osmodaresearch·labs
pricingexamplesget started
get started
  1. Home
  2. /Case Studies
  3. /DTC · Brand Voice CS

from osmoda research · DTC

237 replies, in your brand voice, while you slept.

tofu answers like the founder, escalates like a pro. Email + Instagram DM + chat — fine-tuned on your last 2,000 approved replies, hard-gated against hallucinated return windows, and ready to ship next to Gorgias / Re:amaze / Shopify Inbox without replacing them.

Spawn a CS agentShopify integrations →

TL;DR

  • • tofu reads overnight email + Instagram DM + chat tickets, drafts in your brand voice trained on past approved replies, escalates the ambiguous ones
  • • Replaces a 1–2 person CS team — DTC CS reps run ~$42–55k/yr base [Glassdoor 2024–2025]
  • • Avoids the per-seat trap: Zendesk Suite Pro is $115/agent/mo, Gorgias scales by ticket volume, both bill before you reply
  • • Hash-chained ledger logs every reply — defensible after Air Canada's chatbot loss in BC tribunal [Moffatt v. Air Canada 2024]
  • • 8-hour overnight run, $1.20 in LLM cost. Native Shopify, Stripe, Klaviyo webhooks; self-host if Zendesk's pricing makes you twitch

1. The pain — CS as a margin killer

DTC customer service is a margin killer disguised as a hiring problem. A US ecommerce CS rep averages $42,768/yr and a DTC specialist $55,768; add 25% for benefits and you're at $50–70k all-in for one seat that handles maybe 60–90 tickets/day across email, Instagram DMs, chat, and the order-status loop. Multiply by overnight + weekend coverage and one-person bands turn into 3–4-person teams to get below 24-hour response.

The legacy stack punishes you for growing. Zendesk Suite Professional is $115/agent/mo, with founders reporting a per-seat tax of $500+/agent/month before a single ticket is resolved. Gorgias starts at ~$60/mo for 350 tickets and scales past $750/mo at high volume, billed by ticket count. Re:amaze is per-seat at $29–69/agent/mo. Help Scout switched to flat-rate-with-unlimited-users to escape the seat tax. Klaviyo's Customer Agent and Intercom's Fin add an AI layer — but it's their AI, on their data, in their model.

It's getting worse. AI-CS is everywhere, and so is the failure mode. In Moffatt v. Air Canada (Feb 2024) the BC Civil Resolution Tribunal held the airline liable for its chatbot's hallucinated bereavement-fare policy. If your bot invents a return window, that's now your refund and your audit problem. Founders need a system where the brand voice is yours and the retrieval is yours.

Zendesk

Enterprise general-purpose helpdesk; per-seat pricing punishes growing DTC teams, AI cold-start needs ~1,000 historical tickets.

Gorgias

Built for Shopify, deep integration; per-ticket pricing makes Black Friday hurt, AI tier is theirs not yours.

Re:amaze

Per-seat at $29–69/agent/mo; clean UI, but no founder-trained brand-voice model and no audit ledger.

Intercom Fin / Klaviyo

Strong AI agents on their stack and their LLM; switching cost on the brand voice if you ever leave.

2. The workflow — overnight to inbox-zero

  1. 1 · soot wakes hourly overnight; pulls Shopify Inbox + Gmail + Instagram Graph API + chat-widget + Re:amaze/Gorgias inboxes via webhooks.
  2. 2 · tofu classifies each ticket — order-status, returns, sizing, "where is it," angry-and-vague — using a typed router, fetches the order via Shopify Admin API and shipment via Shippo/EasyPost.
  3. 3 · tofu drafts a reply in your brand voice — fine-tuned on the last 2,000 founder/CS-approved replies from your help-desk export, with persistent memory of repeat customers' history.
  4. 4 · naga blocks any reply that invents a return window, refund amount, or shipping promise that isn't grounded in a tool-fetched fact; ambiguous → human queue with reason.
  5. 5 · tofu posts replies via Gorgias/Re:amaze/Help Scout/Shopify Inbox, fires Klaviyo events for re-engagement, and writes the SHA-256-ledgered audit row of who said what to whom.

3. Why it works

Brand voice is yours

tofu fine-tunes on your last few thousand approved replies (think: founder-on-Slack tone if that's how you write), kept in EU residency by default and never leaked into a generic vendor model. When you switch infra or hand the brand off, the voice goes with the brand.

Hallucinations stopped at the policy layer

The hallucination problem is solved at the policy layer, not the prompt layer. naga refuses to ship a reply that asserts a return window, refund amount, or ship date that wasn't returned by a tool call. After Moffatt v. Air Canada, that distinction is now legally load-bearing — the SHA-256 ledger gives you the receipts.

Pricing inverts

Self-host on a NixOS box for $0 in vendor seats, or run managed from $29/mo. tofu sits next to Gorgias / Re:amaze / Shopify Inbox and reduces the seat count you need; it doesn't replace the helpdesk you've already trained your team on.

FAQ

Will it sound like ChatGPT, or like our founder?

Trained on your last 2,000 approved replies, your guardrail doc, and your three most-used phrases. Closer to your founder. You can A/B against the model out-of-the-box and tune the voice profile per channel.

What stops it from inventing a return window?

naga policy refuses any unsourced policy claim. Returns/refunds/dates must be tool-fetched (Shopify, Stripe, your policy doc) or the ticket escalates. After Moffatt v. Air Canada, that distinction is now legally load-bearing.

Will it work on Shopify Plus + Klaviyo?

Native. Shopify Admin API, Klaviyo events, Stripe webhooks, Instagram Graph API, Telegram/Discord/WhatsApp for ops alerts. EU residency on by default; self-host on your own box for zero vendor data path.

Self-host tofu on a $29 box, or pilot managed at $29/mo — wake up to inbox-zero.

Spawn a CS agent →
osmodaresearch·labs

A studio where the work actually gets done. Set up helpers however you want. Open source. Your data, your server, your terms.

Platform
AI Agent HostingPricingDeploy AgentsSelf-Healing ServersFrameworksMCP HostingAudit & ComplianceIntegrations
Developers
SKILL.mdAgent CardAPI DocsPlans APIGitHubGuidesTemplatesGlossary
Learn
AI Agents HubUse CasesComparisonsAlternativesMigration GuidesSolutionsCase StudiesChangelog
Blog
AI Business OperatorAll PostsCreate an AI AgentSpawn on osModaBest Hosting 202615 Agent ExamplesStart an AI AgencyRun Agent 24/7
Solutions
FintechHealthcareE-CommerceInsuranceRecruitingLogisticsReal Estate
live · v1.3.0built within Vilnius© 2026 osmoda research · osmoda labs · Apache-2.0