An eCommerce founder we work with was spending three hours every night after her kids went to bed answering customer support emails. Orders, refunds, shipping questions. The same types of questions, over and over.
We deployed an agent on Thursday. By Monday it was handling 70% of the support volume.
Her text to us: "I read my daughter a bedtime story last night. First time in months."
That's the pitch. Here's how it actually works.
What "Automated Support" Actually Means
Let's be clear about what the agent does and doesn't do.
What it does:
- Reads every incoming support email
- Categorizes it (order status, refund request, shipping question, product question, complaint, other)
- Drafts a reply using your products, policies, and past responses as context
- Sends routine replies automatically (if you enable auto-send)
- Escalates complex issues to a human with a summary and suggested response
- Tracks which issues are resolved and which need follow-up
What it doesn't do:
- Handle genuinely angry customers who need empathy from a real person
- Make policy decisions (should we refund this? should we offer a discount?)
- Process actual refunds or modify orders (it drafts the communication, a human clicks the buttons)
The goal isn't to replace your support team. It's to handle the 70% of emails that are predictable so your team can focus on the 30% that actually need a brain.
The Math That Makes This Obvious
Most support teams spend their time on a surprisingly small number of question types:
| Category | % of Volume | Automatable? | |----------|-------------|-------------| | Order status / tracking | 25-30% | Yes | | Shipping questions | 15-20% | Yes | | Return/refund process | 10-15% | Partially | | Product questions | 10-15% | Yes (with product data) | | Account issues | 5-10% | Partially | | Complaints | 5-10% | No (escalate) | | Everything else | 10-15% | Depends |
That means 60-75% of incoming support volume follows patterns predictable enough for an AI agent to handle. At an average of 3 minutes per email, that's hours returned every day.
Setting Up the Support Agent
Step 1: Connect to Your Email
Your agent needs access to the email inbox where support tickets land. If you're using a shared inbox (support@yourcompany.com), connect that.
Follow our Gmail integration guide for the OAuth setup. The key scopes you need:
gmail.readonly(read incoming emails)gmail.send(send replies)gmail.modify(label, archive, mark as read)
If you're using a helpdesk like Zendesk or Intercom, OpenClaw can connect to those APIs too. But email is the simplest starting point.
Step 2: Build Your Knowledge Base
The agent is only as good as the information it has. Before it can answer customer questions, it needs:
Product information. What you sell, what it does, pricing, specifications. Export this from your website or product database into a reference file.
Policies. Return policy, shipping times, refund process, warranty terms. Be specific. "We offer refunds within 30 days" is better than "we have a flexible return policy."
FAQ pairs. Take your 20 most common support questions and write ideal responses. These become the agent's training data for tone and approach.
Past responses. If you have a history of support emails, export the best ones. The agent learns your communication style from examples, not instructions.
Put all of this in your OpenClaw workspace:
workspace/
reference/
products.md
policies.md
faq.md
past-responses/
order-status-examples.md
refund-examples.md
shipping-examples.md
Step 3: Configure Triage Rules
Not every email should get the same treatment. Set up triage rules that match your support workflow:
support:
triage:
auto_respond:
- category: order_status
condition: tracking_number_available
action: send_tracking_info
- category: shipping_eta
condition: order_in_transit
action: send_estimated_delivery
draft_for_review:
- category: refund_request
action: draft_response_with_policy
notify: slack_channel
- category: product_question
action: draft_response_from_knowledge_base
escalate:
- category: complaint
action: notify_human_immediately
priority: high
- sentiment: negative
action: flag_for_review
- mentions: ["lawyer", "BBB", "chargeback", "social media"]
action: escalate_urgent
The key insight: start conservative. In the first week, set everything to "draft for review" so you can see what the agent writes before it sends anything. Once you trust its output for a category, flip it to auto-respond.
Step 4: Set Up the Feedback Loop
The agent learns from corrections. When you edit a drafted response before sending, the agent sees the difference between what it wrote and what you actually sent. Over time, its drafts get closer to what you'd write.
Configure the feedback loop:
support:
feedback:
track_edits: true
learning_mode: passive
weekly_report: true
The weekly report is gold. It shows you which categories the agent handles well and which ones still need work. Use it to decide where to expand auto-respond and where to keep human review.
Step 5: Monitor and Adjust
After the first week, check these numbers:
- Auto-response accuracy: How many auto-sent replies were correct? Target: 95%+
- Draft approval rate: How many drafts did you send without editing? Target: 80%+
- Escalation rate: How many emails needed human intervention? This should decrease over time.
- Response time: How fast does the agent reply versus how fast your team used to? This is the number customers care about.
If accuracy is below 90%, your knowledge base needs work. The agent doesn't have enough context to answer correctly. Add more examples, more policy detail, more product information.
Real Numbers From a Real Client
Here's what happened with that eCommerce founder over the first month:
Week 1 (draft-only mode):
- 156 support emails received
- Agent drafted 142, escalated 14
- 89 drafts sent without edits (63% approval rate)
- Average response time: 4 minutes (was 4-8 hours)
Week 2 (auto-respond for order status + shipping):
- 171 emails received
- 78 auto-responded (order status + shipping)
- 72 drafted for review
- 21 escalated
- Approval rate on drafts: 76%
Week 4 (auto-respond expanded):
- 148 emails received
- 103 auto-responded (70%)
- 34 drafted for review (approved 28 without edits)
- 11 escalated
- Effective automation rate: 89%
The founder went from 3 hours of support per night to 20 minutes of reviewing escalations.
Common Mistakes
Not providing enough context. If the agent doesn't know your refund policy, it'll make one up. LLMs are helpful like that. Feed it accurate, specific information.
Auto-responding too early. Run in draft mode for at least a week. The first few days will have embarrassing mistakes. Better to catch them in review than in a customer's inbox.
Ignoring the escalation queue. The whole point of automation is to free up time for complex issues. If you stop checking the escalation queue, you've made things worse, not better.
No fallback for edge cases. The agent will occasionally receive an email it genuinely can't handle. Make sure there's a clear path to a human for those cases.
The Bottom Line
Customer support automation isn't about removing humans. It's about removing the repetitive work that keeps humans from doing the work that actually matters. The empathetic response to a frustrated customer. The creative solution to an unusual problem. The relationship-building that turns a one-time buyer into a repeat customer.
An AI agent handles the predictable 70%. Your team handles the meaningful 30%. Everyone's better off.
If you want this set up for your business without the configuration headache, book a call. We'll connect your support inbox, build your knowledge base, configure the triage rules, and have the agent running within a day. $999, one-time.