Reachly
Tiered AI support deployment for a D2C skincare brand handling 1,200 to 1,800 support tickets a month. AI handles Tier 1 autonomously, drafts Tier 2 for human review, and routes Tier 3 directly to agents, with no bot loops.
Reachly was handling 1,200 to 1,800 support tickets a month with a team of 4 agents. Most tickets were routine: order status, return requests, ingredient questions, and shade matching. But the team had no way to separate these from the harder tickets (complaints, allergic reactions, and high-value wholesale enquiries), so every message got the same queue. Average response time was 18 hours. CSAT was 3.1 out of 5. Two agents were spending most of their day on messages that could have been answered by a well-configured FAQ page.
Week 1: Ticket categorisation
We pulled 90 days of closed tickets from Freshdesk and categorised every one into three tiers. Tier 1 (same answer regardless of who asks, low stakes) turned out to be 61% of total volume. Tier 2 (variable response needed, human judgment required) was 24%. Tier 3 (relationship at risk, escalation, or high value) was 15%. This gave us the scope for the build before writing a line of code.
Week 2–3: Tier 1 automation
We built the Tier 1 knowledge base from 200 resolved tickets, the product catalogue, the returns policy, and the FAQ page. We used a retrieval-augmented generation setup: the incoming message is classified, the relevant knowledge base section is retrieved, and a response is generated and sent automatically. We built in a confidence threshold: if the classifier confidence drops below 0.82, the ticket skips Tier 1 and routes to Tier 2. We also built explicit escalation triggers: any message containing frustration language, a mention of a medical symptom, or the phrase 'speak to someone' immediately routes to Tier 3.
Week 4–5: Tier 2 drafting and handoff logic
For Tier 2, the AI drafts a reply and places it in a human review queue inside Freshdesk. The agent sees the draft, edits if needed, and sends with one click. We measured that agents were editing fewer than 20% of Tier 2 drafts after the first two weeks. For Tier 3, the system routes directly to an agent with the full conversation history attached and a one-line summary of the issue. The agent picks up with context and no requirement for the customer to repeat themselves.
Week 6: Shadow mode and launch
We ran the system in shadow mode for one week, generating responses without sending them, and compared against human responses. Tier 1 accuracy was 94%. We made adjustments to the knowledge base for 11 edge cases, ran shadow mode for two more days, and launched. In the first full month, 79% of tickets resolved without human involvement. CSAT moved from 3.1 to 7.2 out of 10 within 60 days.
"Our support team was burning out doing the same repetitive replies all day. Now they spend their time on the tickets that actually need them. Our customers are happier, our agents are happier, and we are spending less. I genuinely didn't think all three were possible at the same time."
