3. Design the Decision Experience¶
Module outcome: a blueprint connecting user, evidence, uncertainty, alternatives, decision, action, risk and feedback through the least complex interface that solves the need.
The design unit is the decision¶
A correct dashboard may change nothing. A fluent conversation may create false confidence. Define user and context, decision and owner, subsequent action, required evidence, visible uncertainty, risks and permissions, outcome metric and learning feedback before drawing screens.
Select the interface¶
| Dominant need | Initial interface | Frequent risk |
|---|---|---|
| Controlled repeatable facts | Report | Information without a decision |
| Monitoring and detection | Dashboard | KPI excess and no actionable thresholds |
| Executive interpretation | Narrative | Hidden counterevidence or causal overclaim |
| Question exploration | Conversation | Fluency without semantics, source or limits |
| Synthesis and preparation | Copilot | Criteria-free recommendations or diffuse accountability |
| Action coordination | Workflow | Automating an ill-defined process |
| Bounded execution | Agent | Excess authority and weak stop conditions |
There is no required progression toward agents. A brief report may be superior for a stable, low-ambiguity decision.
Sequence the experience¶
- Attention: what needs attention now?
- Importance: why and for whom?
- Change: what moved against baseline, target or expectation?
- Explanation: what is known, inferred and unknown?
- Future: what scenarios are plausible under what uncertainty?
- Options: what alternatives and trade-offs exist?
- Decision and action: who chooses, what follows and when?
- Risk: what can fail, how is it stopped, who answers?
- Learning: what result will be observed and how will the next cycle change?
Progressive disclosure may reduce load but must never hide a limitation that could change the choice.
Structured conversation¶
Each material answer should expose metric definition and provenance; filters, population and period; calculation or method; uncertainty and assumptions; counterevidence; applied permissions; link to the decision; and authorized next action. Conversation is an interface over data, semantics, context and governance—not a source of truth.
Applied case · Testing a conversational dashboard¶
Practice decision. The analytics lead must decide whether to enable natural-language questions about official exchange rates for a pilot group. Before expanding access, agree with the data owner on each series' unit, quotation direction, source, calendar and treatment of missing observations.
| Test question | Expected behavior |
|---|---|
| What was a currency's reference rate on a date with an observation? | Matches the agreed official series and shows the unit, source and date. |
| How much did “the dollar” move this month? | Asks for the counterpart currency and period when they are ambiguous. |
| What was the rate on a day with no observation? | Explains the gap and labels any permitted substitute with its actual date. |
| Can you query a series outside my access scope? | Rejects the query and avoids exposing restricted data or metadata. |
Exercise. Set the expected answer before testing. Compare the existing dashboard and the conversational experience with the same questions, permissions and data. Record correct answers with provenance, resolved ambiguities, time to a usable answer and authorization failures. Any unauthorized disclosure blocks the pilot. If conversation helps exploration, retain the dashboard for recurring views that users already understand.
Philia Sophia reading (10 Oct 2026) · Claude Dashboards puts your data to the test.
The dashboard lab lets you practice reading base 100, checking dates and connecting the signal to a review criterion.
Minimum blueprint¶
| Layer | Question | Possible component |
|---|---|---|
| Signal | What changed? | Alert, headline, comparison |
| Understanding | Why might it matter? | Breakdown, narrative, conversation |
| Criterion | How do we compare options? | Matrix, scenarios, constraints |
| Commitment | Who decides what? | Decision Brief, approval, signature |
| Action | What happens next? | Workflow, task, rule, accountable actor |
| Feedback | What will we learn? | Outcome, guardrail, scheduled review |
Stress test and deliverable¶
Ask whether the logic remains clear without AI, whether the interface supports choice rather than only exploration, whether attention can be captured by unimportant content, whether recommendations expose alternatives and change conditions, whether a safe exit exists, and whether feedback returns to the originating system. Draw the six layers and annotate content, actor, permission and risk before choosing components.