• Jan 14, 2026
  • 4 min read

The Discovery Questions I Ask Before Any Data Project

Started as a 2022 list, rewritten with what I ask now.

Most data projects that disappoint don’t fail in the code. They fail in the first meeting, when everyone nods at “we need a dashboard for sales” and nobody asks what anyone will do differently once it exists.

Back in 2022 I kept a list of questions for working with product teams, split by role: analyst, BI, data scientist, consultant. Re-reading it, the roles didn’t matter much. The questions did. What I’d do differently now: group them by what they uncover, and start with the decision. This is that version.

1. The problem

  • What happens today without this? Who does it by hand, how long does it take, and what goes wrong?
  • What did the last mistake cost? A missed reorder, a late invoice, a bad hire. If nobody can name one, the project is a nice-to-have.
  • Why now? Something changed: growth, a new tool, an audit, a person leaving. That reason sets the urgency.

2. The users

  • Who will open this, and how often? Daily operators and a monthly board meeting need different things.
  • Where do they look today? A spreadsheet, an email, a phone. Meet them there before inventing a new place.
  • Who will not use it, and does that matter?

3. The decision

This is the section that saves the most time.

  • Which decision does this support? Name it as a sentence: “decide which customers to call this week”, not “visibility into customers”.
  • What would they do differently if the number were high? If it were low? If the answer is “nothing”, it’s a report nobody needs.
  • What’s the smallest output that supports that decision? Often one number and a threshold, not a dashboard.

4. The data

  • Where does each number come from? System, table, owner.
  • Who owns it, and who can fix it when it’s wrong?
  • How fresh does it need to be? Real time is expensive. Yesterday is usually enough.
  • What’s sensitive? Personal data, salaries, client names. This decides where it can live and who can see it.

5. Delivery

  • What exists that we should use? The BI tool they already pay for, the database already running. New tools need a reason.
  • What’s the first version? Something usable in days, not a full platform in months.
  • Who maintains it after I leave? If nobody, it has to be simple enough not to need anyone.

6. Success

  • How will we know it worked? A measurable change: hours saved, errors avoided, response time.
  • When do we check? Agree a date for a short review now, while everyone still remembers why they asked.
  • What would make us stop? A project with no exit condition never ends.

7. If there’s a model involved

  • What’s the simplest rule that would do? A threshold or a lookup often beats a model, and anyone can explain it.
  • How will we judge it? Against what baseline, on which data, by whom.
  • What happens when it’s wrong? Who sees the mistake and how it’s corrected.
  • Who could be treated unfairly by it?

From questions to a one-page spec

The answers fit on one page, in three parts, which I now write before any build:

PartComes from
Problemsections 1 and 2: what hurts, for whom, and what it costs
Integrationssections 4 and 5: which systems, which data, which tools
Outcomesections 3 and 6: which decision, and how we’ll know it worked

I call it a PIO (Problem–Integrations–Outcome). It’s short enough that a busy person reads it, and specific enough that a developer, or a coding agent, can build from it.

Takeaways

  • Ask about the decision first. Everything else follows from it.
  • “What did the last mistake cost?” sorts must-haves from nice-to-haves.
  • Agree how success is measured before building, not after.
  • Write the answers down on one page, and build from that page.