If you are building an AI customer service agent, you might be obsessing over model benchmarks when you should be looking at your inbox. A recent case study from a two-location yoga studio highlights a critical truth: the channel determines more about your agent's behavior than the model does. The studio launched the same agent, using the same facts file and model, across Instagram DMs, web chat, and SMS simultaneously. The results were wildly different, proving that technical capability means nothing without channel context.

The Yoga Studio Experiment

On Instagram, the agent handled approximately 300 messages in the first week but failed to reliably send working booking links, a mechanical limitation of the platform. On web chat, conversations died instantly after the first reply, resulting in zero bookings due to silent abandonment. SMS performed beautifully, but only because the studio received just forty messages a week. This disparity wasn't a bug in the code; it was a mismatch between the agent's capabilities and the channel's unwritten social contracts.

Understanding Channel Constraints

Each platform carries specific expectations. DMs feel like texting a friend, where two sentences constitute a full paragraph and latency must be in seconds. Email is treated as a letter, where speed buys no credit. Web chat demands immediate answers or the user closes the tab. SMS is short, urgent, and often unwelcome if abused. When an agent writes email-length paragraphs into SMS, or one-line fragments into email, users perceive the bot as 'bad' even if the underlying model is performing correctly.

Why Multi-Channel Launches Fail

The author advises launching one channel, not four. Multi-channel strategies multiply failure modes, creating separate transcripts, testing requirements, and session handling logic that quickly become unmaintainable. To choose the right starting point, count where enquiries arrive today. If 80 percent of requests come via Instagram DMs, that is your channel, regardless of which platform your engineering team finds interesting. Check for repeatable questions; a channel with 150 repeats of the same twelve queries is far more valuable than one with 500 unique, one-off questions.

Key Takeaways

  • WhatsApp has a 24-hour customer service window; outside it, you can only send pre-approved templates, breaking open-ended follow-up flows.
  • Web chat suffers from silent abandonment where sessions die after the first reply with no identity attached for follow-up.
  • SMS messages split at 160 characters, causing long answers to arrive as fragmented, out-of-order stacks.
  • Never port one script across channels unchanged; packaging must live in the system prompt to match channel norms.

The Bottom Line

Stop treating the channel as a delivery detail. It is half of your agent's behavior, and choosing it first prevents you from discovering your platform's limitations in production.

Action Plan

Pull last month's enquiries and tag them by channel to end internal debates. Select the single channel with the highest volume of repeatable questions. Write that channel's specific constraints, such as length limits and link behavior, directly into the system prompt. Build your facts file against real transcripts from that channel, not your imagination, and schedule the decision to add a second channel for thirty days out, after a month of rigorous transcript analysis.