When is your data AI-ready?

In this article:

• What is AI-ready data?
• Why definitions are the real work
• Half of your data isn't in a database
• Why organisations rarely manage this on their own
• How we approach an AI-ready data project

What is AI-ready data?

Your data is AI-ready when an AI agent answers a question about your business as reliably as your best employee would. That is the bar we hold every data project to at Blis Digital.

Rajeev: "It's AI-ready when an agent can give you the same answer as your best employee."

That employee knows which calculation applies, which business rules hold, which dataset to use and for which period. They built up that knowledge over years, and most of it lives in their head. That is exactly the knowledge you need to unpack and record before a model can take it over.

Ask a model to write an email and a bit of guesswork is fine. Ask it for a number and you expect an exact answer. And a model states a wrong number just as confidently as a right one.

Rajeev once spoke with a director who didn't know how many people worked for him. His right hand spent 3 days calculating it. Once the number was there, the questions started. Do interns count? Temporary staff? All entities? Put AI on top of that and the calculation goes from 3 days to a few seconds. You still don't know what you're measuring.

Why definitions are the real work

Take a simple question: how many customers do we have?

Sales counts the active opportunities in the CRM. The CFO looks at who was invoiced this month. The contact centre counts who called in. All three are correct.

Rajeev: "And then you end up with five different definitions living across the organisation."

It also hides in questions that sound harmless. "How much revenue did we make this year at our building on the Coolsingel?" That question contains 4 assumptions: what counts as a building, which locations fall under Coolsingel, whether revenue is before or after discounts, and whether this year means the financial year or the calendar year. Without those agreements in place, the model guesses.

You don't need one definition for the whole organisation, by the way. Finance and sales can each keep their own 'lens', as long as you record which lens belongs to which question. Give that to your model, and an agent asked "give me today's customers" knows the question comes from finance. It can also state which lens it used.

Rajeev: "Then you're not constantly explaining why the numbers differ from each other."

Half of your data isn't in a database

Everything above is about structured data: your CRM, your ERP, your accounting system. That is the data you can unlock relatively easily and apply those definitions to.

The rest sits on SharePoint, in folders and on laptops. Word files, PowerPoints, PDFs and Excel sheets. The same problem shows up there in a different form. A document that matters can easily exist in 20 versions.

Rajeev: "You might still be able to tell which version is the latest, but which one is the right one?"

A database takes care of that. Outdated records get a status or go to the archive, so you always know the current state. In a folder structure you have a copy of a copy of a copy. There too, someone has to record which document is valid and how a new version replaces the old one.

What helps is pulling structure out of it. One organisation we spoke with has people working full-time on figuring out which permits need renewing, spread across hundreds of PDFs. You can use AI to put that information into a structured database. Then you get an answer with a source and a date again, and the agent doesn't have to guess which paragraph it came from.

Why organisations rarely manage this on their own

Agreeing on what a contract is doesn't sound like rocket science. Yet three months later, that definition is still in draft. Someone wearing a finance hat sees a customer differently from someone wearing a sales hat, and both are right from the perspective of their own process. It gets stuck in back-and-forth and yet another version.

Rajeev: "The magic is in getting those five people who have an opinion on this around the table, getting them to talk to each other and asking the hard questions."

That is our role as co-driver. We keep asking questions for an annoyingly long time, weigh the interests against each other and make the technical consequences visible. A wish can make perfect sense in terms of content and still turn out too expensive technically. When that happens, we say so.

How we approach an AI-ready data project

We start with a blueprint for one process or data domain. In workshops with the stakeholders, we map how that process really runs, which data points it creates along the way and which systems they live in. We record definitions, KPIs and exceptions in a data model. Only then do we look at technology.

Rajeev: "The smaller, the better."

If there's already a data platform, great, we'll work with that. If there isn't, a small database with the data for that one process is enough for your first proof of concept. That keeps the investment manageable, and you quickly see which problems are technical and which need a business decision first.

Then we build. With permission, we record the workshops and capture the decisions and definitions in a second brain. Our engineers can query it during the build, so the context from that workshop stays available when someone wants to know weeks later what exactly the client meant by a contract.

We stay involved until the client's own team is working with it. Definitions don't stand still. At one client, a new sales manager inherited KPIs that had been set up 20 years earlier. He couldn't steer his organisation with them, so they were revised. Your data model has to be able to handle a change like that.

Start with one question

Pick one question that 3 departments answer differently today. Record what the terms mean, which source is leading, which exceptions apply and who owns it. Then build a small pilot that shows whether AI gives that answer reliably.

That is where AI-ready data begins.

Listen to or watch the Blis Bytes episode on AI-ready data

• YouTube: [add link]
• Spotify: [add link]