Discovery audit
Async review of your last ~500 tickets, your help-desk macros, and your refund policy. Output: a written brief on what the agent will and won't own.
Tier 1 offer
AI Support Triage Agent
The first agent every DTC brand should hire. It reads inbound tickets, drafts on-brand replies, looks up orders, issues refunds within your guardrails, and only escalates the edge cases your team should actually see.
$4,997 one-time ~10 days to live 30-day tuning included
What's included
Six things you get for $4,997
Brand-voice spec
Distilled from your top-rated past replies. You sign off on tone, length, and signature moves before a single live ticket is touched.
Production agent build
Connected to Gorgias, Zendesk, or Shopify Inbox. Order lookup via Shopify Admin API. Refund + return actions behind a configurable threshold.
Eval harness
Blind A/B against your team's historical replies. We don't ship until win rate clears 60% with neutral-or-better CSAT.
Shadow-mode launch
Agent drafts every reply but a human approves for the first 48 hours. Auto-send rolls out per-category as confidence clears the bar.
30-day tuning
Weekly written report. Prompt + macro adjustments. Integration patches. We respond to escalations in your Slack within 24h.
Timeline
Day-by-day
- Day 1 Discovery call + data access
- Days 2 to 4 Voice spec drafted, approved
- Days 5 to 7 Agent build + eval harness
- Day 8 Shadow mode in production
- Day 10 Live on a portion of inbound
- Days 11 to 40 Tuning, expansion, weekly reports
How we work
Our process, in one screen
Async by default
One 30-min kickoff and one 30-min launch review. Everything else lives in a shared Notion and Slack channel. No standups, no status meetings.
Eval-first
We don't ship a prompt change without an eval. You see win rates and regressions before anything touches a real customer.
Your data, your code
The agent runs on your cloud (or ours, billed at cost). Conversation logs, prompts, and eval datasets are yours. No data goes back to model training.
Service FAQ
Specifics, before you book
If your question isn't here, ask on the call. We keep them short.
Which help desks do you support?
Gorgias, Zendesk, and Shopify Inbox out of the box. Front, Help Scout, and Kustomer take an extra ~3 days. Custom in-house tools: talk to us, we've integrated with stranger things.
What about refunds, returns, and order edits: does the agent take action?
Yes, with guardrails. The agent can issue refunds up to a configurable threshold, push return labels via Loop or ReturnGo, and edit non-shipped orders. Anything beyond the threshold drafts a reply for human approval.
How do you train it on our brand voice?
We pull a sample of your top-rated past replies, distill them into a voice spec (length, formality, signature moves), and run blind A/B evals before launch. You sign off on the voice spec before a single live ticket is touched.
What does the 30-day tuning window actually look like?
A weekly review of every escalation, every flagged reply, and the deflection vs. CSAT trend. We adjust prompts, add macros, and patch integration edge cases. You get a written summary each Friday.
Where does the agent run, and who owns the data?
Compute runs on Anthropic's API (Claude Sonnet 4.7) plus a thin orchestration layer we deploy to your cloud or ours. You own the conversation logs, the prompts, and the eval dataset. No data is used for model training.
What happens after the 30 days?
Three options. Self-serve: you take it from here, we hand off docs and code. Light-touch maintenance: $1,500/month for monitoring + monthly tuning. Growth retainer: $25K/quarter to add more agents and workflows.
See it on your help desk
Bring 10 anonymised tickets to the call. We'll show what the agent would have replied.