Key Takeaways
- AI vendors increasingly sell access to governed data, not just smarter models. The terms behind that pitch decide whether answers can be trusted.
- Seven terms cover most board-level questions: governed data, data catalogue, permission-aware retrieval, AI gateway, data lineage, data classification and data residency.
- Gartner predicts organisations will abandon 60% of AI projects unsupported by AI-ready data through 2026.
- Use the glossary as a checklist: if a vendor cannot explain how it handles each term, the pilot is not ready to scale.
By the end of this glossary, you will have working definitions of the seven data terms that now sit underneath every serious enterprise AI proposal, and one question to ask a vendor about each.
The timing matters. On 13 November 2026, Databricks and OpenAI described how ChatGPT, Codex and GPT models now work on governed company data, with each user seeing only what they are already allowed to see. OpenAI's Greg Brockman put it plainly: AI without company data is like hiring someone with 20 PhDs and never inviting them to a meeting. The real question for a COO is not which model is smartest. It is whether the data rules underneath are strong enough to let that model into the meeting.
What is governed data?
Governed data is company data that has an owner, an agreed definition, access rules and a quality standard, all enforced by systems rather than by goodwill. For AI, it means a model can only read data that is documented, permissioned and current, so its answers match the numbers your finance team already reports.
What it means for your organisation
Governed data is the difference between an AI answer you can put in a board paper and one you must re-check by hand. Gartner warned in February 2025 that organisations will abandon 60% of AI projects unsupported by AI-ready data through 2026, and governance is most of what makes data AI-ready.
Example
A logistics group asks an AI assistant for on-time delivery by region. Without governance, it blends two versions of the delivery table. With governance, only the certified table is visible, and the answer matches the monthly operations report.
What is a data catalogue?
A data catalogue is the searchable inventory of an organisation's data: what each dataset contains, who owns it, how fresh it is and who may use it. AI tools increasingly read the catalogue first, using it as the map that tells them which tables and documents to trust for a given question.
What it means for your organisation
If you cannot list your critical datasets and their owners, an AI tool cannot either. A catalogue also turns the vague question of whether AI can use our data into a concrete list of approved sources, which is far easier to defend to an auditor.
Example
A property management firm catalogues its tenancy, maintenance and billing data with named owners. When it pilots an AI helpdesk, the project team approves three catalogued sources in one meeting instead of debating spreadsheets for six weeks.
What is permission-aware retrieval?
Permission-aware retrieval means an AI system checks the user's existing access rights before it fetches any document or record to answer a question. Two employees can ask the same question and receive different answers, because the AI only draws on what each person is already allowed to see.
What it means for your organisation
This is the single most important control when AI is connected to shared drives, email or HR systems. Without it, an AI assistant becomes a fast way around your access model, surfacing salary files or client records to anyone who asks the right question.
Example
In a professional services firm, a junior associate and a partner both ask for the fee history of a key client. The partner sees the full billing record; the associate receives only the matters he is staffed on.
What is an AI gateway?
An AI gateway is a control layer that sits between your staff, applications and AI models. Every request passes through it, so the organisation can log usage, set budgets by team, keep vendor API keys out of employees' hands, block risky content and switch models without rebuilding each application.
What it means for your organisation
A gateway answers three questions boards now ask: who is using AI, what is it costing, and where is our data going. Databricks, for example, routes Codex requests through its Unity Gateway so that the OpenAI key stays in the gateway and every request is logged.
Example
A retail chain gives five departments access to two different AI models through one gateway. Finance sees spend by department each week, and IT can cut off a misbehaving integration in minutes without touching the other four.
What is data lineage?
Data lineage is the recorded path a piece of data takes from its source system, through every transformation, to the report or AI answer where it appears. For AI, lineage lets you trace any figure in an output back to the exact table, document and version it came from.
What it means for your organisation
Lineage turns an AI error from a mystery into a fixable defect. It also supports regulators' growing interest in traceability: the Privacy Commissioner's August 2026 guidance on agentic AI expects organisations to keep logs so that what an AI agent did with personal data can be reconstructed.
Example
An insurer's AI summary shows a claims ratio that looks wrong. Lineage shows the figure came from a staging table refreshed before month-end adjustments, so the data team fixes the source rather than retraining staff to distrust the tool.
What is data classification?
Data classification is the practice of labelling data by sensitivity, typically public, internal, confidential and restricted, and attaching handling rules to each label. For AI, the labels decide which data may be sent to an external model, which must stay on private infrastructure and which must never be used at all.
What it means for your organisation
Classification is how a broad AI policy becomes enforceable. Instead of telling staff to be careful with sensitive data, systems can block restricted files from reaching a public model automatically, and auditors can see the rule as well as the intent.
Example
A healthcare administrator labels patient records as restricted and appointment schedules as internal. Its AI scheduling assistant may read the schedules, while anything carrying a restricted label is excluded by default.
What is data residency?
Data residency is the physical and legal location where data is stored and processed, including the servers an AI model uses when it reads your information. It decides which jurisdiction's laws apply, which regulators can request access and what you must tell clients about where their data goes.
What it means for your organisation
Hong Kong's PDPO cross-border transfer provision, section 33, is still not in force, yet clients, banks and sector regulators routinely ask where data sits. For many enterprises, local or private-cloud hosting is less a legal requirement than a condition of winning and keeping contracts.
Example
A wealth management firm evaluates two AI vendors with similar features. One processes data offshore; the other keeps it in Hong Kong. The firm's largest corporate client has a residency clause, so the decision is made before the demo ends.
Bring the glossary into your next AI review
Models will keep improving every quarter. The organisations that capture the value will be the ones whose data rules let those models work safely on real business questions. Deloitte's 2026 State of AI in the Enterprise found that only a quarter of organisations have moved 40% or more of their AI pilots into production, and weak data foundations are a common reason pilots stall.
Use these seven terms as a checklist in your next vendor meeting or steering committee. Ask who owns the data, how permissions carry through, where requests are logged, how answers are traced, which labels block which data and where it all physically sits. Clear answers mean you are ready to scale; vague ones mean you have found your next workstream.
Empowering AI. Centered on you.
Put your first AI employee on governed data
Start with one role and one process, with your data stored in Hong Kong and $0 setup fee. Talk to our team about where to begin.
Frequently Asked Questions
Do we need a full data governance programme before starting an AI pilot?
No. Govern the specific datasets the pilot needs: name an owner, agree the definitions, set permissions and log access. Expand governance as the AI use case expands.
Is governed data the same as data security?
Not quite. Security protects data from unauthorised access, while governance also covers ownership, definitions, quality and approved use. AI needs both to produce answers you can trust.
Does the PDPO require AI data to stay in Hong Kong?
Not currently, because section 33 of the PDPO on cross-border transfers is not in force. Contracts and sector regulators such as the HKMA may still impose residency expectations.
Who should own AI data governance in an enterprise?
Business data owners should own definitions and access decisions, with IT running the controls and compliance setting policy. Leaving it entirely to IT usually leaves definitions unresolved.
How long does it take to make data AI-ready?
It depends on scope. Preparing a few well-defined datasets for one use case can take weeks, while organisation-wide cataloguing and classification is usually a multi-quarter effort.