The AI Data Inventory: Mapping and Classifying Your Data Before AI Touches It
At the Thursday ops review, someone proposes connecting the new AI assistant to the shared drive "so it can answer questions". Before anyone says yes, you need an AI data inventory: a short list of the data sources that one AI use would read, each with a named owner, a classification, a sampled quality check and a note on who can access it. It takes about two weeks of part-time work, it fits in one spreadsheet, and it is the cheapest way to find out what your data will teach an AI system before the system teaches it back to your customers, your staff or a regulator.
The Problem: Nobody Knows What Is in the Data, and AI Reads All of It
Stay in that ops review. The proposal sounds sensible: the drive holds years of know-how. Then somebody asks four questions, and the room goes quiet:
- What is actually on the shared drive?
- Who owns each folder?
- Which folders hold personal data?
- How much of it is ten years out of date?
From there the project usually goes one of two ways. It stalls for months while people argue about a company-wide data clean-up. Or it goes ahead, and soon after launch the assistant helpfully surfaces a salary spreadsheet to anyone who asks the right question. This lesson is the third way: answer those four questions properly, for this one use, in a fortnight.
What happens when nobody asks what the data represents?
Here is the expensive version, from a company with more data, engineers and money than you have. Amazon built a model to screen resumes. To teach it what a good candidate looked like, the team trained it on patterns in resumes submitted to the company over a 10-year period. The model found the pattern in that history and began penalizing resumes that included the word "women's", as in "women's chess club captain". Reuters reported in October 2018 (republished by CNBC on 10 October 2018) that Amazon had scrapped the project; the team had been disbanded by the start of 2017, after executives lost hope for it.
Look closely at what failed. The data was not missing or scattered across forgotten servers. It was plentiful, internal, and clean in the sense a database administrator means: the fields were filled in and the records were real. By every technical test you would normally run, it was good data.
What nobody had written down was what that data represented. It was not a record of "what good candidates look like" but of a decade of past applications and the human judgements around them, with a skew inside it. The model did exactly what it was asked: it learned the history. One line in a spreadsheet, asking "what is this, who owns it, what does it represent, and is it fit for this particular use?", is the cheap step that was missing.
How common is this?
Very. In a Gartner press release dated 26 February 2025, based on a survey of 1,203 data management leaders run in July 2024, 63% of organizations either did not have, or were unsure whether they had, the right data management practices for AI. Nearly two thirds could not say yes.
The same release predicted that through 2026 organizations would abandon 60% of AI projects not supported by AI-ready data. Treat that as what it is: a forecast, not a measurement. Use it as a signal of where analysts expected the trouble, not as proof.
Here is the number that should change how you plan. A Gartner survey published 30 June 2025 (432 respondents, fieldwork in late 2024) found that data availability and quality was a top barrier for 34% of low-maturity organizations and for 29% of high-maturity ones. The gap is five points. Data trouble does not go away as you climb the ladder. You do not solve it once; you build a habit. This lesson is that habit in its smallest form.
Why not just catalogue everything first?
Because you will never finish. Splunk, in a blog post dated 26 March 2025, defines dark data as "the large volumes of unexplored raw data available to an enterprise" and cites its own global survey putting it at 55% of an organization's data. Hold that figure loosely: Splunk sells tools for this problem, and the page does not give the survey year. We found no independent measurement. Even at half that share, a large part of your estate has never been looked at by anyone.
Centralizing it is no quick fix either. The Cisco AI Readiness Index 2025, released 14 October 2025 from a survey of more than 8,000 leaders and summarized by RCR Wireless on 9 December 2025, reports that 19% of organizations have centralized data infrastructure, against 76% of its top "Pacesetter" group, and that 64% struggle to centralize data. That is a vendor survey with self-reported answers (Cisco sells infrastructure), so read it as direction, not precision.
So be practical. A data catalog, the searchable register of every data asset with its metadata, is a fine destination, but as a precondition it turns "catalogue everything, then start AI" into a multi-year programme that is abandoned before a single AI use benefits. The first inventory has to be small, and its rows can seed the catalog later; Level 3 returns to that wider structure in Data Architecture for AI: data contracts, semantic layers and knowledge graphs. The trick now is knowing how to make the inventory small without making it useless.
