The support lead at a mid-size e-commerce brand sent me a screenshot of their help desk dashboard. 847 open tickets. Average first response time: 14 hours. Customer satisfaction score: 2.8 out of 5. Three support agents handling everything from "Where's my order?" to "This product gave me a rash."
She was exhausted. Her team was exhausted. They were hiring a fourth agent, but with a 6-week onboarding period, that wouldn't help for months. They needed something that could start handling tickets today.
We deployed OpenClaw as a customer service agent. Within two weeks, the open ticket count dropped to under 200. First response time went to 45 seconds. CSAT climbed to 4.1. And the human agents finally had time to handle the complex cases that actually required human judgment.
But it wasn't all smooth. Here's the full story.
The Setup
OpenClaw is our open-source AI agent platform. For this deployment, we configured it as a customer service agent with access to the brand's systems:
- Shopify — order status, tracking info, return processing - Gorgias (their help desk) — ticket management, customer history, macros - Their internal knowledge base — product information, policies, shipping details, FAQ
The agent could read customer messages, look up relevant information across all connected systems, draft responses, and — for certain categories — take actions like initiating returns or sending replacement tracking links.
We spent the first day mapping every ticket category from the past 90 days. The breakdown was revealing:
- 38% — "Where is my order?" (tracking and shipping status) - 22% — Returns and exchanges - 15% — Product questions (sizing, ingredients, compatibility) - 12% — Billing issues (charges, refunds, promo codes) - 8% — Complaints and escalations - 5% — Miscellaneous
The first four categories — representing 87% of volume — were highly repetitive and followed clear decision trees. Perfect for an AI agent.
The Build
Configuration took about eight hours across two days:
Day 1: We built the knowledge base and connected systems. This meant ingesting their entire product catalog, shipping policies, return policies, FAQ, and common resolution paths. We also connected OpenClaw to Shopify and Gorgias via API.
Day 2: We built the conversation flows, escalation rules, and safety guardrails. The key design decisions:
- The agent identifies itself as an AI assistant. No deception. - For tracking questions, it pulls the order status and responds immediately. No human needed. - For returns, it can initiate the process if the request falls within standard policy. Edge cases go to a human. - For product questions, it searches the knowledge base and product catalog. If it can't find a confident answer, it escalates. - For billing issues, it can look up charges and explain them, but cannot issue refunds. Refund requests go to a human. - For complaints, it empathizes, apologizes, and immediately escalates to a human agent with full context. - Any message containing certain keywords (legal, lawyer, allergic reaction, injury) gets instant human escalation.
The Results — Week 1
The first week was a rollercoaster.
The good: The agent handled "Where's my order?" tickets flawlessly from day one. These were 38% of volume, and the resolution was nearly instant — pull the tracking number, share the status, provide the delivery estimate. Customer feedback on these interactions was overwhelmingly positive.
The surprising: Return requests went better than expected. The agent could check the purchase date, verify the return window, and either initiate the return or explain why it wasn't eligible. About 85% of return conversations were fully resolved without a human.
The concerning: Product questions were hit-or-miss. The agent was great for straightforward questions — "Is this product vegan?" "What sizes does this come in?" But for nuanced questions about how products work together or subjective recommendations, the answers were sometimes technically correct but unhelpful. A customer asking "Which moisturizer is best for dry, sensitive skin?" doesn't want a list of every moisturizer that mentions "dry skin" in the description.
We spent the weekend refining the product question prompts. Instead of just searching the catalog, we added decision-tree logic: "If the customer is asking for a recommendation, ask about their skin type, concerns, and budget, then recommend based on the brand's specific guidance." That improved satisfaction scores for product questions by about 30%.
The Results — Month 1
After 30 days of operation and continuous tuning:
- Ticket resolution rate (autonomous): 73% - Average first response time: 45 seconds (from 14 hours) - CSAT score: 4.1 (from 2.8) - Open ticket count: Under 200 at any given time (from 847) - Human agent workload: Reduced by roughly 65% - Cost savings: The brand estimates about $8,500 per month compared to hiring additional agents
The three human agents went from overwhelmed generalists to specialists handling complex cases, VIP customers, and quality oversight. Their job satisfaction actually improved because they were doing more interesting work instead of answering "Where's my package?" for the hundredth time.
What Went Wrong
Let me be honest about the failures.
The tone issue. Early on, the agent's tone was too neutral for the brand, which had a playful, casual voice. Customers who were used to getting responses with emojis and personality suddenly got clinical, professional replies. We had to spend time fine-tuning the agent's voice to match the brand. This isn't a set-and-forget thing — brand voice requires iteration.
The edge case problem. One customer asked for a return on an item purchased 91 days ago — one day outside the 90-day window. The agent correctly denied it per policy. The customer got angry. A human would have made an exception for one day. We added "soft boundary" logic for cases within 10% of a policy limit, where the agent now escalates to a human with a recommendation to make an exception.
The Saturday night incident. On a Saturday at 11 PM, the Shopify API had an intermittent outage. The agent couldn't look up orders and started responding with "I'm unable to access your order information right now" to every tracking request. We didn't have alerting set up for API failures. Twenty-three customers had a bad experience before we caught it Sunday morning. We immediately added health checks and failover logic.
Lessons Learned
1. Start with the highest-volume, lowest-complexity tickets. Get those right first. The quick wins build confidence with leadership and customers. 2. Brand voice is not optional. An AI that doesn't sound like your brand creates a jarring experience. Budget time for voice calibration. 3. Build escalation paths before you need them. The escalation rules were the most important thing we built. A customer who gets smoothly handed to a human is satisfied. A customer who gets stuck with an AI that can't help them is furious. 4. Monitor actively for the first two weeks. We reviewed every single ticket the agent handled for the first 14 days. That's how we caught the tone issue, the edge cases, and the recommendation problem. 5. API reliability is your reliability. If a connected system goes down, the agent goes down. Build health checks, fallbacks, and alerting from day one.
FAQ
Does the AI tell customers it's an AI? Yes, always. The first message includes "I'm [Brand]'s AI assistant." We've found that transparency actually increases trust. Customers set appropriate expectations and appreciate the instant response.
What percentage of tickets still need a human? About 27%. These tend to be complaints, complex multi-issue tickets, VIP customers, and edge cases that fall outside standard policy. The goal isn't 100% automation — it's handling the routine work so humans can focus on cases where empathy and judgment matter.
How long before it was fully operational? Two days for initial deployment, two weeks of active tuning to reach stable performance. After that, it's mostly maintenance — updating the knowledge base when policies change and reviewing escalated tickets weekly for improvement opportunities.
Can it handle multiple languages? Claude supports many languages natively, and we've deployed multilingual agents for other clients. This particular brand only needed English, but the architecture supports it.
What if a customer specifically requests a human? Instant escalation. No argument, no persuasion, no "Are you sure?" The agent says "Absolutely — I'm connecting you with our team now" and hands off with full context.
Your support team deserves to do work that matters.
If your team is buried in repetitive tickets, we can deploy an OpenClaw customer service agent that handles the routine and escalates the rest. Your customers get instant responses. Your team gets their time back.
Start with a Strategy Audit and we'll map your ticket data to build a custom deployment plan.