AI-Ready Data: How to Fix Data Quality One Use Case at a Time
AI-ready data is not a state your whole company reaches after a heroic clean-up. It is a small, provable thing: the handful of sources one use case needs, shown to be owned, current, single-versioned, correctly labelled for access and still identifiable after it has been cut into pieces. This lesson shows you how to prepare data for AI in about a week, for one use case, and how to avoid both of the expensive mistakes: loading everything, and cleaning everything.
You are now at Level 2 of the L1 to L5 ladder, AI Capable, where pilots have to prove something. Level 1 left you with an AI data inventory that maps and classifies your sources (Lesson 2) and, in the previous lesson, an acceptable use policy and an honest AI readiness assessment (Lesson 5). Those told you what you have and what is allowed. This lesson turns them into a short list of sources that a real pilot can stand on. The next lesson, building your first RAG system and finding out why naive retrieval fails, puts a retrieval pipeline over exactly that list.
Why AI-Ready Data Stalls Pilots, and What It Costs
Start with one sentence. It comes from a worked illustration in Anthropic's "Contextual Retrieval" write-up (19 September 2024), not from an incident at any company, and it is the most useful sentence I know for explaining data quality to people who think they already understand it:
"The company's revenue grew by 3% over the previous quarter."
Read it as an auditor would. Is it accurate? Yes, every word. No typos, no duplicates, no broken encoding. By every traditional data quality test, this sentence passes.
Now read it as a person who asked, "How did our biggest supplier do in the second quarter?" Which company? Which quarter? Those two facts lived in the title and header of the original report. When the report was cut into small passages so an AI system could search it, the title and header stayed behind. The passage arrived alone, like a page torn out of a binder with no cover.
That cutting is called chunking: splitting a long document into short passages, called chunks, so that a search step can find the few that matter and hand them to the model. It is a normal and necessary step. It is also the moment a perfectly good document forgets who it is.
Anthropic reports that adding a short note of situating context to the front of each chunk (50 to 100 tokens, a token being roughly a word fragment) cut retrieval failures by 35% on its own mix of benchmarks. Treat that number with care: it is a vendor measuring its own method on its own tests. Take the direction, not the digit: content that still knows what it is gets found and used correctly more often than content that does not.
So here is the lesson of the sentence. Data quality for AI is not mainly about typos. It is about whether each piece of content still knows what it is, where it came from, how old it is and who may read it, after it has been cut up.
What does unready data look like inside a pilot team?
It looks like a team asking for "the warranty documents" and receiving a mess nobody is embarrassed by, because it has always been this way.
They get a shared drive export with four versions of the same policy. They get a folder of scanned PDFs that nobody has opened since a reorganization. They get a wiki whose owner left the company. Nobody can say which version is authoritative, when it was last reviewed, or who is allowed to see it. The demo is on Friday, so the team loads all of it. The assistant then answers, confidently and politely, from the 2022 policy.
Notice that nothing in that story is a model problem. A human temp handed the same folder would make the same mistake.
How much does unready data cost AI projects?
Nobody has published a measured abandonment rate caused by data readiness, so be suspicious of anyone who quotes one. What exists is a pair of surveys and a pair of forecasts, and they all point the same way.
- The newest survey. In a release of 16 April 2026, Gartner reported a survey of 353 data and analytics and AI leaders, run in November and December 2025. Organizations that reported successful AI initiatives invested up to four times more, as a percentage of revenue, in foundations such as data quality, governance, AI-ready people and change management than those reporting poor outcomes. Success was self-reported and the link is a correlation, so read it as direction, not proof.
- The earlier survey. A Gartner survey of 1,203 data management leaders, run in July 2024 and released on 26 February 2025, found that 63% of organizations either do not have, or are unsure whether they have, the right data management practices for AI.
- The forecast. In the same February 2025 release, Gartner predicted that through 2026 organizations would abandon 60% of AI projects that are unsupported by AI-ready data. Read the wording slowly. It is a prediction, and it is conditional: it describes projects that lack AI-ready data, not all AI projects.
