--- title: "AI-Ready Data: What It Means and How to Test for It" description: "A concrete definition of AI-ready data: one central store, staged cleaning, one governed definition per term, and controlled access for models." canonical: https://shipshapedata.com/resources/ai-foundations/what-ai-ready-data-means/ language: en-GB author: "Emile Van Zyl" author_role: "MarTech Automation Consultant" author_url: https://www.linkedin.com/in/emile-vanzyl-enhancing-your-human-advantage/ date_published: 2026-10-26 date_modified: 2026-10-26 --- # AI-ready data: what it means and how to test it > A concrete definition of AI-ready data: one central store, staged cleaning, one governed definition per term, and controlled access for models. Section: Home > Resources > AI foundations Canonical page: https://shipshapedata.com/resources/ai-foundations/what-ai-ready-data-means/ By Emile Van Zyl, MarTech Automation Consultant (https://www.linkedin.com/in/emile-vanzyl-enhancing-your-human-advantage/) | published 2026-10-26 Many organisations now describe themselves as close to AI-ready, or say they will be once a new platform is in place. The phrase is used loosely. Often it means that the organisation holds a lot of data, which is not the same thing. A large volume of data spread across systems, with conflicting definitions, gives a model more ways to be wrong. AI-ready data is data that a model can use to produce a correct, checkable answer. In practice that means four things: the data sits in one place, it has been cleaned in stages from bronze to silver to gold, every business term has one governed definition, and models reach it through controlled access. This guide is for the person who has been asked whether the organisation is ready for AI and needs a definition they can test against. It stays on the data. The organisational questions, such as skills, ownership and change, are covered in our AI readiness checklist. ## Why AI-ready needs a definition The term is used in several different ways. Hardware makers use it to describe a PC with a chip for running models locally. Training providers use it to mean staff who have completed a course. Inside organisations it most often refers to data, and there it tends to mean volume: years of transactions, a large document store, or a data lake that grew without a plan. Volume is not readiness. A model given more data of uneven quality does not average out the errors. It repeats them with confidence. Two systems that count customers differently will give a model two different answers, and the model has no way to know which one the finance team reports to the board. Published definitions agree on the qualities. The UK government's 2026 guidance on making government datasets ready for AI describes AI-ready data as accurate, complete, consistent, secure and enriched with metadata, so that both people and machines can trust and understand it. Those qualities are correct, but they describe a state. They do not tell you what to build to reach it. The rest of this guide sets out the four things we build, and the order we build them in. > AI-ready is not a quantity of data. It is a set of conditions that let a model give the right answer. ## The test: a correct answer from your data A useful definition needs a test. Ours is simple. Pick a real business question, ask it of your data through a model, and check whether the answer is correct and whether you can show where it came from. Choose a question that a manager asks every month, such as revenue by region last quarter, or the number of open support tickets older than a week. You already know the right answer, because someone produces it by hand. If a model working from your data gives the same figure, and you can trace that figure back to source records, the data that answer depends on is AI-ready. If it gives a different figure, or the right figure for the wrong reason, it is not. ### What the test exposes - Location. The model could not find all the records, because some sit in a system it cannot reach. - Quality. The model found the records, but they contain duplicates, gaps or stale values. - Meaning. The records were clean, but the model used a different definition of the term from the one the business uses. - Access. The model reached data it should not have seen, or could not reach data it needed, because there was no governed route. Each failure points to one of the four conditions below. That is why we treat them as a definition rather than a wish list. A failed test tells you which condition is missing and where to start. ## Centralised: a cloud data warehouse The first condition is that the data a model needs sits in one place. For most organisations that place is a cloud data warehouse, or a lakehouse that behaves like one. Operational systems such as a CRM, an ERP and a ticketing tool are built to run the working day to day. They are not built to answer questions that span all three. A warehouse brings those sources together into one store designed for analysis. Our guide to data warehousing explains how that works and when you need one. We build on Snowflake, Databricks and Google BigQuery, and on Microsoft Fabric where an organisation already runs on Microsoft. The choice of platform matters less than the decision to have one governed store rather than several partial copies. ### Why a model needs one place A model asked a question across scattered systems has two poor options. It can query each source separately and try to join the results itself, which is slow and prone to error. Or it can work from an export someone made last month, which is out of date and ungoverned from the moment it is saved. A central store removes both problems. It also gives you one place to apply security, one place to monitor and one place to check when an answer looks wrong. Centralising does not mean moving everything at once. We