Guide

Data Practice Readiness

Answer nineteen questions and find out which of four maturity stages you are in — and what to ask IT for.

Last reviewed About 45 minutes, as a team Free — no sign-up

Your IT organization can say yes to almost anything now. That is the problem.

Tooling is no longer the constraint — a Fabric capacity, a Copilot license, an agent framework are each a purchase order away. What is scarce is a clear, agreed account of what your data practice should look like and in what order to get there. Without one, every conversation with IT is a ticket: one tool, one workspace, one permission, queued behind every other team's requests. With one, you arrive as a partner carrying a plan.

So the data and business teams have to lead the asks — not wait to be enabled. Turn up with an as-is picture everyone in the room recognizes, a to-be picture everyone has agreed to, and a clean split between what you will do yourselves and what you genuinely need from IT. That is what gets IT on board, and it is what this assessment is for.

How to use it

  • Answer as a team, in one room — data engineering, BI and analytics, and at least one business owner. Separately-filled copies disagree with each other, and that disagreement is itself the first finding.
  • Three answers: No · Partly · Yes. For the staging, Partly counts as No. Scoring generously defeats the instrument — it is only worth running if it can tell you something you did not want to hear.
  • Budget about 45 minutes for the nineteen questions, and longer for the roadmap at the end.
  • Every question states what a Yes looks like, concretely. A question without that bar gets answered optimistically.

The questions

Nineteen, in four groups. Each is tagged with the maturity stage it is evidence for. Your answers stay in this browser — nothing is sent anywhere.

0 of 19 answered

A Data engineering

All six are evidence for Stage 1.

A1 S1 When the business brings a new use case or data source, can you bring it on board in under a week?

A Yes looks like: A new source lands and reaches a business-ready table along a known path, without a project being opened for it.

A2 S1 When a source system changes, do you find out from monitoring rather than from a broken report?

A Yes looks like: A schema or volume change raises an alert before a user notices anything.

A3 S1 Do you have logical checks that alert when the data stops making sense — not just when a job fails?

A Yes looks like: Rules on the content, not the run. An invoice with no lines. A serial number shared by twenty-five unrelated customers. A 40% week-over-week move in a measure that never moves.

A4 S1 Can the data steward who owns that data correct the issue themselves, without a developer?

A Yes looks like: A named business owner per domain, with a way to fix the record at source or override it, and no ticket in between.

A5 S1 Do you have a dependency map and a failure policy for your pipelines — which failures can be tolerated, and which must stop everything downstream until fixed?

A Yes looks like: It is written down, and when a given load fails someone can immediately say which reports are now wrong.

A6 S1 Do you know what each pipeline costs to run, and would you notice if one doubled?

A Yes looks like: Capacity consumption attributed to workloads, and a review or alert that would catch the regression in days rather than at renewal.

B Semantic models and shared meaning

B1–B3 are evidence for Stage 2; B4–B6 for Stage 4.

B1 S2 Do you have shared, domain-focused semantic models — Rental, Equipment Sales, Service, Parts, Finance — rather than one model per report?

A Yes looks like: A report author's default is to pick an existing model. New models are rare and deliberate.

B2 S2 When someone asks “what does this metric mean?”, is the answer written into the model itself?

A Yes looks like: The definition travels with the measure, so it appears everywhere the measure is used — not in a side document that goes stale.

B3 S2 If a user has a question about a number — where it came from, what logic was applied — can they answer it themselves?

A Yes looks like: Lineage and the transformation rule are visible to the person asking, with nobody from the data team in the loop.

B4 S4 Could you point an AI tool — Claude, Copilot, a Copilot Studio or Foundry agent — at those models and get consistent answers for your key domain metrics?

A Yes looks like: You have actually tried it, on real questions, and checked the numbers against a known-good source.

B5 S4 If two people ask an AI the same question in different words, do they get the same number?

A Yes looks like: The model constrains the answer, so phrasing cannot change it.

B6 S4 Do the models enforce who can see what, so an AI answering a question cannot leak past the asker's permissions?

A Yes looks like: Row- and object-level rules live in the model and apply through every path to it — including the agent's path, not just the report's.

C Self-service and tooling

C1–C4 are evidence for Stage 2; C5–C6 for Stage 3.

