In almost every first call about AI, someone says it. "We'd love to do this, but our data isn't ready yet."

Sometimes it is the IT head. Sometimes it is the CFO. It always comes from a good place. Nobody wants to build something on a mess.

But in most cases, the answer is no. You don't need to fix your data first. And waiting for that clean-up is one of the main reasons AI work never starts.

Where the idea comes from

For years, "doing AI" meant training your own model. That really did need a lot of clean, labelled data. If your records were scattered across five systems with three spellings of every customer name, you were stuck.

That world has mostly gone for the kind of work most companies want help with. Reading invoices. Sorting supplier emails. Checking contracts against a list of rules. Pulling details out of claims or onboarding files. Today's models can already read and write. They don't need to learn your business from millions of rows. They need clear instructions, the right context and good tests.

So the question changes. It is no longer "is all our data clean?" It is "do we have enough real examples to tell whether this works?"

What we actually need

To start a pilot, we ask for three things.

A set of real examples. Usually 100 to 200 real cases of the task. Real invoices, real emails, real claims. Messy ones included. Especially messy ones. Ode's CTO, Eddie Siegel, put this well in a recent interview: you need enough real examples to find the edge cases, not a mountain of perfect data.

The right answers for those examples. What should have happened with each one? Often this already exists, because your team did the work by hand. The approved invoice. The reply that went out. The claim decision.

An hour with the people who do the work. They know the odd cases. The supplier who always sends two PDFs. The customer who writes in Arabic and English in the same email. That hour saves days.

That's it. With those, we build tests from your own cases. Every change we make is checked against them. You see the score.

Why messy data helps

This sounds backwards, but it matters. If we only test on clean examples, the system will look great in the meeting and fall over in real life. That is exactly how pilots stall.

Real work is messy. Scans are crooked. Fields are missing. People write "as discussed" and attach nothing. If your tests include that mess, you find out early what the system can handle and what it can't. Then you decide, with numbers, which cases go to AI and which go to a person.

What a data project as a first step costs you

A big data clean-up before any AI work has three problems.

It takes a long time, often longer than planned. Siegel warned against treating a multi-year data journey as a blocker, and I agree.

It is hard to know when it's done. Clean enough for what? Without a real task to test against, nobody can say.

And the business sees nothing while it runs. No hours saved. No money number moving. By the time the data is "ready", the budget and the patience are often gone.

When data really does matter

I don't want to oversell this. There are cases where data is the problem.

If the information the task needs simply isn't recorded anywhere, AI can't invent it. If two systems disagree and nobody knows which one is right, someone has to decide. If the access rules around a system are unclear, we need to sort those out with your security team before anything moves.

The good news is that a short pilot finds these problems quickly. In ten days you learn exactly which data gaps block the task in front of you. That is far more useful than a general clean-up. You fix what the work actually needs, in the order it needs it.

How this works in a pilot

In our 10-day pilot, the first days go on measuring how the task runs today and building the tests from your real cases. Then we build the simplest thing that works, inside your tools, and run it against those tests. By the end you have the scores, a money number and a clear list of anything in your data that needs fixing before full rollout.

Your data stays inside your systems and follows your security rules. Where it needs to stay in the UAE or Saudi Arabia, it does.

Start with what you have

If your team does a task by hand today, you already have the data you need to test AI on it. It is in the inbox, the ERP and the shared drive.

If you have a task in mind, see how the 10-day pilot works. Bring a set of real examples. We'll tell you honestly what they can and can't support.