AI Consulting in Hong Kong: Six Questions Before You Buy

A growing number of Hong Kong SMEs are adopting AI. Some have purchased ChatGPT Plus or Claude subscriptions. Some have reviewed vendor proposals. Some have completed initial pilots. But production-grade AI systems that are fully deployed and operating stably remain the exception.

The gap is rarely a technical problem. It is usually a procurement decision problem.

The market for AI consulting services is expanding quickly. Some vendors resell tool licences. Some can deliver working custom systems. Many fall somewhere in between. In this market, proposal language has converged — "intelligent AI," "RAG," "automated workflows" — making evaluation difficult.

The six questions below are not a technical audit checklist. They are a decision framework for procurement. Any business owner can use them before engaging a vendor.

1. Can they analyse and take over your business processes?

AI consulting is not software installation. A system that produces real business impact must be built on a clear understanding of existing workflows — where documents come from, who handles them, what format the output takes, and where it goes next.

If a vendor leads with a demo in the first meeting and treats process analysis as secondary, they are usually selling a generic solution rather than building something for your business.

A verifiable question to ask: "Can you walk me through each step of our business process after AI integration?" A vendor who can answer clearly has the foundation to design a working system. The difference between delivery capability and advisory capability is examined in detail in Why Enterprises Need an AI Consultant, Not Just Claude or Poe.

2. Who owns the AI memory and conversation data?

Over time, enterprises accumulate significant volumes of conversation data, prompt designs, and business context through AI tool usage. Where is that data stored? Who owns it?

Major AI platforms (ChatGPT, Claude, Gemini) store conversation history on platform servers by default. Memory is typically not shared across conversations automatically. The context an enterprise builds up in an AI system often cannot be taken away or migrated.

When procuring AI consulting services, these questions need answers: On whose infrastructure is the memory layer designed? If the relationship ends, can that data be fully migrated? What is the switching cost of changing vendors in future?

Custom systems can deploy the memory layer on enterprise-owned infrastructure, making accumulated context a company asset rather than a platform asset. The architectural design is examined in HKSoka's Memory System: An Engineering Breakdown.

3. How will internal documents and knowledge bases be integrated?

Most enterprise knowledge lives in PDF reports, internal manuals, email records, or instant messaging conversations. An AI system that cannot access these data sources is substantially limited in practical use.

Retrieval-Augmented Generation (RAG) is the current mainstream integration approach — when answering questions, the AI retrieves relevant content from the enterprise document library in real time, rather than relying on the static knowledge from model training.

But RAG system quality varies enormously. Questions to clarify include: Can the system handle mixed Chinese and English documents? When documents are updated, how quickly does the system reflect the latest version? When multiple documents are queried simultaneously, how is relevance ranked?

Hong Kong enterprise documents commonly mix Traditional Chinese and English. This has a meaningful impact on RAG retrieval accuracy. The specific challenges and verification methods are covered in The Practical Limits of Bilingual RAG.

4. Can it connect to existing systems and workflows?

The practical value of an AI system often depends on integration depth. An AI assistant that only operates within a standalone interface — requiring staff to switch tools actively to use it — tends to see low adoption rates.

Integration may involve: internal ERP or CRM systems, instant messaging platforms (WhatsApp Business, etc.), email workflows, Google Workspace or Microsoft 365 document libraries, existing approval or reporting processes.

Vendors should be able to describe the technical implementation of the integration specifically, not merely commit to "we can integrate." Agent architecture and fixed pipelines differ significantly in integration depth and cost. Different use cases suit different designs; there is no universally optimal choice. The relevant analysis is in AI Agent or Fixed Pipeline: Choosing the Right Architecture for Your Use Case.

5. How will you measure usage and return on investment?

After an AI system goes live, how do you confirm staff are using it? Which features see the highest usage? Which query types most often trigger errors or low-confidence responses?

An AI system without usage tracking cannot be meaningfully optimised over time, and cannot provide management with objective data on return on investment.

Before procurement, confirm: Does the system have built-in usage logs and analytics? Can errors or low-confidence responses be identified and recorded? Is there a regular performance evaluation mechanism?

Large language models have a known limitation called hallucination — generating plausible-sounding but factually incorrect information. Production-grade systems manage this risk through retrieval grounding, confidence boundary design, and audit logs. The relevant design principles are in When LLMs Get It Wrong: Accountability and Validation Design in Production AI Systems.

6. How is access control designed in a multi-user environment?

Enterprise AI systems typically serve multiple departments or roles, with different data access scope and functional requirements for different users.

An HR AI assistant should not be able to access financial data. Junior staff queries should not reach executive decision documents. These restrictions need to be implemented at the system architecture level, not left to user self-discipline or after-the-fact management.

Vendors should be able to explain specifically: How are user permissions defined and managed at the technical level? How is departmental isolation implemented? What is the process for changing permissions after an employee leaves or changes roles? Can administrators manage these settings independently, without needing to involve the vendor each time?

Decision framework: the practical differences between three procurement paths

Hong Kong SMEs typically have three paths for AI procurement, each with its own suitable scenarios.

Tool subscriptions (ChatGPT Plus, Claude Pro, and similar monthly services): suitable for individual use or initial exploration. Low startup cost, no technical team required. Memory cannot be migrated across platforms, integration capability is limited, and business process customisation is not available.

AI consulting engagement (commissioning a consultant or engineer to design a custom system): suitable for enterprises with a defined business problem that needs a systematic solution. The key evaluation criterion is whether the vendor can deliver a working system, not just a proposal document or strategy report.

Building in-house: requires internal engineering capability or a long-term technical partner. Suitable for scenarios with highly specialised business processes or very high data sensitivity. Higher upfront investment; the system is fully owned by the enterprise.

No path is universally superior. The right choice depends on the combination of business scale, data sensitivity, budget structure, and internal technical capability. The Six Functional Gaps Between an AI Consultant and Directly Using an LLM Platform is a useful starting point for clarifying your own requirements.

Levi is an independent AI engineer based in Hong Kong, building production-grade LLM applications and RAG pipelines for enterprises in Hong Kong and the Greater Bay Area. Academic background includes a Master's in Business Analytics and AI from Vlerick Business School, Belgium. Production deployments include an insurance industry RAG comparison system and HKSoka (hksoka.com), a multi-layer memory AI platform. Past work has also been applied in client projects (under NDA).

For an initial conversation about AI procurement or consulting engagement:
smartai.hk+ai.consulting@proton.me
linkedin.com/in/levi-innovation