AI Use Cases for Business
Submit
PlaybookBuilt from the public setups listed at the bottom

Approve-to-send support desk: AI drafts every reply, a person clicks send

For stores and small SaaS teams: an AI agent sweeps the inbox, drafts replies from your playbook and requests refunds through approval cards, and widens its scope only on evidence.

Grok BotGorgiasHermes AgentShopify & ecommerceStartupsLocal & service
Setup
Half a day, plus a week of draft-only comparison
Your time
Design target: two or three approval passes a day

This is what the setup aims for, not a measured result.

  1. Write the support playbook from real tickets

    Pull your last 200 tickets and group them by question type. For the top 10, write the ideal answer and the policy behind it.

    Mark what is never automated: refunds over a set amount, health or legal questions, security, and angry customers.

    PromptWritten by our editors
    Read our last 200 support tickets in [helpdesk or inbox].

Group them by question type and show the count for each, largest first.
For the top 10 types, draft the ideal answer in our voice and name the policy or help page it relies on. Say "no policy found" where there isn't one.

List anything that should always go to a person: refunds over $[amount], health, legal or security questions, and upset customers.

  2. Pick the tool that fits your stack

    Shopify stores with order-heavy questions: Gorgias’s AI Agent answers from order data. Orthofeet reports 56% of tickets automated in under two months, a vendor-reported figure.

    Small SaaS or service teams: a Grok Bot connected to your helpdesk, email and Stripe, set up the way Claire Vo runs Holly Helpdesk.

    If you want to self-host: Hermes Agent with a draft-only helpdesk token, which one owner is still testing.

  3. Start draft-only for a week

    The AI drafts; you send. Count how many drafts go out without edits.

    Wherever you make the same edit twice, fix the playbook instead of the draft.

  4. Route money through approval cards

    Refunds and credits reach you as approval cards with the customer, order, amount and reason.

    Never set Always allow on refunds, credits or cancellations.

    PromptWritten by our editors
    You may request refunds and credits but never issue them.

For each one, send me an approval card with: customer, order or charge, amount, reason, and what our policy says.
If the request is outside policy, say so on the card and suggest a reply.
Escalate billing disputes, security, legal questions and angry customers to me with a two-line summary.

    Based on Claire Vo's support Bot, which can request refunds but not issue them.

  5. Review weekly and widen scope on evidence

    Every Monday, have the agent review last week’s tickets and propose three additions to the playbook or help docs.

    Let a ticket type go out without approval only after its drafts consistently go out unedited, and only if your tool supports that safely.

    PromptWritten by our editors
    Every Monday at 8am, review last week's tickets.

Tell me: tickets handled, the share of your drafts I sent without edits (show the count and the division), the 3 question types that took the longest, and 3 proposed additions to the playbook or help docs.
Don't change the playbook yourself.

Guardrails

  • No refund, credit or cancellation without a person.
  • Escalate health, legal, security and angry customers, every time.
  • Keep a readable log of every AI-drafted reply and who approved it.
  • Model per-resolution or usage fees against your ticket volume before switching on automation.

Business fit

  • Shopify & ecommerceStores take 'where's my order' and returns off the team while refunds stay human.
  • StartupsSaaS teams get first-line support covered without a support hire.
  • Local & serviceService businesses answer repeat questions faster without anything going out unseen.

Use cases behind it