--- name: ai-refund-approval-experience type: experience journey: workflow persona: Support lead domain: Customer support technology: AI agent + web console core: "1.0" description: "Experience skill: Refund approval with an AI agent — the best-practice workflow for a support lead in Customer support on AI agent + web console. Use when designing, building, reviewing or writing copy for any step of this journey. Brand-agnostic: pair with any Celestial design skill for the look. Also use it for decks, research readouts, specs and documents about this journey. Outputs: PowerPoint / Keynote / Google Slides decks, Word and Google Docs documents, PDF reports, emails, social posts, charts and diagrams." --- # Refund approval with an AI agent Let an agent draft and run refunds while a human stays in charge of anything irreversible — with a trace anyone can read. **Outcome** — Refunds under threshold run automatically; those above are approved by a human in one decision with full context, and every action is traceable. This skill says *what happens and in what order*. It never specifies colour, type or spacing: describe screens as compositions of Celestial core v1.0 components and patterns, and let the active design skill render them. --- ## Persona — Amara, Support lead > “I’m happy for the agent to do the work. I’m not happy finding out what it did from an angry customer.” Runs a team of 12; approves refunds over a threshold; audited quarterly. **Goals** - Clear the refund queue without losing control - See why the agent proposed each action - Undo or adjust before money moves **Pains** - Black-box automation - Approvals without context - Bulk actions that cannot be reversed **Behaviours** - Batches approvals twice a day - Spot-checks the agent’s reasoning - Exports reports for finance **Access needs** - Uses a screen reader for long reviews - Prefers keyboard shortcuts for approve / skip --- ## Domain — Customer support - Financial actions above threshold require human approval - Every automated action logged with actor, input and outcome - Customer data minimised in logs (no full card numbers) Vocabulary to use: `ticket`, `refund`, `threshold`, `approval`, `trace` ## Technology — AI agent + web console Inputs: Keyboard shortcuts, Mouse, Natural-language instructions - Agent output is non-deterministic — always show inputs and reasoning - Long-running steps must show live progress and be cancellable --- ## Journey (5 steps) | # | Step | Goal | Compose with | Copy | |---|---|---|---|---| | 1 | **Inbox** | See what needs attention, highest risk first | `inbox` | “3 need approval” | | 2 | **Triage** | Label, assign and set priority | `form + chips` | “Suggested: Billing · High” | | 3 | **Approve** | Decide with the agent’s draft and its reasons | `AI approval` | “Approve $54.00 refund” | | 4 | **Run** | Watch it happen and cancel if needed | `progress` | “Running · step 2 of 3” | | 5 | **Report** | Keep an auditable record | `table + export` | “Export CSV” | ### 1. Inbox - Sort by risk, not arrival - Badge items that need a human - **empty:** Nothing to approve — show what the agent handled today ### 2. Triage - Agent suggests labels; human confirms - Suggestions marked as suggestions ### 3. Approve - Show what will happen, to whom and how much - Approve, edit or skip — one decision per item - **error:** If the agent is unsure, say so and ask a question instead of guessing ### 4. Run - Steps listed with live status - Cancel available until money moves - **loading:** Live status per step; announce completion ### 5. Report - Plain-language trace per action - One-click export for finance --- ## Principles 1. Humans approve anything irreversible 2. Show the agent’s inputs, reasoning and confidence 3. One decision per item, reversible until committed 4. Every action leaves a readable trace 5. Autonomy grows with evidence (thresholds), not by default ## Do / Don't **Do** - Label AI suggestions as suggestions until a human accepts them - Offer “Why?” on every proposed action - Provide keyboard shortcuts for approve / edit / skip **Don't** - Bulk-approve without per-item context - Hide failures inside a success summary - Let the agent take irreversible actions above threshold ## Accessibility - Approval cards are articles with headings; shortcuts listed and remappable - Live progress uses a polite live region; errors use assertive - Reasoning text is real text, not an image or tooltip only Core minimums always apply: 4.5:1 text, 3:1 UI and focus, visible focus on every control, 44px touch targets, no meaning by colour alone, reduced motion respected. ## Content and voice - Say what will happen: “Refund $54.00 to Theo’s original card” - Confidence in words: “Likely duplicate charge (2 matching payments)” - Never anthropomorphise: “The agent found…”, not “I think…” ## AI interactions - Show sources the agent used (order, payment, policy) - Ask a clarifying question when confidence is low - Let the human edit the draft before approving - Log model, prompt version and outcome for each action ## Edge cases - Agent and human disagree — record both and the final decision - Payment provider down — pause queue, notify, retry - Partial refund — show remaining balance before approval --- ## Measures | Measure | Target | How | |---|---|---| | Auto-resolved under threshold | ≥ 70% | Refunds completed with no human step | | Approval time | < 30 s | Median time per approval decision | | Reversal rate | < 1% | Approved refunds later reversed | | Trace completeness | 100% | Actions with input, reasoning and outcome logged | Research: Shadow-mode comparison: agent draft vs human decision; Think-aloud review sessions with leads; Quarterly audit sample. --- When building this journey: follow the steps in order, compose each screen from core components and patterns only, keep the persona's access needs and the domain rules non-negotiable, and take every visual value from the active design skill. ## Using this journey in other outputs - **Deck or readout**: one slide for the persona (name, role, quote, goals, pains), one per step (goal, screen, the principle that matters most, its error state), then measures with targets. 16:9. - **Spec or document**: a section per step with goal, screen composition, copy, states and acceptance criteria; principles and accessibility rules as a checklist at the end. - **Look and feel**: this skill is brand-agnostic. If a design skill is named or installed (e.g. "in Mars"), take every colour, font and shape from it; otherwise use plain neutral styling and say which design skill would apply.