- The earlier forecast. On 29 July 2024 Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, and named poor data quality first among the causes, alongside inadequate risk controls, escalating costs and unclear business value.
- The explanation. IDC research with Lenovo, reported by CIO.com on 25 March 2025, attributes the low conversion from proof of concept to production to "the low level of organizational readiness" in data, processes and IT infrastructure, and lists insufficient AI-ready data as a cause. The coverage does not state the sample or method, so use it as a direction too.
A word on age. The two forecasts were made in 2024 and early 2025: one covered the period to the end of 2025, the other runs to the end of 2026, and I found no published scoring of whether either came true. They stay here because they are the best-dated statements of the problem from a named analyst firm, not because they measure it. Quote them as forecasts, with their dates, and nobody on your steering committee can catch you out.
That is the first cost: pilots that demo well on Friday and die in the steering committee three months later because nobody trusts the answers.
The second cost is the opposite mistake, and careful organizations make it. Somebody hears "data is the problem" and launches a "clean all our data first" program. It runs for a year, produces a governance framework, a tooling evaluation and many meetings, and delivers nothing a pilot can use, because it was never aimed at one.
Why do organizations get data readiness wrong?
Five habits, all of them reasonable until you look closely.
- They treat readiness as a property of the data. It is a property of data for a use. More on this in the next section, because it changes everything.
- They treat it as a project with an end date. Gartner's February 2025 release says readiness is a process and a practice, and that AI-ready data is not "one and done".
- They rely on traditional data management. The same release says it is too slow, structured and rigid for AI teams, and that organizations lack the practices and metadata to judge whether data is ready.
- They never settled ownership and access. In CIO.com's reporting of 5 December 2024, EchoStor's Daniel Clydesdale-Cotter warned that many enterprises may be rushing to implement AI "without properly considering who owns the data, where it resides, and who can access it through AI models", and IDC's Daniel Saroff put much of the lack of success in generative AI down to "bad data".
- They see metadata as paperwork. Metadata is data about a document: its title, date, owner, product, jurisdiction. The 3% sentence shows it is not paperwork. It is the part that makes the content usable.
How AI-Ready Data Works: Fit for One Use Case, Then Kept That Way
Here is the definition in one sentence, following Gartner's February 2025 description: AI-ready data is data that has been shown to be fit for one specific AI use case, and kept that way, with metadata, governance and monitoring as the proof.
Three ideas are packed into that sentence. Each one removes a common argument from your next meeting.
The kitchen analogy: prep for tonight's menu
Walk into a good restaurant kitchen an hour before service. You will see rows of small containers: chopped shallots, portioned fish, a sauce base, each one labelled and dated, each one within arm's reach of the cook who needs it. Chefs call this mise en place, which just means "everything in its place".
Now notice what the chef did not do. The chef did not empty, scrub and relabel the entire walk-in fridge before cooking a single dish. They took tonight's menu, pulled exactly the ingredients those dishes need, checked that each one was fresh, labelled it, and put it where it would be used. Tomorrow's menu gets its own prep.
"Fix all the data first" is cleaning the whole warehouse while the dining room sits empty. "Load everything, the demo is Friday" is cooking from whatever is nearest the fridge door without checking the date. And a label that falls off a container is the 3% sentence: perfectly good stock that nobody can identify, so it either gets thrown out or, worse, gets used in the wrong dish.
Hold on to that picture. Every step in the workshop is a piece of it.
Is readiness a property of the data, or of the use?
Of the use. The same data can be ready for one job and unready for another on the same afternoon.
Picture an HR policy folder. For the use case "summarize our leave policy for head office staff", it may be completely ready: one current document, a named owner, reviewed within the last year. For the use case "answer payroll questions for staff in all 14 countries", the same folder is nowhere near ready, because twelve of the country documents are missing and two are in draft.
Nothing about the folder changed. The menu changed.
This is why "is our data clean?" is a question nobody can answer and "is it ready for which set of questions?" is one you can settle in a two-hour meeting. Gartner's definition makes the same move: the data must be representative of the specific use case. When a colleague says the data is not ready for AI, your reply from now on is: ready for what?
