Mozabrik: cutting Amazon customer support first-response from 4 hours to 3 minutes

Mozabrik is a custom photo-mosaic brand on Amazon. Customers send a photo, get a printed mosaic poster back. Sounds simple. The customer support surface is anything but: every order is custom, every customer has questions about photo quality, sizing, shipping windows, and "will this look right with my photo."

Customer support used to be a 4-hour first-response surface. Now it's 3 minutes. Same SLA on edge cases, just AI takes tier 1. Here's how the stack works, what broke, and what's running today.

The starting problem

Mozabrik scaled from launch to 700+ verified Amazon reviews. The product worked. The customer support load didn't.

Volume pattern: 30-50 customer messages per day, 80% of them tier-1 (shipping status, photo specs, return policy, sizing questions). 20% required actual human judgment (custom artwork requests, photo quality borderline cases, refund disputes).

Handling tier-1 manually meant: 4-hour first-response on average, customer satisfaction soft, occasional 24-hour gaps on weekends. Amazon penalizes slow response. Reviews start mentioning it.

Hiring a part-time CS rep at $20/hour, 4 hours per day = $1,600/month before benefits or scaling. Not the right move at this revenue band.

What we built

AI customer support stack tuned for Mozabrik's specific product surface. Reads the customer message, classifies intent, drafts a response using the product knowledge base, sends if confidence is high, escalates to me if confidence is low.

Stack components:

• Amazon Seller Central message ingestion via API polling (no third-party plugin)
• Product knowledge base built from FAQ + order history + 700+ review patterns
• LLM-drafted response with confidence score per message
• Auto-send threshold tuned per intent class (shipping = high, refund = low)
• Telegram escalation for low-confidence messages with full context
• Weekly summary of edge cases that broke confidence, used to retrain the knowledge base

What broke (first month)

The honest part.

Week 1. Model drafted responses that referenced product variants that don't exist. Fixed by tightening the knowledge base to the actual SKU catalog with explicit refusal pattern on novel variants.

Week 2. Auto-send threshold was too aggressive. Sent a refund acknowledgement on a message that turned out to be a sizing question. Fixed by lowering the auto-send threshold and adding a second-pass intent classifier.

Week 3. Telegram escalation ping rate hit 40 per day during a holiday gift surge. Fixed by adding batching (max 5 escalations per hour) with priority sort by message age.

Week 4. Amazon API polling rate-limited during peak. Fixed by switching to webhook-based intake with API polling as fallback.

By end of month one, the system was stable. Now it handles tier-1 autonomously with first-response under 3 minutes. Escalates to me on novel cases. I read the daily summary in 5 minutes over morning coffee.

What's running today

30-50 messages per day flow through the stack. Median first-response: 3 minutes. Median resolution time on auto-sent responses: 7 minutes. Escalation rate: 15-20% of messages, well-below the original 100% manual load.

Review patterns mentioning support speed shifted from "slow" to "fast" within 60 days. The 4.6 star rating held through the rollout, then climbed.

Most AI customer support gets sold as "deflection." The actual win at SMB scale is faster first-response on tier-1, not deflection. Customers don't want to avoid talking to you. They want to know you heard them in under 5 minutes.

What I'd do differently

1. Build the knowledge base from your own review patterns, not generic FAQ. The 700+ Amazon reviews already showed me what customers ask. Generic FAQ misses the texture.
2. Lower the auto-send threshold for the first 30 days. Escalate more, auto-send less. Get the confidence calibration right before you let the system run hot.
3. Telegram escalation with full conversation context. Email is too slow. SMS is too narrow. Telegram with message threading is what works.
4. Webhook intake, polling fallback. Don't trust polling at peak. Webhook-first or get rate-limited at exactly the moment volume spikes.

What this would cost a buyer

Hiring a CS rep to cover the same surface: $1,600+/month at $20/hour part-time, more at full-time. Plus turnover risk, training, scheduling, and weekend gaps.

The AI customer support stack: $2,500 productized AI Growth Audit to validate fit, then a 4-week Sprint to ship the system. Client keeps the infrastructure. Runs on their own Amazon Seller Central account. No retainer required after launch.