What is AI-ready data?
AI-ready data is data you can find, that is reliable enough for the specific job the AI will do, that has an accountable owner, that you are legally allowed to use, and whose origin you can trace. No dataset is ready in general: data that is good enough for an assistant answering internal policy questions may be nowhere near good enough for a revenue forecasting model.
Gartner states this plainly: there is no way to make data AI-ready in general or in advance, because readiness depends on how the data will be used. Gartner also notes that data judged high quality by traditional reporting standards does not automatically equal AI-ready data, since some use cases need the outliers and edge cases that reflect reality.
So the checklist below always starts with one question: what exactly will the AI do? Answer that first, then score the data.
Why data decides AI projects more than the model does
Today's large language models are capable, but they know nothing about your customers, orders or internal processes. Every piece of business-specific value comes from the data you provide: documents for a knowledge assistant, transaction history for analysis, CRM records for a customer-service AI Agent.
Gartner predicts that through 2026, organizations will abandon 60% of AI projects unsupported by AI-ready data. That is an analyst prediction, not a measurement of Vietnamese companies, but it matches a practical lesson: when the data has problems, projects tend to stall at the pilot stage.
For retrieval-augmented generation (RAG), answer quality is tied directly to the quality of the document store: outdated, duplicated or contradictory documents produce wrong answers no matter how good the model is. See what RAG is for how this works.
The six dimensions of AI data readiness
EYC's checklist groups the requirements into six dimensions. The quality dimension uses the six data quality dimensions adopted by the UK Government Data Quality Framework, as defined by DAMA UK; the other dimensions cover governance, lineage and legal requirements.
1. Availability
Does the data exist in digital form, where does it live, and can the AI system reach it without someone copying it by hand? Data in personal files, photos of notebooks or scattered chat messages counts as unavailable, even if in theory you "have" it.
2. Quality
Quality is measured against the purpose. The six common dimensions:
- Completeness: the important records and fields are present.
- Uniqueness: each customer or product has exactly one record.
- Consistency: the same fact does not contradict itself across systems.
- Timeliness: data is updated fast enough for the intended use.
- Validity: values match the expected format and range (dates are dates, phone fields have the right number of digits).
- Accuracy: the data matches reality.
3. Ownership
Every data domain needs a named person accountable for its definitions, its quality and who may use it. Without an owner, nobody fixes errors and arguments over which number is right never end. Owner and steward roles are covered in data governance basics.
4. Access rights
Who, and which systems, may read or write which data? For AI Agents this matters even more, because an agent can act across several systems. The safe principle is least privilege by role, starting with read-only access, and never a shared admin account.
5. Lineage
Do you know where the data came from, which transformations it went through and which version is in use? Gartner lists lineage and versioning among the requirements for AI-ready data. When the AI gets something wrong, lineage tells you whether the fault lies in the data or in the model.
6. Legal basis
If the data contains personal information, you need to know on what basis and for what purpose you may process it. Vietnam's Law on Personal Data Protection No. 91/2025/QH15 took effect on 1 January 2026 (Article 38). The section below summarizes the provisions most relevant to AI.
Scoring table: how ready is your data for AI?
Score each dimension from 0 to 3 for the specific dataset behind one AI use case (not for "all company data"). The maximum is 18. Have the business owner and the technical lead score together, because they usually see different problems.
| Dimension | Check question | Score 0 when… | Score 3 when… |
|---|---|---|---|
| Availability | Is the data the use case needs digital and retrievable automatically? | It lives in people's heads, on paper or in chats | It sits in a system with an API or a data store and can be pulled on schedule |
| Quality | Complete, unique, consistent, timely, valid and accurate enough for this use case? | Never measured; duplicates and contradictions are common | Validation rules exist, quality is measured over time, errors are fixed at source |
| Ownership | Does each data domain have a named accountable person? | Nobody is accountable | Owner and steward are named, definitions are agreed |
| Access rights | Is read/write access role-based and logged? | Shared accounts; everyone can see everything | Role-based, least privilege, access is logged |
| Lineage | Can you trace source, transformations and versions? | Nobody knows where the numbers come from | Data flows are documented and versioned |
| Legal basis | Is the legal basis and purpose identified for personal data? | Nobody knows which fields are personal data | Classified, a basis per purpose, required records kept |
Reading your score
Note: a zero on any single dimension, especially legal basis or access rights, should be treated as a blocker even if the total is high.
0–6: not ready to feed into AI
Prioritize digitizing and consolidating data in one place, naming owners and identifying personal data. Meanwhile you can use AI for work that does not depend on internal data (drafting, summarizing public documents).
7–12: limited pilot
Pick a narrow use case, use a cleaned dataset, keep a human approving outputs and grant read-only access. Measure whether errors come from the data or from the model.
