How Tendo works
Last updated: 25 April 2026
This is the technical-but-friendly version of what happens when a customer support email reaches Tendo. If you'd rather see the nuts-and-bolts of connecting your sending inbox, jump to{" "} Connecting your inbox.
The flow at a glance
- Inbound. Your customer emails your support address. You forward that mailbox to a per-store Tendo address (we generate it during setup). Postmark receives the email and posts it to Tendo as a webhook.
- Sanitize. Tendo strips quoted-reply blocks and signatures so the AI sees only what your customer actually wrote, not a re-quoted copy of every previous round-trip.
- Classify. Claude Haiku reads the cleaned ticket and tags an intent: refund_request, shipping_question, policy_inquiry, product_issue, escalate_to_human, or other.
-
Retrieve. Tendo looks up your knowledge-base
chunks via vector embeddings — return policy, shipping, FAQs.
If the ticket mentions an order number (e.g.,{" "}
#1004), Tendo fetches the order snapshot from Shopify webhooks (financial status, fulfillment, total) so the AI has real context. - Generate + judge. Claude Sonnet drafts a reply, proposes a refund (with a specific dollar amount), or escalates. A second pass — the "judge" — re-reads the draft for accuracy, citation correctness, and policy alignment, then returns a confidence score.
- Guardrail. Hard rules check things the AI doesn't decide: maximum refund cap, daily refund cap, your first-N-actions-require-approval setting, escalation keywords (chargeback, lawyer, BBB), and whether the order even exists. Anything the guardrails reject becomes an escalation, not an autonomous action.
- Approve. The proposed action lands in your inbox. You see the draft, the confidence score, the cited knowledge chunks, and the order context. You approve, edit inline, or reject.
- Execute. Approved replies send through your connected Gmail / Microsoft 365 / SMTP account (your domain, your reputation). Approved refunds go through the Shopify Admin API directly.
- Thread. When the customer replies, the round trip closes back through Tendo's inbound address (set as Reply-To on the outbound) and threads to the same ticket. No new tickets for follow-ups.
What Tendo never does
- Reply or refund without a guardrail check — even on autonomous paths the cap is in effect.
- Train AI models on your data — Anthropic and OpenAI commercial terms prohibit training on API content.
- Send mail on a domain that isn't yours — the From: address is always the inbox you connected.
- Read inboxes — Postmark forwards what merchants explicitly forward to it; Tendo doesn't pull from any IMAP folder.
What Tendo gets wrong
AI is wrong sometimes. Tendo's strategy is: when in doubt, escalate. Most failures we see are over-cautious, not over-confident — the AI escalates a ticket the merchant could have handled with a one-liner. We tune toward that side deliberately.
Edge cases worth knowing: tickets in languages other than English may classify with lower confidence; multi-topic emails ("two questions, one about returns and one about a missing package") almost always escalate; carrier-side damage may escalate when your KB doesn't explicitly cover it.
Where to next
- Connecting your inbox — Gmail, Microsoft 365, or SMTP. Walk-through with screenshots.
- Set up your knowledge base inside the Tendo admin (paste your return policy, shipping FAQ, sizing guide).
- Calibrate your voice — paste 3–5 of your past replies and the AI will match your tone on every draft.