usually start with the sources behind one use case and add others as the next use case needs them. This keeps the first piece of work small enough to finish and measure. ## Cleaned in stages: medallion architecture The second condition is that data is cleaned in stages, with each stage kept separate. The pattern we use is medallion architecture, which has three layers: bronze, silver and gold. - Bronze holds raw data as it arrived from the source system, unchanged. It is the record of what the source said, and you keep it so you can reprocess or audit later. - Silver holds data that has been cleaned and conformed: duplicates removed, formats standardised, records matched across systems and quality rules applied. - Gold holds data shaped for use: tables built around business entities and measures, ready for a dashboard, a report or a model. The separation is the point. When a figure in gold looks wrong, you can trace it back through silver to the raw record in bronze and see where the change happened. When a cleaning rule changes, you rerun it over bronze rather than patching gold by hand. ### What cleaning means in practice Cleaning is measured against specific qualities: accuracy, completeness, consistency, timeliness, validity and uniqueness. Our guide to data quality dimensions defines each one and explains how to measure it. The silver layer is where those checks run, so that problems are caught before any model sees them. Records that describe the same customer, product or supplier in different systems need particular care. Matching them into one trusted record is the work of master data management. A model that cannot tell that two records describe the same customer will count that customer twice. ## Agreed: a semantic layer The third condition is that every business term has one agreed definition, held in one place and used by every tool. This is the semantic layer: one governed definition of your business. Most organisations have several definitions of their most important terms. Sales counts revenue when a contract is signed. Finance counts it when it is recognised. Operations counts an active customer as anyone with an open account, while marketing counts anyone who has bought in the last year. People cope with this because they know which team produced a report. A model does not know. It reads the column called revenue and uses it. A semantic layer settles each definition once. It states what revenue means, which tables it comes from, which filters apply and who owns the definition. Dashboards, reports and models then all draw on the same logic, so a question asked of a model and the same question asked of a dashboard return the same figure. ### Where the definitions start The work begins with a written record of what each field and measure means. A data dictionary is the usual starting point. The semantic layer turns those written definitions into logic that systems can run, so the definition and the calculation cannot drift apart over time. > If two teams would give different answers to the same question, a model will give both of them. ## Reachable safely: governed access The fourth condition is that models reach the data through a controlled route. Clean, agreed data is still not AI-ready if the only way a model can use it is through a copy pasted into a prompt or an export loaded into a separate tool. That copy has no access controls, no audit trail and no updates. Governed access means three things. - Access controls. A model, and the person using it, can see only what that person is allowed to see. A sales assistant should not return payroll data because the underlying table happened to be reachable. - Business logic. The model works through the definitions in the semantic layer rather than writing its own calculations against raw tables. - Lineage. Every answer can be traced to the data it used, so you can check it and explain it. One way to provide this route is Model Context Protocol (MCP), an open protocol that lets models call approved tools and data sources through a defined interface. The model asks for what it needs, the server applies the rules, and the request is logged. We design and build these routes as part of our MCP connections work. When this route is in place, a model can answer from an organisation's own content with confidence. At Slimstock, a custom AI chat interface answering from the company's own content provides 24/7 support in 11 languages. That result depends on the content being current and the route to it being governed. ## When data is not AI-ready When the four conditions are missing, the failures follow a pattern. They are rarely a problem with the model itself. - Confident wrong answers. A model working from inconsistent data produces fluent, plausible figures that are incorrect. Our guide to AI hallucination covers how ungrounded answers arise. - Pilots that do not scale. A demonstration works on a hand-picked extract and fails when it meets the full, uncleaned data. - Loss of trust. Once one wrong figure reaches a senior meeting, people stop using the tool, and the next AI proposal faces more doubt. - Exposure. A model with broad, ungoverned access returns information to people who should not see it, which creates a data protection problem as well as a technical one. AI works as a magnifying glass. Over strong foundations it amplifies the gains. Over weak foundations it accelerates failure and exposes risk that was already there. Many of the recurring causes in our guide to why AI projects fail trace back to data that was never made ready. ### Where to start You do not need every source in the organisation to be AI-ready before you build anything. Choose one use case with a clear metric. Identify the data it depends on. Make that data meet the four conditions, build the use case on top, and measure the result. The first win then pays for the