Is my data ready for AI?

It depends on what you want AI to do with it. Using an off-the-shelf assistant to draft an email needs almost nothing from your data. Grounding AI in your own systems — so it answers from your contracts, your tickets, your numbers rather than the open internet — is where data readiness starts to matter, and where most of the value for a business lives. Confuse the two and you will either over-prepare for a trivial use or under-prepare for a serious one.

This is a readiness check for the second case: the point at which you connect AI to your own data.

Ready for what?

There is no single bar for "AI-ready data." There is a bar for each use.

A general assistant working from public knowledge needs no access to your data at all. A tool that summarises a document you paste in needs only that document. But an assistant that answers "what did we agree with this client on renewal terms" has to reach into your contracts, your CRM, and your email — accurately, with the right permissions, and without inventing the parts it cannot find. That is a materially higher bar, and it is the one worth planning for.

So the first move is not to audit all your data. It is to name the questions you want AI to answer, then check only the data those questions actually touch.

The four things that actually matter

For grounding AI in your own systems, readiness comes down to four questions.

Most organisations are stronger on quality than they fear and weaker on access and governance than they hope. Which brings us to the trap.

The permissions trap

The single most common failure when connecting AI to internal data is not bad data. It is over-permissioned data meeting a tool that respects no boundaries.

In most companies, far more people technically can open far more files than anyone intends — years of broad shares, open folders, and "just give everyone read access" have piled up quietly. It causes little harm while access depends on someone knowing a file exists. Point an AI assistant at the same store and that changes: ask it the right question and it will happily surface whatever it is allowed to reach, including the salary spreadsheet in the wrongly-shared folder.

The fix is not to lock down the entire estate before you start. It is to connect AI through a governed layer that enforces each user's own permissions, and to fix the worst oversharing on the data you are actually exposing first, rather than everywhere at once.

A pragmatic readiness check

You do not need a finished data lake, a governance programme, and a two-year cleanup before AI can return value. That counsel of perfection is how these efforts stall.

Start narrow. Take one high-value question, map the handful of systems it draws on, and check those four boxes — access, quality, findability, governance — for that slice only. Fix what the slice needs. Ship it. Then let the next use pull the next slice of data into shape. Readiness earned this way compounds, and it pays its way at each step rather than asking the business to fund a cleanup on faith.

The honest headline: your data is readier than a big-bang project would suggest, and less ready than a careless connection assumes. The work is to connect the right slice, the right way, first.