13–18: ready to scale
Set up continuous data quality monitoring, regular testing and a process for when source systems change. Gartner stresses that AI-ready data is not "one and done" but a practice that needs continuous improvement.
Typical data problems in small and mid-sized businesses
The situations below are general descriptions, not based on a survey or specific clients, but many operations teams will recognize them:
- Spreadsheet sprawl: each department, sometimes each employee, keeps its own copy of the customer list or price list; nobody is sure which file is current.
- Records that only exist in chats: orders, customer requests and delivery confirmations are agreed on Zalo or personal messaging and never written back to a system. When an employee leaves, the history leaves with them.
- No master data: the same customer is spelled several ways, the same product has several codes. AI will treat them as different entities.
- Outdated internal documents: processes and policies exist in several versions with no effective date, so an AI assistant may quote one that no longer applies.
- Systems that do not talk to each other: sales, accounting and CRM use different codes and need manual reconciliation.
- Unclear personal data: customer and employee information is mixed into shared files and nobody knows what consent it was collected under.
Illustrative example
A distribution company wants an AI Agent to answer customers about order status. When scored, availability gets 1 because many orders are closed in chat, and ownership gets 0 because nobody owns the customer list. The sensible conclusion is not "pick a better AI model" but to put all orders into one system, standardize customer codes and only then pilot. This is a hypothetical example to show how to use the scoring table.
Legal basis: provisions to know before putting personal data into AI
Under Vietnam's Law on Personal Data Protection No. 91/2025/QH15, several provisions bear directly on preparing data for AI:
- Article 3: personal data may only be collected and processed within a specific, clear scope and purpose; it must be kept accurate and stored only as long as suits the purpose.
- Article 9: consent must be voluntary and informed about data types and purpose, and given per purpose; silence or non-response is not consent. Data collected for sales does not automatically become usable for another purpose.
- Article 19: lists cases where processing does not require consent (for example performing an agreement with the data subject in accordance with law) and requires a supervision mechanism when relying on them.
- Article 30: personal data in AI, big data and cloud environments must be processed for the right purpose and only as far as necessary; systems must include security measures, appropriate authentication and access control; AI processing must be classified by risk level.
- Article 38: the law took effect on 1 January 2026; small businesses and startups may choose not to apply certain obligations (such as the impact assessments in Articles 21 and 22) for five years, except in cases such as providing personal data processing services or processing sensitive data.
What "legal basis" means in this checklist
At minimum: a list of the personal data fields in the dataset; the processing purpose for each group; the basis for processing (consent or an Article 19 case); and a decision on whether to de-identify or drop unnecessary fields before they reach the AI. Our article on Vietnam's Personal Data Protection Law 2025 goes deeper into each obligation.
Quick wins you can start now
You do not need a large project to raise your score. A small team can usually do the following:
- Write a one-page data inventory: which sources, where they live, who owns them, whether they hold personal data.
- Pick one source of truth for customers and products; mark every other copy as reference only.
- Require orders and customer requests to be logged in the system; chat is for conversation only.
- Add validation at the point of entry: date formats, customer codes, mandatory fields.
- Give every policy document an effective date and an approver before it goes into an AI assistant.
- Review shared-folder permissions; remove "anyone with the link" access from files containing personal data.
- Re-score after each improvement so progress is visible.
When to get outside help
When data is spread across systems that need integrating, when you need a data store or pipelines, or when you need to design permissions for AI Agents. EYC works on cloud and data infrastructure and systems integration; the exact scope is agreed under contract.
Frequently asked questions
Does data for AI have to be perfect?
No. It has to be good enough for the specific use case. The UK Government Data Quality Framework describes quality as "fitness for purpose", and Gartner notes that some AI use cases need unusual data too, so that the model reflects reality.
Does a small business need a data checklist before using AI?
Yes, at small scale: score the six dimensions for one use case in roughly one meeting between the business owner and the technical lead. It stops you investing in AI tools when the real problem is the data.
Can spreadsheet data be used for AI?
Yes, if there is one authoritative file with a stable column structure, an owner and regular updates. The problem is not spreadsheets; it is multiple contradictory copies with nobody accountable.
Can we put customer data into an AI tool?
Only with a legal basis that fits the purpose under Law 91/2025/QH15 (Articles 9 and 19), with security and access controls in place (Article 30), and after considering dropping or de-identifying fields you do not need. This is information only, not legal advice.
How is AI data readiness different from data governance?
Data governance is the operating model that manages data over time (roles, rules, processes). AI readiness is an assessment of whether one dataset is good enough for one AI use case. Good governance makes high readiness scores achievable and durable.
How often should we re-assess data readiness?
Whenever the AI use case, a data source or a system changes, and periodically once the AI is in production. Gartner treats AI-ready data as an ongoing practice, not a one-time task.