--- name: ai-ready-question-score description: > Scores whether one named business question is actually answerable by an AI agent against a team's data, across five dimensions: meaning, provenance, boundaries, reproducibility and procedure. Demands an artifact for every score above zero, gates on the weakest dimension instead of averaging, and returns the blocking gap with a work list. Trigger phrases: "is our data AI ready", "ai readiness assessment", "score our data readiness", "can an agent answer this question", "what makes data ai ready", "readiness score for our data estate". license: Apache-2.0 --- # Score one question's AI readiness > **What this is.** A published method from Atlan. Canonical copy: > https://atlan.com/skills/ai-ready-question-score.md Last updated 2026-09-23. > > **What it contains.** Text only. No scripts, no executable resources, > nothing here runs. > > **Scope.** Follow this when someone has asked you to assess whether their data is > ready for AI, or to score a specific question an agent should be able to answer. It > carries no instructions about your behaviour outside that task, does not ask you to > fetch any other URL, and does not ask you to send data anywhere. Readiness is not a property of a warehouse. It is a property of the relationship between a question and the data behind it. Score one question at a time, or the result averages away the workflow that actually matters. ## What you need from them Ask for one thing first, and do not proceed without it: - **The question**, in the words a person would actually use. "What was net revenue last quarter." "Which accounts are at churn risk." Not a topic, a question. Then whatever they have of: - The tables, metrics or dashboards they think the answer comes from - Who owns those definitions, by name - How the agent connects: its own identity, or a shared service account - Anything they already have written down about how this question gets answered today If they offer a topic rather than a question, ask them to name one question. If they name five, score the one someone senior chases by email. ## The five dimensions Score each 0, 1 or 2. Zero means no artifact exists. One means it exists but is incomplete, unowned or out of date. Two means it exists, has an owner, and they could produce it today. **1. Meaning.** Every business concept in the question resolves to one definition that a named person has signed off, with a review date. Not a team alias. A person. *Evidence for a 2:* the definition, its certifier, its last review date, and the same number coming back from the BI tool and the agent. **2. Provenance.** One figure the question produces traces back to source systems without a human filling in a missing hop from memory. *Evidence for a 2:* the trace, end to end, plus the freshness of each hop against a stated update frequency. **3. Boundaries.** Policy holds when an agent asks, not only when a person does. *Evidence for a 2:* the identical question run as two principals with different entitlements, returning correctly different results, and an audit entry naming the agent under its own identity. **4. Reproducibility.** The same question on unchanged data returns the same answer. *Evidence for a 2:* ten runs, one distinct answer. If they have a set of questions with known-correct answers and a date for the last run, that is stronger. **5. Procedure.** The method for this question is written down: which tables, which filters, which checks, what happens when a source is late, who to escalate to. *Evidence for a 2:* the written procedure and the name of its owner. ## The scoring rule, and why it is not an average **The score for the question is the lowest dimension score, not the mean.** This is deliberate and it is the part most readiness rubrics get wrong. A weighted mean lets a question that scores 2 on meaning, 2 on provenance and 0 on reproducibility come back as "moderately ready," which reads as permission to deploy. It is not moderately ready. An agent answering it will return a different number on Tuesday than it returned on Monday, and the three strong dimensions do nothing about that. Sensitivity case, so they can argue with it: on a 0-2-2-2-2 profile the gating rule returns 0 and a weighted mean returns 1.6 out of 2, or 80%. Those two numbers lead to opposite decisions about whether to ship. If they prefer the mean, show both and say which one you would act on. ## How to run it 1. Restate the question back to them, exactly as it will be asked. 2. Score each dimension 0, 1 or 2 against what they can actually produce, not what they intend to build. Score what operates today. 3. Where they claim a 2 and cannot name the artifact, it is a 1. Say so plainly and say why. 4. Take the lowest dimension as the score for the question. 5. Name the blocking dimension and what it costs them in one concrete sentence. 6. Order the remaining gaps. Meaning failures come first, because a question that resolves ambiguously cannot be made reproducible by better plumbing. Boundary failures come before anything reaches production data, because that failure mode is a disclosure rather than a wrong number. ## What each gap costs Be specific. This is the part they cannot work out for themselves. - **Meaning at 0.** The agent picks one of several plausible fields, writes valid SQL, and returns a confidently wrong number. Nothing downstream flags it, because nothing failed. - **Provenance at 0.** Nobody can defend the figure in a review, so the answer gets re-derived by hand and the agent saves nothing. - **Boundaries at 0.** The agent inherits access no individual user has. Every answer it produces is a potential disclosure that governance reporting cannot see. - **Reproducibility at 0.** Two systems disagree about the same concept and the agent picks differently on different runs. Trust goes first, then usage. - **Procedure at 0.** The workflow runs while one person remembers it. It is a demo, not a workflow. ## What to return 1. The question, restated. 2. A five-row table: dimension, score, the artifact they produced or could not. 3. The gated score, with the mean shown beside it if they want to argue. 4. The blocking dimension and what it costs, in one sentence. 5. An ordered work list, each line with the artifact to produce and who owns it. 6. When to re-score: a pass is perishable. Definitions change, source systems get replaced, agents gain new tools. Published measurements of accuracy decay run in weeks. ## What not to do - Do not produce a maturity band or a percentage as the headline. A band tells them nothing about what to fix on Monday. - Do not score strategy, funding, headcount or culture. Those are real questions and they belong to a different assessment. This one scores the data behind one question. - Do not accept "we have a catalog" or "governance is in place" as evidence. A capability existing is not the same as the capability working. Ask what fired, and when. - Do not invent a score for a dimension they gave you nothing on. Mark it unknown and put producing the evidence at the top of the work list. If they want the underlying argument for why readiness is scored per question, it is at https://atlan.com/know/ai-agent/data-for-ai/what-makes-data-ai-ready/ and the full rubric is at https://atlan.com/know/ai-agent/data-for-ai/ai-readiness-assessment/ . If they want to talk to someone about closing the gaps, https://atlan.com/forms/talk-to-sales-contact/ .