How an AI Data Inventory Works: Four Ideas, One at a Time
You are at Level 1 of the ladder (AI Aware), in the chapter about seeing what you actually have. The previous lesson, on how to find and govern the shadow AI your employees already use, gave you a Shadow AI Register and a one-page Data-to-Tool Rule. One column of that register was hard to fill in honestly: "highest data class involved". Nobody could complete it with confidence, because the data itself had never been listed or classified. This lesson fixes that, for one use case. The next lesson, on choosing your first AI use case and avoiding pilot purgatory, then scores candidates on value, feasibility and risk, and your inventory ratings become the data part of that feasibility score.
Four ideas follow. Each removes a specific way the work goes wrong.
Idea one: why scope by use case and not by company?
Scope by use case because readiness is a property of data in relation to a job, not a property of data on its own. Data mapping for AI means tracing which sources one AI use would read and how they would reach it, and for a single candidate use that is typically 5 to 15 sources.
Think of your attic. Nobody clears all of it before using it. If you need the box of winter coats, you clear a path to that box, and you label each box you have to move on the way. After a few trips, a surprising amount of the attic is labeled, and you never spent a weekend on it.
This is not a shortcut we invented; it is how Gartner describes the term. AI-ready data is data that is fit for a specific AI use. In the February 2025 release Gartner puts it this way: "The data must be representative of the use case", including the patterns, errors and outliers that use needs, and readiness is shown through a process backed by metadata to "align, qualify and govern the data".
Notice the surprising phrase in there: errors and outliers. Data that has been scrubbed until it is spotless can be less ready, not more. A fraud model needs to see fraud. An assistant that drafts replies to angry customers needs to have seen angry emails. So "is our data AI-ready?" has no answer. "Is this data ready for this use?" does. That was Amazon's missing question.
Idea two: why must the owner be a person's name?
The owner must be a named person because a department cannot answer a question on Thursday afternoon. A data owner is the business person who can say what a field means, whether it can be trusted, and who should be allowed to see it.
Most systems already have a systems administrator in IT, who can tell you where the data lives, how big it is and who has a login. They usually cannot tell you that the "delivery date" field means promised date for one carrier and actual date for another. Only someone who uses the data daily knows that.
So every row gets two kinds of knowledge: the IT administrator for the plumbing, and a business owner for the meaning. If the owner column says "Finance" or "IT", the row has no owner. In a Level 1 organization there is usually no chief data officer to run this, so the inventory falls to the CIO, with one business data owner per system.
Idea three: how do you classify data without a nine-level scheme?
You use four plain classes, the same four from Lesson 1: Public, Internal, Sensitive and Prohibited, each tied to which tools may receive it. Data classification is simply putting a label on data that says how carefully it must be handled.
Many firms skip classification because the corporate scheme has nine levels nobody understands. Four is enough to decide. Picture moving house with four colors of tape:
- Green tape (Public): anyone can carry this box. Published price lists, website copy, press releases.
- Yellow tape (Internal): stays with the family. Process documents, shipment events, internal reports.
- Red tape (Sensitive): only you carry it. Customer names and addresses, contracts, anything with personal data.
- Black tape (Prohibited): this box does not go on the van at all. HR files, payroll, credentials. No AI tool receives it. (This is a data class, not a reference to the EU AI Act's prohibited practices; the two are unrelated.)
The useful thing about tape is that you do not need to know everything inside a box to choose the right color. You need to know the most delicate thing in it. If you are unsure whether something is Public, ask the question from the Tines guide to acceptable use policies: "Would I post this publicly on the internet?" Tines calls this the Public Test, and its own rule is that if the answer is no, you do not put the data into an AI tool. This program adapts it for classification: if the answer is no, the data is not green, and the Data-to-Tool Rule decides which approved tools may receive it. (The four-tier scheme is one vendor's practitioner advice, not a formal standard. It earns its place because people can remember it.)
Idea four: how do you check quality and access without a big tool?
