What is AI-ready data?

Data that is complete, current, easy to find and allowed for the specific task the AI will do. Readiness depends on the task: the same files can be ready for one process and not for another.

Do I need a data team to get a process ready for an AI skill?

Usually not. For a single process, the people who run it can do most of these checks themselves in an afternoon. IT is needed mainly for access, and the data owner or privacy lead for what may go into the AI tool.

Which processes are the best first candidates for an AI skill?

Processes that are repeated often, have an agreed way of working, and pass most of the nine checks. A process with messy data is better tidied first and turned into a skill afterwards.

What if the process involves personal or confidential data?

Check with the data owner or your privacy lead whether that data may be used in your AI tool. If it may not, the skill can still prepare the rest of the work and leave the sensitive parts for a person to fill in.

ResourcesMethod

Is your process ready to become an AI skill? A data checklist

A skill can only be as good as the data it works with. Nine things to check before you turn a process into an AI skill, each with a quick test and a fix.

The Fygment team

24 October 2026, 7 min read

Most teams pick the process they turn into an AI skill by asking whether the work is repetitive. That is a good start, but it skips the question that decides whether the skill will work: is the data around the process ready?

A skill works from the files and information it is given. If the price list exists in three versions, the template lives in someone's inbox and the rules for choosing an option sit in one person's head, the best-written skill will still produce the wrong result. It will just produce it faster.

Gartner expects organisations to abandon 60% of AI projects that lack AI-ready data through 2026. That prediction is about whole AI projects rather than single skills, but the logic holds for one process too: when the data has problems, the AI cannot fix them for you.

Below are nine things to check before you turn a process into a skill, in three groups. Each comes with a test you can run in a few minutes and a fix for when the answer is no. As an example we use the client proposal skill from What makes a good AI skill.

Where the data lives

The proposal skill needs six inputs: the brief in the CRM, the case study library, the service catalogue, the rate card, the proposal template and the tone guide. Try to paste a link to each one. In many companies at least one turns out to be a file on someone's laptop, an attachment in an inbox, or knowledge in an account manager's head. A skill cannot use any of those.

Structure matters too. If every project folder is organised differently, the AI has to guess where things are. Microsoft's guidance on preparing data for Copilot recommends logical folders and consistent file names, because the AI uses them to work out what a file is and how files relate. The same guidance says to remove old drafts and outdated versions, because they confuse the AI. If a colleague cannot tell from the name which rate card is current, the AI cannot either.

And someone has to keep the data current. A rate card that was right in January is wrong by October unless somebody owns it. Microsoft recommends naming a person who looks after the data over time, and Gartner describes AI-ready data as something that needs continuous improvement, not a one-off clean-up.

1
The data is in a shared place, not in someone's head, inbox or personal drive. Test: paste a link for each input. If not: write it down, or save it in the shared place.
2
The data follows the same folder structure in every project. Test: find the same file in three projects. If not: agree one structure, apply it and write it down.
3
There is one clear, correct version of every file. Test: a colleague picks the current file by its name. If not: name the right one and archive the other copies.
4
The data is current and has an owner. Test: name who updates each input and when it was last checked. If not: name an owner, add a "last checked" date and archive what is out of date.

“The best-written skill will still produce the wrong result. It will just produce it faster.”

What good looks like

A skill learns what a good result looks like from examples. If the proposal template exists and there is at least one finished proposal everyone agrees is good, the skill has something to work towards. If there is no template yet, build one from your best recent example.

Then write down the choices. Every process has fields and options, and rules for when to pick which: which case studies may be used for regulated clients, when to offer a smaller first phase, which discounts need approval. A simple test is to give a colleague a past case and see whether they fill it in the way it was actually done. If they do not, the rules live in someone's head, and the skill will not know them either.

Finally, collect a handful of real past cases with the result that was accepted, including at least one messy one. These become the skill's test cases later. Anthropic recommends building tests before writing much of a skill, and Gartner stresses that data for AI has to represent the real work, errors and outliers included.

5
The template exists, with at least one finished example. Test: open both now. If not: build the template from your best example.
6
Fields, options and when to choose which are written down. Test: a colleague fills in a past case the same way it was done. If not: write the table, with the terms your team uses.
7
Real past cases with known correct results, including a messy one. Test: collect three to five past cases with the result that was accepted. If not: ask an expert to mark the right result for each.

Whether the AI may use it

The last two checks are about access. First, can your AI tool actually reach the data? Many AI tools only see what the person using them can open, as Microsoft documents for Copilot. Ask the AI to open each input. Scanned PDFs and photos of documents often fail even when access works, so check that the AI can read the text, not just find the file.

Second, is the data allowed in that tool? This is where companies get caught out. In Cisco's 2024 privacy study, 48% of the privacy and security professionals surveyed admitted entering non-public company information into generative AI tools, and 63% said their organisation had limited what data may be entered. Ask the data owner or your privacy lead before a skill starts handling client or personal data. If the answer is no, the skill can still help: let the AI draft, and fill in the sensitive parts yourself.

8
The data is reachable by your AI tool. Test: ask the AI tool to open each input and read it back. If not: ask IT for access, or replace scans and photos with the original files.
9
The data is allowed in your AI tool. Test: ask the data owner whether this data may go into this tool. If not: leave it out, or let the AI draft and enter the sensitive parts yourself.

What to do with the answers

If all nine are a yes, the process is a good candidate for a skill. If several are a no, fix the data first. Most of the fixes are small: writing something down, agreeing a folder structure, archiving old copies, naming an owner. They pay off without AI as well, because colleagues find the right file faster too. Even a government AI readiness scorecard from Canada asks whether the data is available, accurate and up to date before any AI tool is brought in.

Once the data is ready, the work moves to the skill itself, which we cover in What makes a good AI skill. Fygment keeps that skill with its owner, its versions and its approvals in one company library, and the past cases you collected for check seven become its test cases: a new version has to pass them before anyone uses it.

If you have a process that passes all nine checks, Fygment is free to start.

Questions we often get

What is AI-ready data?: Data that is complete, current, easy to find and allowed for the specific task the AI will do. Readiness depends on the task: the same files can be ready for one process and not for another.
Do I need a data team to get a process ready for an AI skill?: Usually not. For a single process, the people who run it can do most of these checks themselves in an afternoon. IT is needed mainly for access, and the data owner or privacy lead for what may go into the AI tool.
Which processes are the best first candidates for an AI skill?: Processes that are repeated often, have an agreed way of working, and pass most of the nine checks. A process with messy data is better tidied first and turned into a skill afterwards.
What if the process involves personal or confidential data?: Check with the data owner or your privacy lead whether that data may be used in your AI tool. If it may not, the skill can still prepare the rest of the work and leave the sensitive parts for a person to fill in.

Keep reading

What makes a good AI skill: the elements every skill needs
A reliable AI skill has a clear job, a process with steps, data sources and checks, tests with expected results, and an owner who keeps it current. Here are the 23 elements we look for, with a worked example of a skill that drafts client proposals.
AI skills aren't one-and-done: the mistakes companies make
Most AI skills are written once, tried twice and shared with the whole team. These are the six mistakes we see most often, why they happen, and what the teams that get skills right do differently.