Aeons lets AI send 63% of support messages.
AI CSAT reached 4.39 out of 5, above the 4.33 team-wide average. Richpanel closed 529 conversations with no human touch while higher-risk cases kept a human review layer.
100,000+ customers. Health-adjacent questions. No room for careless AI.
Aeons needed speed without giving automation permission to improvise on sensitive topics.
- Repeat topics dominated the inbox. Delivery, subscriptions, returns, and product questions followed predictable patterns.
- Some questions carried real risk. Ingredient, dosage, complaint, legal, and financial topics could not be treated like ordinary order status.
- Draft-only AI left the queue intact. Aeons needed AI that could send and close, not merely suggest.
Routine email runs autonomously. Higher-risk replies get human review.
Richpanel gives Aeons two operating paths, both governed by the team’s SOPs.
- Autonomous for repeat work. Delivery, subscriptions, returns, and product questions can be answered and closed by AI.
- AI-led with human review for risk. Richpanel gathers context and drafts; a person reviews legal, financial, complaint, and sensitive cases before send.
- One policy layer governs both. Refunds, replacements, brand voice, and escalation boundaries are encoded up front.
529 conversations closed with no human touch.
Another roughly 850 conversations were AI-led and human-reviewed, keeping the safety layer exactly where Aeons wanted it.
- 63% of outbound messages sent by AI. Richpanel produced 5,753 of 9,138 support messages in the measured window.
- AI CSAT reached 4.39 out of 5. The team-wide average was 4.33.
- 529 conversations closed autonomously. No human touch was needed.
- Roughly 850 more were AI-led. A person reviewed and sent the higher-risk replies.
- About five FTE of capacity returned to the team. The output matched five full-time agents combined.
The first failure was tagging, not a customer reply.
The AI began creating near-duplicate tags, which distorted reporting. A knowledge audit caught the pattern, and Aeons restricted the system to an approved tag list.
- The issue was visible in operations. Similar tags split one category into several reporting buckets.
- The fix became a rule. The AI now checks an approved list before applying a tag.
- Reporting returned to a clean state. The same control now governs future tagging work.
Five FTE of capacity went back to the team.
The 5,753 AI messages roughly match the output of five of Aeons’ full-time agents combined. That capacity moved to work the team could not reach while repeat email filled the day.