Your AI project probably runs on data you already own
Most companies don't need new data to start with AI — they need to make the information they already have findable, current and owned.
Most companies planning their first AI project assume the hard part is acquiring data. It usually isn't. The contracts, quotes, support tickets, product specs and policy documents that a useful system needs already exist inside the business. The problem is that they sit in six places, in four versions, with no one able to say which one is correct.
That gap is where AI projects stall. Not in the model, in the filing.
Start from a question, not an archive
The instinct is to inventory everything before starting. Resist it. A full data audit takes months, produces a spreadsheet nobody reads, and delays the thing that would have taught you the most.
Pick one decision or one recurring task instead. How we price a non-standard job. What we tell a customer whose order is late. Which clause we can concede in a supplier contract. Then prepare only the information that feeds that one task. You will learn more about the state of your data from twenty documents you actually use than from a complete map of documents you don't.
What "ready" actually means
For a given set of documents, four things have to be true:
- Findable. The system can reach it without someone emailing a file. A shared drive, a database, an exported folder — the mechanism matters less than the fact that it exists in one reachable place.
- Current. There is one authoritative version, and everyone agrees which one it is.
- Permissioned. Access is controlled by who is asking, not by whether the system happens to have the file.
- Described. Each document carries who produced it, when, and what it applies to.
Nothing on that list is technical. All of it is organisational, which is why it tends to be handed to the wrong people.
The version problem is the expensive one
If three price lists exist and two are obsolete, a person notices the dates and picks the right one. A system will not. It will answer from whichever it read, in the same confident tone it uses when it is correct, and someone will quote a customer from a 2023 rate card.
The fix is unglamorous and durable. Name one owner per document set. Have them declare the canonical version, archive the rest somewhere the system cannot reach, and take responsibility for keeping it current. Do this and your AI project improves. Skip AI entirely and your company still improves, which is a good sign you are working on the right thing.
Most of your best knowledge was never written down
The answers your best people give are rarely in the manual. They are in support tickets, sales email replies, and the message where a colleague explained why this particular customer gets an exception.
Start capturing that in the format it already takes. Collect the twenty questions your team answers most often and write down the actual answers — short, dated, attributed. Thirty good answers will do more for a first system than a two-hundred-page policy document that nobody has opened since the last reorganisation.
Decide what the system may never see
Before connecting anything, draw the line. Salary data, health information, unredacted customer records, matters under legal privilege. Decide what stays out, write it down, and make the exclusion structural — separate storage, separate permissions — rather than an instruction the system is asked to follow.
Then decide how access works. If a supervisor can see a document and an intern can't, that has to stay true when both ask through a chat window. Systems that answer the same way to everyone will eventually answer the wrong person.
Set an expiry date
Old information is the quiet failure mode. A policy that lapsed in 2024 sits in the folder indefinitely and reads as authoritative. Give every document set a review date and a named reviewer. When a document expires, it leaves the pool until someone re-approves it.
What to expect
Preparing data for one workflow takes weeks, not quarters, and the second workflow is faster because the ownership habits are already in place. Judge progress on two things: how often the system's answer is correct, and how often you can trace that answer back to a specific document. The second one matters more early on. An answer you can't trace can't be corrected.
This is the least visible part of adopting AI and reliably the heaviest. It is also where we spend the first weeks of most engagements, because a good model reading disorganised information produces confident, well-written mistakes — and those are harder to catch than obvious ones.