One AI sales agent,
seven channels, anti-ban limits per channel
For a sports nutrition brand we built an AI sales agent that answers customers in LINE, Telegram, Meta, Instagram, TikTok Shop, Shopee and Lazada from one brain: a 3,665-line playbook with 14 dialogue scenarios and 46 objections, a closed tool list, per-channel anti-ban limits, upsells and follow-ups, order alerts to Telegram. 846 unit and 48 integration tests. Live in production on LINE.
The task
Customers write in seven places. Managers answer in working hours in one language. The brand wanted sales conversations handled around the clock in the customer’s language, with real prices, upsells and a human available at any moment, without getting the accounts banned.
How we built it
The playbook first. Five AI agents in parallel wrote the chapters (product, audience, dialogue scenarios, objections and FAQ, voice and compliance), an assembler merged them, and a reviewer agent checked every line against the catalogue and the regulatory rules. Result: 3,665 lines, 14 scenarios, 46 objections, 62 FAQ, 63 corrections and 19 contradictions removed. The owner answered 27 open questions, 10 of them critical, before any code.
One brain, many channels. A Claude-based agent with a closed tool list and the playbook as its system prompt. Channel adapters for LINE, Telegram, Meta and Instagram, TikTok Shop customer chat, Shopee and Lazada. Each adapter knows its platform’s reply windows and rate limits, so the agent cannot be banned for being too eager.
Sales logic. Qualification, product recommendation from the catalogue, upsells based on real basket data, scheduled follow-ups, hand-off to a human with full context, order alerts to a separate Telegram bot for the team.
Testing. 846 unit tests and 48 integration tests, replayed real conversations, measured hallucination rate and escalation rate before going live.
The result
The agent went live on LINE, the brand’s main channel in Thailand, and the other channels are connected as the platforms issue keys. Managers now handle escalations and approvals instead of every message.
FAQ
How do you stop it from inventing prices or promising stock?
The agent has a closed list of tools: look up a product, check a price, create a lead, book a follow-up. Prices and stock come from the catalogue database, never from the model's memory. A reviewer agent checked every price in the playbook against the catalogue before launch and found 25 to verify.
What is anti-ban protection?
Every platform has rules on when and how often a business may message a customer: reply windows, rate limits, push quotas. The agent tracks these per channel and per customer and will not send what the platform would punish.