Start with work that already has a clear input, a clear output, and a cost you can name. Not with the most impressive technology, and not with the process that annoys people most.
When this work disappoints, the technology is rarely the reason. The order is. The right things get done in the wrong sequence, and the sequence is what decides whether anything compounds.
Why order matters more than choice
Two businesses can buy the same tools, spend the same money, and end up in completely different places. The difference is usually not which tool. It is what they did first.
Do the hardest process first and you spend eighteen months proving nothing, while the people who were willing to try lose interest. Do a visible, easy process first and you get a win nobody can build on, because the next process shares none of its plumbing.
Order is the actual decision. It is rarely treated as one.
But you were told to take whole domains
You will read advice to pick one to three complete business domains, claims or underwriting or distribution, and transform them end to end rather than scattering small pilots across the business. That advice is right about the scattering. Scattered pilots are how a programme dies quietly.
It is also answering a different question. Choosing which domain to commit to is one decision. Sequencing the work inside that domain is another, and it is the one almost nobody writes down. You can commit to claims end to end, spend the budget, and still get the order wrong inside it. That is the failure this page is about.
So take the domain advice. Then use what follows to decide what actually happens in week one.
The three questions that rank any candidate
Take every process someone has suggested for AI. Score each one on these three, honestly.
1. Can you state the input and the output in one sentence each?
If you cannot, the process is not one process. It is several wearing a single name, and nothing automated will survive contact with it. Split it first.
2. Do you know what it costs today?
Hours, headcount, cycle time, error rate, anything countable. Without a number from before, you will have an argument after, not a result. The number does not need to be perfect. It needs to exist and to be written down before you start.
3. Does it repeat often enough to matter?
A task done four hundred times a month teaches you something and pays for itself. A task done twice a month teaches you nothing you can rely on, whatever the outcome.
Anything scoring well on all three is a candidate. Anything failing the first question is not ready, and no amount of budget changes that.
Then sort the survivors by what they leave behind
This is the step that gets skipped, and it is the one that makes the difference between a pilot and a programme.
Of the processes that passed, some will leave something reusable behind: a cleaned data feed, a connection to a system, a review step people trust, a way of measuring quality. Others will leave nothing but themselves.
Do the ones that leave something behind first, even when they are slightly less exciting. The second project then costs less than the first, and the third costs less again. That is what compounding looks like, and it is the only way the fifth project ever happens.
The order that produces the best first result and the order that produces the best third result are rarely the same order. Pick for the third.
What not to start with
Not the thing in the board pack
The process with the most visibility tends to be the one with the most exceptions, the most people with opinions, and the least tolerance for being wrong while it learns.
Not a tool you bought first
Buying before you have ranked your processes means the ranking gets written backwards to suit the tool. It happens quietly and it is very hard to undo.
Not everything at once
Three simultaneous pilots teach you a third as much as one finished piece of work, because nobody can tell which change caused which result.
What good looks like after ninety days
One process live, not three in pilot. A before number and an after number that a finance person accepts. One piece of shared plumbing built that the next project will use. And a ranked list of what comes next, with the reason for the ranking written down.
That is an unglamorous quarter. It is also the only version of this that is still working in two years.