second, and each new use case reuses the foundation already in place. Sometimes the honest answer is that a use case should wait. If the data it needs does not exist, or cannot be governed, building the AI first will produce a tool that nobody trusts. We would rather say so at the start than after the budget is spent. ## Where this leaves you AI-ready describes the condition of your data, not its size. Data is AI-ready when it sits in one place, has been cleaned in stages from bronze to silver to gold, carries one governed definition of every business term, and can be reached by models through a controlled route. The test is whether a model can give the correct answer to a real business question and show where that answer came from. If you want a quick view of where you stand, our free AI readiness assessment has 16 questions, scores strategy, people, data and change on the spot, and asks for no email address. It will show whether data is the first gap to close. Our data foundation service builds the four conditions described in this guide: the cloud data warehouse, medallion architecture, the semantic layer and the governed route that lets models reach the data. We scope the work around one use case, tie it to a real metric and stick around past go-live. Talk to us if you would like to test one of your own business questions against your data. We reply personally and will say if we are not the right fit. ## Frequently asked questions AI-ready data is data a model can use to give a correct answer that you can trace back to its source. It sits in one central store, has been cleaned in stages from raw to conformed to business-ready, uses one governed definition for every business term, and is reached by models through a controlled route with access rules and lineage. Volume alone does not make data AI-ready. The quickest check is to ask a real business question, one whose answer you already know, of your data through a model. If the model returns the correct figure and you can trace it to source records, the data behind that question is ready. If the figure is wrong or cannot be traced, the cause is usually location, quality, meaning or access. Our free readiness assessment gives a wider view in 16 questions. Infrastructure is ready when the data a model needs sits in one governed store, such as a cloud data warehouse or a lakehouse, rather than spread across operational systems and exports. Without one central store, a model must join scattered sources itself or work from out-of-date copies. The store then needs staged cleaning, a semantic layer that defines each business term, and a governed access route for models. Start with one use case and the data it depends on. Bring that data into a central store, clean it through bronze, silver and gold layers, agree one definition for each business term in a semantic layer, and give models a controlled route to it. The time this takes depends on the number of sources, their quality and how far definitions differ between teams, so scope it around one use case first. An AI-ready dataset is one a model can use without manual fixes. It is accurate, complete and consistent, documented with clear definitions and metadata, and held under access controls. In an organisation this usually means gold-layer tables in a governed warehouse, built around business entities and measures, with every term defined in a semantic layer so that a model and your dashboards return the same figures. ## Frequently asked questions ### What is AI-ready data? AI-ready data is data a model can use to give a correct answer that you can trace back to its source. It sits in one central store, has been cleaned in stages from raw to conformed to business-ready, uses one governed definition for every business term, and is reached by models through a controlled route with access rules and lineage. Volume alone does not make data AI-ready. ### Is your data AI-ready? The quickest check is to ask a real business question, one whose answer you already know, of your data through a model. If the model returns the correct figure and you can trace it to source records, the data behind that question is ready. If the figure is wrong or cannot be traced, the cause is usually location, quality, meaning or access. Our free readiness assessment gives a wider view in 16 questions. ### Is your data infrastructure ready for AI? Infrastructure is ready when the data a model needs sits in one governed store, such as a cloud data warehouse or a lakehouse, rather than spread across operational systems and exports. Without one central store, a model must join scattered sources itself or work from out-of-date copies. The store then needs staged cleaning, a semantic layer that defines each business term, and a governed access route for models. ### How do you make data AI-ready? Start with one use case and the data it depends on. Bring that data into a central store, clean it through bronze, silver and gold layers, agree one definition for each business term in a semantic layer, and give models a controlled route to it. The time this takes depends on the number of sources, their quality and how far definitions differ between teams, so scope it around one use case first. ### What are AI-ready datasets? An AI-ready dataset is one a model can use without manual fixes. It is accurate, complete and consistent, documented with clear definitions and metadata, and held under access controls. In an organisation this usually means gold-layer tables in a governed warehouse, built around business entities and measures, with every term defined in a semantic layer so that a model and your dashboards return the same figures. --- Shipshape Data is a London AI consultancy: we build the data foundation your AI depends on, then the AI on top. Site guide for agents: https://shipshapedata.com/llms.txt | Contact: hello@shipshapedata.com