Does AI-ready mean clean?
No. It means representative, and representative includes the mess. Gartner's definition says the data must be representative of the use case, "of every pattern, errors, outliers and unexpected emergence" that is needed to train or run the model for that specific use.
That sounds odd until you think about the kitchen again. If real customer emails arrive misspelled, half-finished and angry, then the emails you test your assistant on must be misspelled, half-finished and angry too. A tidy test set that looks nothing like Monday morning proves nothing about Monday morning.
The limit on this comes from Gartner analyst Roxane Edjlali in the same release: "Above all, if the data has issues, then the data is not ready for AI." The two statements fit together. Keep the natural mess of real inputs, because the system must cope with it. Remove the defects in your reference content, such as a superseded policy or an unowned wiki, because the system will repeat them as truth.
What are the five readiness checks for a document source?
For use cases that answer questions from documents, five checks decide it. This is a practical checklist assembled from the sources in this lesson, not a published standard, so adapt the wording to your own environment.
- A named owner. One person who approves changes to the source. A team name or a mailbox does not count.
- A freshness date. You know when each document was last reviewed, not just last saved.
- A single authoritative version. One live copy. The others are retired or clearly marked as superseded.
- A true access label. The classification from your Lesson 2 inventory, so you know who may see answers drawn from it, and actual sharing that matches it. More on the second half below.
- Document-level metadata that survives chunking. Title, date, product, jurisdiction or whatever your use case turns on, attached to every chunk, not only to the file.
Two pieces of independent evidence show why checks 2, 3 and 5 are not bureaucracy.
On metadata: Deakin University's paper "Seven Failure Points When Engineering a Retrieval Augmented Generation System" (arXiv, 11 January 2024) reports from its case studies that adding the file name and chunk number to the retrieved context helped the model read the passages. Something as plain as a file name did real work.
On freshness and versions: the Stanford RegLab and HAI study of commercial legal research tools (May 2024), which opens the lesson on why naive retrieval fails, found a commercial tool presenting a standard overturned by the Dobbs decision as good law, which is your four versions of the warranty policy with higher stakes.
Why do access labels for AI data need checking against real permissions?
Because an AI assistant reads everything the asking person can open, and summarizes it without being asked. An access label is the classification a source carries (Public, Internal, Sensitive or Prohibited, the four classes from Lesson 2): it says who may see answers drawn from that source and which tools may receive it. A permission is something different: who can actually open the file today.
The two drift apart quietly. A Sensitive folder gets a company-wide link for a 2022 project, and nobody takes it back or notices, because nobody goes looking. An assistant goes looking on behalf of everyone. Microsoft's SharePoint guidance for Copilot (Microsoft Learn, 2024) puts it plainly: "Copilot for Microsoft 365 utilizes all data that a user has access to, which may include broadly shared files that the user is unaware of." The same holds for any assistant that answers with the asking user's rights.
So the access check has two halves. First, the label must allow the tool you plan to use, under your acceptable use policy. Second, the actual sharing on the source must match the label. A source that passes the first half and fails the second is a leak waiting for the right question. In Lesson 7 the label travels into the retrieval pipeline; here you make sure it is true before it travels.
What changes when the use case reads tables, not documents?
The checks stay; their names change. Not every pilot is a document assistant. Some query a sales database, classify claims or extract fields from invoices. The five checks apply, translated:
- Owner: the named owner of the table or feed, not the platform team that hosts it.
- Freshness: the load time and the agreed refresh service level, rather than a review date.
- Single version: one agreed definition of each business term. If finance and sales count "active customer" differently, the model will use whichever definition it finds first, and one team will call the answer wrong.
- Access label: a classification on the columns and rows the use case touches, not only on the database.
- Metadata: column descriptions, units and allowed values, written for a stranger to the schema.
