Is your data ready for AI? Check these five things before you start
The conversation usually starts with the model. Which one is best, which one is cheapest, whether you need your own or can use one that's already out there. It's an understandable place to start, because that's where the news is.
It's rarely where the project gets stuck. When an AI feature disappoints, the cause is almost always further down: the information it needed was in the wrong place, out of date, contradictory, or something it should never have been allowed to see. The model did what it was asked. It was asked to work with data nobody had looked at closely.
The good news is that you can find most of these problems before you spend anything on AI. Here are the five checks we would run first.
1. Can you say where the answers live?
Pick the questions you want the AI to answer, and write down where a person would go to answer each one today. Be specific: not "the CRM" but "the notes field on the customer record, plus whatever the account manager remembers".
This exercise tends to surface three things. Some answers live in a system you can connect to. Some live in documents, email threads or spreadsheets scattered across shared drives. And some live only in people's heads, which no AI tool can read.
That last group matters most. If the honest answer to "where does this live?" is "ask the person who has been here longest", an assistant won't fix it. You either write that knowledge down first, or you leave those questions out of the first version.
2. Is there one version of the truth?
Ask two departments for the current price list, the list of active customers, or the latest version of a policy. If you get two different answers, an AI tool will get both too, and it has no way to know which one is right.
People handle this without thinking. They know the sales team's spreadsheet is always a week behind, or that the PDF on the intranet was replaced last spring. A tool that searches everything treats every copy as equally true and gives a confident answer from whichever one it found.
Before connecting anything, decide which source is the master for each kind of information and retire or clearly mark the rest. It's dull work. It is also the single biggest thing you can do for the quality of the answers.
3. Who is allowed to see what, and will the AI respect it?
This is the check that carries real risk.
Most companies have documents that only some people should read: salaries, contracts, HR cases, board papers, customer details. Often those documents are protected less carefully than everyone assumes. A folder shared with "everyone" years ago for convenience. A file sent to the wrong channel. A permission nobody removed when someone changed roles.
While finding those files took effort, the gaps stayed hidden. An AI assistant that can search across your files removes the effort. Ask it an innocent question and it may quote from a document the person asking was never meant to see, because technically they could open it all along.
So before you switch an assistant on, review who can access what, starting with the sensitive areas. Make sure the tool works with each user's own permissions rather than a single account that can see everything. And decide which data should stay out of reach completely, whatever the permissions say. If your business is in a regulated field, check what your obligations say about where that data may be processed and stored, too.
4. Is the data current, and who keeps it that way?
An AI answer is only as fresh as the data behind it. If product details are exported to a spreadsheet once a quarter, the assistant will quote last quarter's details with full confidence.
For each source, find out how often it's updated and who owns it. If the answer is "nobody in particular", that's the job to sort out first. Then plan how the AI tool picks up changes: a nightly sync, a direct connection, or a manual refresh someone is responsible for. Also plan how something gets removed, such as a withdrawn product or a superseded policy, so it stops showing up in answers.
5. Can you tell a good answer from a bad one?
Before you build anything, collect twenty or thirty real questions your staff or customers ask, with the answers an expert would give. This takes an afternoon with the right people in the room.
That list becomes your test. Run it against any tool or prototype and count how many answers are right, how many are wrong, and how many are confidently wrong, which is the dangerous kind. Run it again every time something changes: a new data source, a new model, a new version of the prompt.
Without a list like this, the project is judged on a demo, and demos are always run on the questions that work. With it, you can tell whether a change made things better or worse, and you have a fair way to compare options before you pay for one.
What a week of preparation buys you
None of these checks needs an AI budget. Most need someone who knows the business, someone who knows your systems, and a few days of honest looking.
What you get back is a clear picture of what an AI feature can realistically do with the data you have today. Sometimes the answer is "a lot, start now". Sometimes it's "fix the permissions and pick a master source first". Occasionally it's "this problem doesn't need AI at all; a better search page or a cleaner report would do". All three are worth knowing before the money is spent rather than after.
Where to start
If you only do one thing this week, do the second half of check 3: find out who can see your most sensitive folders, and fix what you find. You'll want that done whether or not you ever add AI, and it's the one problem that gets worse the moment a capable search tool is pointed at it.
If you're planning an AI feature and want a straight view of whether your data is ready, talk to us.