The slow part of an AI project is rarely the software. It's getting permission to reach your own data. Score readiness against one named use case, ask for read-only access to one system at a time, and tie each stage of funding to the access you've actually been given.
Most organisations think their AI problem is about choosing the right technology. Often it’s about getting permission to look at their own data.
I found this out the practical way. I wanted to test whether a set of manual checks could be automated. These are the checks a team does before taking action on a customer account, each one meaning a search across a different system or public register. So I built a rough prototype in an evening, with no access to any internal systems at all, using only public data.
It answered a handful of the checks outright. It flagged everything else for a person to review. And it didn’t get a single answer wrong.
That’s a small result. But it showed something useful. The approach worked. The software wasn’t the thing standing between a handful of checks and all of them. Access was.
Readiness is about one use case, not "AI"
Many organisations score their AI readiness in general. Data, people, governance, technology, strategy. When I did this honestly, the pattern was clear: plenty of data, and almost no way of turning it into decisions.
A general score like that is interesting but not much use. Readiness means something only when you score it against one named use case. Do we have the data this particular job needs? Can the tool reach it? Is it good enough? Who has to say yes?
It’s also worth remembering who fills in most readiness assessments. It’s usually the people asking for the money. A quick build on real, or realistic, data is much harder to flatter.
Data has the longest lead time
Buying a tool takes weeks. Getting access to the systems it needs can take months. In every AI project I’ve been close to, the software was the quick bit. The hard bit was the conversations about who could see what, and why.
That changes how you should plan. If access is the slowest thing, start asking for it first. Don’t wait until the tool is chosen and the budget is approved.
It also gives you a useful test. If a tool can’t do the job on your own data, that’s a data problem, not an automation opportunity. Better to find that out in week two than in month nine.
Make each approval small
Asking for access to a dozen systems at once is a big decision, and big decisions get deferred. Asking for read-only access to one system, for a defined purpose, with a clear way to switch it off, is a small decision. Small decisions get made.
So ask for access one system at a time. Show what you did with the last one before asking for the next. Tie the next stage of funding to each approval, so nobody is spending money on a build that can’t reach its data.
This is slower on paper. In practice it’s usually faster, because nothing sits in a queue waiting for someone to feel comfortable with a large, vague request.
What to do next
Pick the use case you care about most. Write down the data it needs and where that data lives. Then find out who controls access to each source, and how long it takes to get a yes.
If the honest answer is “we don’t know”, you’ve just found your first job. It isn’t a cost problem or a technology problem. It’s an access problem, and it’s fixable.
Not sure who controls the data your use case needs?
That’s a common place to be stuck. Tell me what you’re trying to build and we’ll work out who needs to say yes first.