For extraction and classification use cases, the question set in step 1 of the workshop becomes a sample of 30 to 50 real inputs, such as a recent month of invoices, chosen to include the messy ones. When those translated checks have to hold across many teams and pipelines, they become written agreements: data contracts, a semantic layer and lineage, covered in Data Architecture for AI (Lesson 12, at Level 3). For a first pilot, the scorecard is enough.
The Workshop: Run an AI Data Readiness Assessment for One Use Case
The method is called the use-case data readiness sprint: an AI data readiness assessment scoped to one use case, with the fixes done inside the same week. It takes one 2-hour working session, about four working days of clean-up, and one 45-minute review. You finish holding a signed document called the Use-Case Data Readiness Scorecard.
Who needs to be in the room, and for which steps?
Not everyone needs the whole two hours. Match the people to the steps, and do not accept a delegate who cannot decide.
- The business owner of the use case, for the whole session. They decide what "good enough" means and sign at the end.
- One or two frontline experts who answer these questions by hand today, for steps 1 and 2. They know where the real answers live. Release them once the mapping is done.
- The data owner or steward for each candidate source, for steps 3 to 5. Only they can say which version is authoritative.
- A security or privacy representative, for step 4. Thirty minutes with the sharing settings open on a screen beats a written review a fortnight later.
- The engineer who will build the pipeline, for step 6.
Bring three inputs: the data inventory and classification labels from Lesson 2, the acceptable use policy and readiness assessment from Lesson 5, and one chosen use case. One. If you cannot pick one, that is the real finding and no amount of data work will fix it.
How to prepare data for AI: the seven steps
Gartner's February 2025 release sets out five steps: align data to use cases; identify governance requirements, including how sensitive data is detected and protected when fed to models; evolve metadata from passive to active; prepare pipelines that build a dataset for training as well as a live feed to production (for a retrieval assistant, the first is your test set); and test and monitor the data. The seven steps below cover all five and make them concrete.
- Write the use case as a question set. Collect 30 to 50 real questions the system must answer. Take them from the service desk queue or the shared inbox. Do not invent them in the room, because people in rooms invent tidy questions and customers do not ask tidy questions.
- Map each question to the source that answers it today, and scope to those sources. Ask the frontline experts, "Where do you look this up?" Some questions will have no source: the answer lives in somebody's head. Mark those. They are your first finding, because content that does not exist cannot be retrieved by any system at any price. The sources that do appear are your scope. Everything else is out, by decision, and you write down why. This is the chef pulling tonight's ingredients.
- Score each source on the five readiness checks. Owner named, freshness date known, single authoritative version, access label present and true, document-level metadata present. Yes, no, or partly. Do not debate percentages.
- Check every access label against the policy and against the real sharing. The security representative confirms that each label is permitted for the tool under the acceptable use policy, then opens the sharing settings and compares who can actually open the source with what the label allows. Agree here how sensitive data will be detected and kept out of the model's input. A mismatch is a blocker until it is fixed.
- Fix only what blocks the question set. Retire duplicates. Assign owners. Add missing titles, dates and product or jurisdiction tags. Remove over-broad sharing. Then log everything you decided not to fix. That log is what lets you defend the scope later.
- Decide how metadata and freshness are kept, and give each monitoring check an owner. Ask the engineer two questions. When a document is split, what goes with each piece? And when the policy changes, how does the new version get in and the old one get out? A first load with no refresh plan is a stale system on a timer. Then set two or three checks, such as a stale-document count, an orphaned-source count and a monthly sample audit, each with a named person.
- Sign the scorecard and hand over the question set. The business owner and the data owners sign. The question set goes forward to the next two lessons, where it seeds your retrieval tests and your golden set of reference answers.
A filled-in example
This example is invented for illustration. It is not a real company and the numbers are not results.
Picture a 900-person industrial equipment distributor. The service desk wants an assistant that answers warranty and parts questions from field technicians, who currently phone the desk from customer sites and wait.
