Start with the decision, not the data

The most useful question at the start of any analysis is what decision it's meant to inform. Not what data is available, not what model to use. If nobody can name the decision, the work will probably be interesting and then sit in a folder.

Why we start with data instead

Starting with data feels productive. You can open a dataset today, explore it, find patterns, make charts. Starting with the decision means an awkward conversation first, and sometimes finding out the decision has already been made.

LLMs make this pull stronger. It now takes minutes to go from a dataset to a polished set of charts. That's genuinely useful, but it also means it's easier than ever to produce a lot of analysis that answers questions nobody was asking.

Questions that help

Before starting, I find it helps to get answers to a few questions from whoever is asking:

What decision is this for? Something specific: fund this or that, change a process, keep going or stop.
Who makes it, and when? An answer delivered after the meeting is worth very little.
What are the options? If there's only one realistic option, the analysis isn't deciding anything.
What would you need to see to choose differently? This is the most important one. It turns a vague question into a threshold.
How sure do you need to be? Some decisions are cheap to reverse and some aren't. That should set how much rigor the work needs.

If no result would change the decision, the analysis is decoration. That's worth saying out loud, kindly, before anyone spends weeks on it.

This has a name in decision science

Economists and health researchers have a formal version of this called the expected value of information: roughly, how much better would our decision be, on average, if we knew more? It's used in health technology assessment to decide whether more research is worth funding at all.

You don't need the maths to use the idea. If extra information is unlikely to change what you do, it isn't worth much, however interesting it is.

What changes when you work this way

A few things tend to happen:

The scope gets smaller. Most decisions don't need every variable. They need a few numbers, done well.
The bar for precision becomes clear. If the decision flips at 20% and every plausible estimate is over 40%, you don't need a tighter interval.
Bad news arrives earlier. Sometimes the data can't answer the question. Better to find out in week one.
The output looks different. Less "here's everything we found", more "here's the answer, here's how sure we are, here's what would change it".

It's also kinder to analysts

Working without a decision in mind is draining. Analysts keep producing, the goalposts keep moving and nothing they make seems to land. Naming the decision gives the work an edge, and it gives the analyst a reason to push back when the request grows.

The takeaway

Next time someone asks for an analysis, ask what they'd do differently depending on the answer. If they can tell you, you know what to build. If they can't, that conversation is the most valuable thing you'll do that week.