Building Trustworthy Data Products Module 1 · Decide What to Build

Name the Consumer and the Decision

Updated

Beginner

What you'll learn

~18 min
  • Convert a vague data request into a named consumer, a decision, and a consequence
  • Use freshness and wrongness cost to size the work before committing to it
  • Recognize the requests that should stay ad-hoc queries rather than becoming data products

The request as it actually arrives

Nobody asks for a data product. They ask like this:

“Can we get a dashboard for circuit reliability?”

That sentence contains no consumer, no decision, and no definition of reliability. Build directly from it and you will produce something reasonable that nobody opens after week three — the most common outcome in this line of work, and an expensive one, because it consumes a month and teaches the organization that data work does not pay off.

Three questions convert it into something buildable. Ask them before you open a query editor.

Prompt first: interrogate the request

The temptation is to hand the request to your workspace agent and ask it to build the thing. Resist that for one more lesson. Switch your workspace agent to Plan mode on your working folder and start with a short plain-English request (if you do not have live exports or files handy, ask it to generate a small Meridian practice sample from the lesson’s scenario first):

I received a request for a "circuit reliability dashboard." Do not design anything yet. Plan the intake questions I need answered before this is buildable, and ask clarifying questions about any unknown context.

Check that the plan or draft covers:

  • Consumer by name or role, and the cadence on which they read it
  • The decision that changes when the number moves, and the grain that decision implies
  • Freshness derived from the decision cadence rather than asserted from an adjective
  • The cost of a wrong value, what checking that requires, and what to do when the answer is “nobody knows”

If the first pass misses how to handle unknowns, follow up with: Add what I should do for each question when the requester answers that nobody knows.

You are using the agent as a checklist that does not get tired or polite, not as a designer. The output is a list you take into a fifteen-minute conversation with the supervisor. That conversation is the actual work of this lesson, and no tool does it for you.

💡The refusal is part of the skill

An agent asked to design will always design. It will produce a plausible schema for a request with no consumer, and the plausibility is the danger — it looks like progress. Getting comfortable telling an agent “do not build yet, help me interrogate this” is one of the highest-leverage habits in this training.

Stop and escalate when nobody can name a reader after a real attempt: back to the requester, not onward to a build. A freshness dispute that survives the anchoring conversation goes to whoever owns the decision, because it is a claim about the meeting, not about the data.

KNOWLEDGE CHECK

A manager asks for 'a real-time view of meter health.' You ask what decision changes, and learn the meter team reviews health in a monthly planning session. What should you conclude?

1. Who reads this, by name?

Not “leadership.” Not “the ops team.” A person or a specific role who will look at it on a known cadence.

At Meridian the honest answer turned out to be the distribution planning supervisor, who runs a Monday morning crew-assignment meeting. That single fact settles a surprising number of downstream decisions: it must be right by Monday 7am, it has to be readable in eight minutes, and it needs to identify circuits, not summarize the fleet.

If nobody can name a reader, you have found something important early. Either the request is exploratory — in which case run the query, answer the question, and do not build anything — or the requester is asking on behalf of someone they have not consulted, which you want to discover now rather than at delivery.

2. What decision changes because of it?

The test is whether you can complete this sentence: “When this number is X rather than Y, we do Z instead of W.”

For Meridian: “When a circuit’s degradation score crosses the threshold, it goes on this week’s proactive maintenance list instead of waiting for a failure call.”

That sentence is doing more work than it looks like:

  • It tells you the output is a ranked list of circuits, not a trend chart. Nobody assigns a crew to a trend line.
  • It tells you the grain is one row per circuit, not per meter or per outage event.
  • It tells you a threshold exists, which means somebody has to own what the threshold is — and that is a decision, not a calculation.
⚠If the sentence cannot be completed

A request that survives “what decision changes?” with no answer is a request for reassurance, not information. Those are legitimate — people want to know the fleet is not on fire — but they are satisfied by a one-time query and a short written answer, not by a maintained data product. Building the product anyway is how organizations acquire dashboards nobody reads.

3. What does it cost when it is late or wrong?

Two separate costs, and they size two different parts of the work.

Late sets your freshness requirement and therefore most of your engineering effort. “Right by Monday 7am” is a weekly batch, which is straightforward. “Right within five minutes” is a fundamentally different and much more expensive system. Ask, and then push back on the first answer — the reflexive response to “how fresh?” is “real-time,” and the reflexive answer is almost never the true one. Anchor it to the decision instead: crew assignments happen once a week, so anything fresher than weekly is engineering nobody will use.

Wrong sets how much of Module 4 you need. A wrong circuit on a maintenance list costs a wasted truck roll — recoverable, annoying, and worth ordinary checks. A wrong number in a rate filing costs a regulatory finding and is worth reconciliation against an authoritative source before every publication. Same effort spent in different places, decided by this one question.

Key takeaway

Three questions turn a request into something you can build: who reads it, what decision changes, and what it costs when it is late or wrong. The second question quietly determines your output shape and grain; the third splits your effort between freshness engineering and correctness checking. A request that cannot answer the second one should be satisfied with a query and an email, not a data product. Lesson 1.3 takes the definition problem hiding inside the word “reliability.”

Search lessons