You look. For quality, you pull a sample of 50 to 100 records and read them. For access, you open the permission list and read that too.
The most common quality check in business is an assumption: "it is in the ERP, so it is right." A food inspector does not accept "it came from a good supplier". They open a few crates. Fifty to a hundred records will not give you a statistically perfect error rate, and need not. It will tell you in an hour whether you have a 2% problem or a 20% problem, and that is the decision you need to make before anything is connected.
Access is the part people forget, and it is the part that causes the salary spreadsheet incident. An AI assistant connected to a source inherits whatever permissions that source already has. If the folder is readable by the whole company today, the assistant can quote from it to the whole company tomorrow. Today nobody goes looking; an assistant makes looking effortless.
Access cuts the other way too: the door into the AI system matters as much as the doors it opens. In June 2025, security researchers Ian Carroll and Sam Curry found that McHire, the AI hiring chatbot platform McDonald's uses and Paradox.ai runs, accepted the password 123456 on a test admin account. Combined with a flaw that let them read other applicants' records, their write-up says this potentially exposed more than 64 million job application records. Read it precisely: it counts applications, not unique people, and records reachable, not records confirmed viewed. Paradox.ai fixed the flaw within about a day of disclosure and said in its update of 9 July 2025 that the researchers opened seven chat records, five of them containing personal information, and that only the researchers had viewed them. The AI was not the weak point. The door to the data was. That is why the access column in your inventory matters as much as the quality column.
The Workshop: Build Your AI Data Inventory in a Two-Week Sprint
One honest note before you start. This sprint is not a published standard; it is a working method built from Gartner's per-use-case framing, Splunk's dark data discovery steps and the four-class scheme. Adjust the pace, but never drop the sampling or the permission check: that is where the findings come from.
Who needs to be in the room, and how long does it take?
The room follows the data, not the org chart, and most people are needed for one session only.
| What | Detail |
|---|---|
| People | The sponsor of the candidate use case (kickoff and review); the IT administrator of each in-scope system (their source's hour); one business owner per source (sampling and the meaning question); the privacy lead or data protection officer (one classification session); whoever kept the Shadow AI Register (kickoff). If the AI is built into your office suite, add the administrator of your collaboration platform for the whole sprint. |
| Time | A 90-minute kickoff, about 1 hour per source for sampling, a 90-minute review |
| Inputs | The Shadow AI Register and Data-to-Tool Rule from Lesson 1, a list of your systems, read access to each source |
| Output | One spreadsheet: the AI Data Inventory for one use case |
What are the steps?
- Pick the scope. Write the candidate AI use in one sentence, then list the systems that would feed it. Stop at 15. If you cannot write the use in one sentence, you are not ready to inventory; you are still choosing. If the list passes 15, the use case is too broad, so narrow the sentence, not the rigor.
- List each source. One spreadsheet row per source: system, named owner, what it holds, format, approximate volume, update frequency, where it lives. Approximate is fine. "About 40,000 emails, growing daily" is a good answer.
- Tag the class and check the purpose. Give each source one of the four classes and flag whether it holds personal data. If a source mixes classes (a shared drive with both rate cards and HR files), split it into separate rows: a row that needs two colors of tape is really two boxes. For every row flagged as personal data, the privacy lead records two answers in their own columns: is this AI use compatible with the purpose the data was collected for, and is a data protection impact assessment (DPIA) needed? Under the EU GDPR these are the purpose limitation principle in Article 5(1)(b) and the DPIA duty in Article 35; outside the EU, use your own law's equivalent.
- Sample the records. For each key source, pull 50 to 100 records and score four simple checks: completeness (are the fields this use needs filled in?), duplicates (does the same record appear more than once?), staleness (how much is too old to be true?), and obvious errors (values that cannot be right). Sample; do not scan everything. Record the sample size.
- Record access. Who can read this source today? Count sharing links too: anything shared with the whole company, or through an anyone-with-the-link link, is readable by everyone. Then the AI question: if an assistant were connected to it, would it show people things they should not see?
- Ask the Amazon question. Sit with the owner and ask: what does this data represent, and what would a model learn from it that we would not want? Write the answer down, even if it is "nothing we can think of". This is the one step no tool can do for you.