The team pulls 40 real questions from the previous month of tickets. Things like "Is the hydraulic pump on a model 40 covered after a third-party repair?" and "What replaced part 7731 in the northern region?" In the 2-hour session, the two senior service agents say where they look up each one. Two things fall out of that mapping straight away. Only four sources matter. And 6 of the 40 questions have no written answer anywhere: the agents simply know, or they ask a colleague who has been there for twenty years.
Then the room scores the four sources.
| Source | Owner | Freshness | Single version | Access label | Doc metadata | Verdict |
|---|---|---|---|---|---|---|
| Warranty policy PDFs | Head of service (named) | Reviewed March 2026 | No, 4 versions on the drive | Internal; shared with all staff, which matches | Title and date present, no product line | Fix: retire 3 versions, tag product line |
| Parts catalog export | Supply chain lead | Nightly | Yes | Internal; matches | Part number, model, region | Ready |
| Service bulletins wiki | None (owner left) | Unknown | Unclear | None | Title only | Blocked: assign owner, label, date every page |
| Old email archive of "how we handled it" | None | 2019 to 2023 | No | Sensitive (customer names); shared mailbox open to 60 people | None | Out of scope by decision |
Read the verdict column, because each row teaches something different.
The warranty PDFs are nearly ready. They have an owner, a recent review and sharing that matches their label. The problem is the one from the start of this lesson: four versions. The fix is small and specific. Retire three, and add a product line tag so a pump question does not retrieve the compressor policy.
The parts catalog is ready as it stands. Do not touch it. The urge to improve things that already pass is how one week becomes one quarter.
The service bulletins wiki is blocked. It has no owner, and a source with no owner should not go into a pilot, because nobody can tell you whether a page is still true. It stays out until someone is named, the wiki is labelled, and every page carries a date.
The email archive is the tempting one. It is full of real answers. It is also undated in any useful sense, unowned, and full of customer names. In step 4 the security representative finds it in a shared mailbox that 60 people can read, including two former contractors, far wider than a Sensitive label allows. The room excludes it on purpose, records why, and reports the sharing problem as a separate finding. That one line in the scorecard will save you an argument in three months, when someone asks why the assistant does not know about the 2021 recall.
The outcome: two sources ready, one fixed in four days, one excluded on purpose, one access problem found, and six content gaps handed to the head of service. Nobody cleaned the data warehouse.
What goes in the data readiness scorecard?
Copy this template into your controlled-document system. One scorecard per use case; two pages is plenty.
| Section | What you record |
|---|---|
| Use case and business owner | One sentence describing the use case, and the name of the person accountable for it |
| Question set | 30 to 50 real questions (or real inputs), each with the source that answers it today |
| Sources in scope | Only the sources named against the question set |
| Sources excluded | Each excluded source and the reason |
| Per-source scores | Owner, freshness date, authoritative version, access label, document-level metadata |
| Access and governance check | Label permitted for the tool, actual sharing compared with the label, and how sensitive data is detected and kept out of the model's input |
| Fixes done and fixes deferred | What was changed, by whom, and what you chose to leave |
| Content gaps | Questions with no source, and who will write the answer |
| Monitoring checks | Two or three checks, each with an owner and a cadence |
| Sign-off | Names and date |
| Next review date | A real date in a real calendar |
The 45-minute review at the end of the week has one job: go down the scores, confirm the fixes happened, and sign. If a blocked source is still blocked, it stays out and the pilot starts without it. A smaller pilot on sources you trust beats a bigger one on sources you hope are fine.
Failure Modes and the Metrics to Track
Seven ways this goes wrong. Each has a tell you can spot in a meeting or in ten minutes at a screen.
How do you spot a data readiness effort going wrong?
- Boiling the ocean. The plan mentions "enterprise data quality" and no question set exists. Ask to see the questions. If there are none, there is no scope, and the work will expand to fill the budget cycle.
- Orphaned sources. Ask, "Who approves a change to this document?" Silence is your answer. So is "the team".
- Duplicates and stale versions in the index. Two retrieved passages contradict each other on a date or a figure. When a tester reports that the assistant "changes its mind", look for two live versions before you blame the model.