C1 S2 Can an analyst or citizen developer change the dashboard or app they use every day, without filing a ticket and waiting?

A Yes looks like: The person closest to the work makes the change, the same day.

C2 S2 Are those changes review-gated before they reach production — without the review taking weeks?

A Yes looks like: A pull request a peer approves the same day. Not a release board that meets monthly, and not nothing at all.

C3 S2 Can users build their own reports on trusted models, instead of exporting to Excel?

A Yes looks like: You could show from usage data that exports have declined.

C4 S2 Are models, reports and AI settings in source control, so you can see who changed what and roll it back?

A Yes looks like: Git integration on the workspace. A bad change is reverted in minutes, not rebuilt from memory.

C5 S3 Does everyone who could use these tools know they exist, and how to use them safely?

A Yes looks like: Onboarding and refreshers aimed at the business, not only at the data team.

C6 S3 Can a business user publish something other people rely on, inside guardrails that stop them breaking it?

A Yes looks like: A defined place they are allowed to publish to, with the certification, permissions and review that make it safe for someone else to trust.

D The business test

Evidence for Stage 4 — and the only question spoken in the business's voice.

D1 S4 If you knew the exact agent you needed for your workflow — a marketing content manager, a quoting assistant, an ordering specialist — would you have the framework, the data and the technical setup to build it? Even if you were willing to pay for it out of your own budget, could you?

A Yes looks like: The second sentence is the whole question. Cost is the assumed blocker and is usually not the real one. If the honest answer is “I would fund it tomorrow and still could not get it built”, then everything above this line is the reason.

Where you land

You are at the highest stage where every question at or below it is a Yes. A floor, not an average — and Partly counts as No.

Answer all 19 questions and your stage appears here, with the gap to the next one.

19 to go.

Why a floor and not a score

A score would let one successful AI pilot average out a warehouse nobody trusts. Maturity here is a foundation ladder and the rungs are load-bearing: business users cannot safely self-serve on models that disagree with each other, and an agent cannot give consistent answers over data whose definitions live in people's heads. The lowest unmet rung is the one that matters — it is where the work goes, regardless of how advanced anything above it looks.

Turning it into a roadmap for IT

Three moves. The third is the one that changes how the meeting goes.

  1. Write the as-is in business terms, not deficiencies

    For each No or Partly, one sentence of consequence. “A source change broke the month-end rental report twice last quarter” lands; “gaps in pipeline monitoring” does not.

  2. Pick one to-be stage — the next one

    A roadmap asking for Stage 4 while Stage 1 is failing reads as a wish list, and wish lists go in the queue. One rung at a time is what makes it credible.

  3. Split the asks

    Separate what your team will do from what only IT can provide. This is what turns a request into a plan.

The asks that are genuinely IT's

Use this as a prompt list — a missing one of these is what stalls a whole roadmap.

Capacity
A Fabric SKU or trial capacity, and who carries the compute cost
Tenant settings
Which Fabric, Copilot and agent features are switched on at tenant level
Identity
Entra groups, guest accounts, service principals, and who administers them
Source control
An Azure DevOps or GitHub organization, and permission to connect workspaces to it
Licensing
Power BI Pro or PPU, Copilot Studio capacity, agent licensing
Data access
Gateways, private endpoints, and credentialed access to the source systems

What stays yours

Model design, metric definitions, data-quality rules, steward ownership, publishing standards, training, and the review path. Claiming this is how you earn the asks.

The table to fill in

Gap Business consequence What we own What we need from IT By when

If you can run this yourselves and agree the answers, do — that is the point of giving it away. We are worth more to you on the build than on the conversation.

Not ready to answer all nineteen? The four maturity stages explain what each one takes, and the workshops page covers all three.

Take it with you

Printable, and built to be filled in around a table. No email required — the point of a guide is that it is useful before anyone buys anything.

Download the printable workbook (PDF, 5 pages)

Change log

This page is maintained. It changes when the practice changes, and every revision is listed here.

2026-10-29
First version. Built from a thirteen-question draft: the alerting and steward-correction question was split in two, the lineage question moved to the semantic-model group, and six questions were added — pipeline cost, model-enforced permissions, awareness and training, guarded publishing, and the closing business test. Every question gained a “what a Yes looks like” bar.

Last reviewed 2026-10-29. Something out of date? Tell us.