- Rate each source for this use. Green: usable as it is. Amber: usable after a named fix. Red: not usable until something changes. Fix or exclude every red. The rating is for this use case only; the same source can be green for one use and red for another.
- Set review dates. Give each source a date when its row will be checked again, and note any source nobody has opened in 12 months. Those unopened sources are your own small patch of dark data.
- Repeat for the next use case. Reuse the rows that already exist and add only what is new. This is the attic path getting wider with each trip.
What does a finished inventory look like?
Here is an illustration, not a real company. Picture the same 420-person freight forwarder we followed in Lesson 1. The candidate use, in one sentence: "Draft replies to customer shipment-status emails for a human to check and send."
The team lists what a drafted reply would need to read: shipment events, past emails, customer details, rates. Then somebody points out that the rate cards sit on the shared drive, next to two folders nobody had planned to discuss. Those go on the list too, precisely because a connection to the drive would reach them. They sample 80 records from each source they intend to use.
| Source | Owner | Class | Sample finding (80 records) | Access today | Rating for this use |
|---|---|---|---|---|---|
| Transport management system, shipment events | Operations manager | Internal | 6% of events missing a timestamp | Operations and customer service | Amber: fix timestamps for 2 carriers |
| Shared customer service mailbox | Customer service lead | Sensitive (names, addresses) | 14% duplicates from forwarding chains | 44 people | Amber: de-duplicate, limit to last 24 months |
| CRM customer master | Sales operations lead | Sensitive | 9% of contacts stale by more than 3 years | Sales and service | Amber |
| Rate cards on the shared drive | Pricing analyst | Internal | 3 conflicting versions of the current card | Whole company | Red: exclude until one version is marked current |
| Carrier contracts folder | Procurement lead | Sensitive | Not needed for this use | Whole company | Excluded; access flagged for tightening |
| HR folder on the same drive | HR manager | Prohibited | Not sampled | Whole company (misconfigured) | Excluded; raised as an access incident |
(In the real spreadsheet the owner column holds each person's name; job titles appear here only because the company is invented.)
Read the table the way the steering committee would. Six sources listed, six owners named. Three are usable with fixes, one is red, two are excluded. Nothing is green, and that is normal for a first pass; an inventory where every row is green on day one was probably filled in from memory.
Now look at what each row bought them:
- The timestamps. A reply that says "your shipment departed" with no time, or the wrong one, is worse than no reply. The fix is limited to two carriers, so it is a task, not a project.
- The mailbox. Forwarding chains meant the same conversation appeared several times, which would overweight a few noisy threads. Limiting to the last 24 months also answers the Amazon question: older emails reflect service promises the company no longer makes.
- The purpose check. The privacy lead marked the mailbox and the CRM as personal data and wrote, in their judgement: "Compatible: same customers, same shipments, a person sends every reply. DPIA screened, not required for drafts; reassess if replies are ever sent without review."
- The rate cards. Three files each claim to be current. An assistant would pick one and quote it with complete confidence. Red, until the pricing analyst marks one version as current.
- The HR folder. Readable by the whole company through a misconfiguration. Nobody would have found it without reading the permission list. It has nothing to do with AI, and it was the most valuable finding of the fortnight.
Tell your sponsor in advance: an inventory regularly pays for itself by finding an ordinary access problem before an assistant makes it visible.
What changes when the AI is the assistant built into your office suite?
The scope changes, and the permission check becomes the main event. For many firms the first sanctioned AI is not a custom tool but the assistant built into the office suite they already pay for. A custom assistant reads the 5 to 15 sources you connect. A suite assistant reads, for each user, everything that user can already open. Microsoft's guidance on configuring a secure foundation for Microsoft Copilot (Microsoft Learn, updated 28 October 2026) says it answers using "data that the user already has permission to access". So the freight forwarder's HR folder stops being one row. It is a pattern you have to find everywhere.