- Metadata lost at chunking. Read ten random chunks cold, with no file names showing. If you cannot tell which document, product and date each belongs to, the model cannot either. This is the 3% sentence test, and it takes ten minutes.
- Sensitive content swept in. The scoped sources include material above the classification your acceptable use policy allows for the tool. The email archive full of customer names is the classic case.
- Labels that disagree with permissions. Pick one Sensitive source and open its sharing settings. A company-wide link, or a group called "All staff", on a Sensitive source means the assistant will hand its contents to anyone who asks the right question.
- One and done. The scorecard has no next review date and no named person for the monitoring checks. It was true on the day it was signed and has been decaying since.
Which metrics show that data is ready for a use case?
Six numbers, all of which you can count by hand for a first pilot. None of them needs a platform.
| Metric | How to measure it |
|---|---|
| Question coverage | Share of the question set with an identified authoritative source (count it in step 2) |
| Owner coverage | Share of in-scope sources with a named owner |
| Freshness | Share of in-scope documents with a review date inside the agreed window |
| Duplicate rate | Documents with more than one live version, divided by in-scope documents |
| Metadata completeness | Share of sampled chunks carrying title, date and the agreed tags |
| Label and permission match | Share of in-scope sources whose actual sharing matches their access label (count it in step 4) |
In the distributor example, question coverage at the first session was 34 of 40, which is 85%. That one number tells the business owner something no data quality dashboard would: six things technicians ask about have never been written down. Fixing that is a job for the head of service, not for the data team.
Report these six to your steering committee beside the pilot's results. When answers get worse in month four, look at freshness and duplicate rate first, before anyone proposes buying a different model.
Frequently Asked Questions
What is AI-ready data?
AI-ready data is data shown to be fit for one specific AI use case, and kept that way, with metadata, governance and monitoring as the evidence. Gartner's February 2025 description says the data must be representative of the use case, including its errors and outliers, and that readiness is an ongoing practice rather than a one-off project.
Do we have to clean all our data before starting with AI?
No. Scope the work to the sources one use case needs, fix what blocks that question set, and record what you chose to leave. A year-long clean-up with no use case attached delivers nothing a pilot can use.
How many AI projects fail because of data?
Nobody has measured it; the most quoted figure is a forecast. Gartner predicted in February 2025 that through 2026 organizations would abandon 60% of AI projects unsupported by AI-ready data, a conditional prediction rather than a failure rate for all AI projects. Gartner's April 2026 survey points the same way: organizations reporting successful AI initiatives invested up to four times more, as a share of revenue, in foundations such as data quality and governance.
How long does an AI data readiness assessment take?
For one use case, about a week: a 2-hour working session, roughly four working days of targeted fixes, and a 45-minute review to sign the scorecard. An assessment that tries to cover the whole data estate has no natural end, which is why it should be scoped to one question set.
Who should own AI data readiness?
The business owner of the use case and the data owner of each source own it together, with the data or AI lead running the process. A source with no owner should not go into a pilot.
Key Takeaways
- AI-ready data is a property of data for one use case, proven by metadata, governance and monitoring. When someone says the data is not ready, ask: ready for what?
- Avoid both expensive mistakes. Do not load everything because the demo is Friday, and do not clean everything before any pilot starts. Prep for tonight's menu.
- Start from 30 to 50 real questions taken from the queue. The sources those questions point to are your scope, and questions with no source are content gaps to write, not data to clean.
- Score each source on five checks: named owner, freshness date, single authoritative version, a true access label, and document-level metadata that survives chunking. Fix it or exclude it.
- An access label is only true if the real sharing matches it. Assistants read whatever the asking user can open, so check the sharing settings before the pilot, not after the first leak.
- Representative does not mean spotless. Keep the natural mess of real inputs in your tests, and remove defects such as stale or duplicate versions from your reference content.
- You leave with a signed Use-Case Data Readiness Scorecard with a next review date, because AI-ready data is not "one and done". The next lesson, Building Your First RAG System, builds a retrieval pipeline over exactly those sources and uses the questions to find where retrieval breaks.
Skill.re