is the draft useful enough for staff to edit?
- Does the system record AI suggestions and human edits?
- Did it reduce missed follow-ups, missing information or repeated manual work?
- Do staff know when to trust, edit or stop the AI?
If a small workflow cannot be validated, do not rush AI into higher-risk processes.
9. Does real AI mean the company no longer needs people?
No.
For Hong Kong SMEs, the healthier role for AI is not to replace all staff. It is to multiply the stability and output of people.
AI can help with lookup, organization, drafting, reminders, task preparation and exception spotting. People still remain responsible for sales judgment, customer trust, exception handling, approvals and business commitments.
A practical first-stage pattern is:
AI prepares -> staff review -> system records -> workflow continues
If an AI system encourages you to remove human approval from high-value customers or sensitive decisions from day one, the buyer should be more cautious, not more excited.
10. How oneflash looks at real AI systems
oneflash does not treat an AI system as a single chatbot or a standalone AI agent. For Hong Kong SMEs, the more important layer is the business system foundation where people and AI can work together.
That foundation includes data, workflow, permissions, approval, audit trail, CRM, ERP, WhatsApp, forms, email and task handoff. The value of AI is not to perform on the side. It is to help people prepare work inside controlled workflows.
Before buying an AI system, the first step is not necessarily choosing the strongest model or counting how many agents the product has. A more practical starting point is to choose one real workflow and ask:
- Where does the data come from?
- Who follows up?
- What can AI read?
- What can AI draft?
- Which step requires approval?
- How is the final action recorded?
A system that can be evaluated with these questions is closer to a practical AI-ready business system.
11. Further reading
- Cornell Wex: AI washing
- SEC: false and misleading statements about AI
- FTC: Keep your AI claims in check
- OpenAI: A practical guide to building agents
- OpenAI API Docs: Agents SDK
- Microsoft: Copilot and AI agents
- IBM: What is agentic AI?
- Google Cloud: What are AI agents?