The inventory method still works; you change what a row is. Instead of one system, a row is one site, shared drive or library with a wide audience. This is sensitive data discovery, finding where delicate data lives, driven by permissions rather than by content: you hunt for places where Sensitive or Prohibited data meets a broad audience. Microsoft's guidance runs in this order:
- Find the high-risk places. Use the suite's sharing and access reports to list sites shared with everyone in the company, company-wide and anonymous links, broken permission inheritance, and sites that are inactive or have no owner. Rank them by the most delicate class inside.
- Hide the worst from the assistant for now. The guidance uses interim controls to exclude sensitive sites from Copilot discovery while the real fix happens, and says to remove them once access is fixed. Treat this as a tourniquet, not a cure.
- Fix access with the owners. Site owners run access reviews, remove excess users and company-wide links, rescope sharing links to approved people or groups, and correct inheritance. Every remediated site gets a confirmed owner: the named-person rule again.
- Label, then change the defaults. Apply sensitivity labels, restrict company-wide and anonymous links at tenant level, and require a label when a new site is created, so the mess does not grow back.
Microsoft sells both the assistant and the tools its guidance names, so treat the tool list as one vendor's implementation. The sequence (assess, contain, fix, then harden) is the part that transfers to any suite. Our practical addition: switch the assistant on first for a pilot group whose access you have reviewed, not for the whole company on day one. Level 3 returns to this risk once assistants can act as well as read, in defending against prompt injection and oversharing.
What goes in the data inventory template?
The artifact you finish with is the AI Data Inventory for one use case. Copy these fields into a spreadsheet as column headings, one row per source.
| Group | Field | What to write |
|---|---|---|
| Scope | Use case | The one-sentence candidate AI use |
| Source system | The system, mailbox, folder or site | |
| Description | Named owner | A person, not a department |
| Contents | What it holds, in plain words | |
| Format | Database table, email, documents, spreadsheets | |
| Approximate volume | A rough count is enough | |
| Update frequency | Live, daily, monthly, no longer updated | |
| Location | Where it lives | |
| Classification | Class | Public, Internal, Sensitive or Prohibited |
| Personal data flag | Yes or no | |
| Purpose compatible | For personal data: does this use fit the purpose it was collected for? Privacy lead's answer | |
| DPIA needed | For personal data: yes, no or screened, with the date | |
| Quality | Sample size | How many records you actually read |
| Completeness | Share of sampled records with needed fields filled | |
| Duplicate rate | Share of sampled records that repeat | |
| Staleness | Share too old to rely on, and what "too old" means here | |
| Error notes | Obvious errors you saw | |
| Meaning | What the data represents and known skews | The owner's answer to the Amazon question |
| Access | Who can read it | Groups and rough headcount today |
| Broad sharing links | Company-wide or anyone-with-the-link links, and how many | |
| Overexposure risk if connected to AI | What an assistant would reveal, and to whom | |
| Decision | Rating | Green, amber or red, for this use |
| Fix or exclusion | The named fix, or why it is left out | |
| Review date | When the row is next checked |
Failure Modes, How to Spot Them, and the Metrics to Track
Each failure leaves a visible mark in the spreadsheet, so you can audit your own inventory in five minutes.
| Failure mode | How to spot it | What to do |
|---|---|---|
| Boiling the ocean | More than 15 sources in scope; it is week 3 and no source has been sampled | Rewrite the use case sentence more narrowly and cut the list |
| Owner is a department | The owner column reads "IT" or "Finance" instead of a person | Ask the sponsor to name one business person per source |
| Classification by guess | Every row is "Internal"; no row was ever split; personal data rows have no purpose answer | Bring the privacy lead into one session and walk the rows |
| Quality asserted, not sampled | The sample size column is empty | Book the hour per source and read the records |
| Access ignored | No row says who can read the source; the assistant would inherit company-wide permissions | Read the permission list with the IT administrator |
| Suite assistant switched on first | The licences are assigned, but sites shared with the whole company have never been listed | Hide the worst sites, run owner access reviews, then enable a reviewed pilot group |
| Inventory as a one-off | No review dates; the next use case starts a new spreadsheet | Keep one inventory and reuse rows |
Which failure should you watch for first?
Watch for boiling the ocean, because it kills the work before the other failures get a chance to. It arrives disguised as diligence: "while we are at it, let us cover all of finance too". The answer is the attic: you are clearing a path to one box. Finance gets its turn when a finance use case is on the table.
Which metrics belong on your AI data readiness checklist?
Four numbers, each one countable straight from the spreadsheet or the access logs, and one extra you report loudly.
| Metric | How to measure it |
|---|---|
| Percentage of in-scope sources with a named owner | Count rows with a person's name over total rows |
| Percentage of in-scope sources classified | Count rows with one of the four classes over total rows |
| Error rate in sampled records | Errors found over sample size, reported per source |
| Percentage of sources no one has opened in 12 months | Last access logs |
| Overexposed sources found and closed | Access problems raised and fixed, including broad-access sites holding Sensitive or Prohibited data |
The first two should reach 100% before anything is connected. The third has no universal pass mark: the owner and the sponsor decide what error rate this particular use can tolerate, which is the whole point of judging readiness per use case. When a pilot later needs its amber rows fixed for real, AI-Ready Data: how to fix data quality one use case at a time picks them up at Level 2. Report the fifth number loudly: a board understands it without any explanation of AI.
An inventory does not make your data good. It makes your data known. You can plan around known problems; the unknown ones are what end up in an incident report.
Frequently Asked Questions
What is an AI data inventory?
An AI data inventory is a list of the data sources one specific AI use would read, each with a named owner, a classification, a sampled quality check and a note on who can access it. It is scoped to one use case at a time, not the whole company, and normally covers 5 to 15 sources in a single spreadsheet.
How is an AI data inventory different from a data catalog?
An AI data inventory covers only the sources behind one AI use and records whether each is fit for that use; a data catalog aims to describe every data asset in the organization. Start with the inventory, because Gartner (February 2025) frames AI-ready data as representative of a specific use case, and let its rows seed a catalog later.
How do you classify data before using AI?
Give each source one of four classes (Public, Internal, Sensitive or Prohibited), chosen by the most delicate thing it holds, and tie each class to the tools allowed to receive it. Split any source that mixes classes into separate rows, and for personal data record whether the AI use fits the original purpose and whether a DPIA is needed.
Do we need to fix oversharing before turning on Microsoft 365 Copilot?
Yes, fix the worst of it first, because Copilot answers from whatever each user already has permission to open. Microsoft's guidance (October 2026) says to find high-risk sites, temporarily hide sensitive ones from Copilot, have owners fix access and sharing links, and apply sensitivity labels before setting lasting defaults.
What happens if we skip the data inventory?
The most likely result is an assistant that surfaces data people should never have seen, because a connected assistant inherits the permissions of whatever it reads. Gartner also forecast in February 2025 that through 2026 organizations would abandon 60% of AI projects that lack AI-ready data, but that is a prediction, not a measured result.
Key Takeaways
- Scope to one use case. Write the AI use in one sentence and list only the 5 to 15 sources it would read. Clear the path to the box you need, not the whole attic, and reuse the rows next time.
- Name a person per source. A systems administrator knows where the data lives; a business owner knows what it means. "Finance" is not an owner.
- Use four classes and check the purpose. Public, Internal, Sensitive, Prohibited, labelled by the most delicate thing in the box. Split mixed rows, and for personal data record purpose compatibility and whether a DPIA is needed.
- Check by looking. Sample 50 to 100 records for completeness, duplicates, staleness and obvious errors, and read the permission list and sharing links, because an assistant inherits whatever access the source already has.
- Ask the Amazon question. What does this data represent, and what would a model learn from it that we would not want? Ten years of clean resume data still carried a skew. Then rate each source green, amber or red for this use only, with a review date.
- Clean up permissions before a suite assistant. An assistant built into your office suite reads everything each user can open, so find broadly shared sites, hide the worst, have owners fix access and links, then label and pilot.
- What comes next. You now hold one candidate use case whose data is known, owned and rated. Lesson 3 asks whether it is the right first use case by scoring it against others on value, feasibility and risk, and your green, amber and red ratings become the data part of that feasibility score.
Skill.re