Pick the channel before you pick the platform

A two-location yoga studio rolled its AI agent out on three channels in a single week: Instagram DM, web chat, and SMS. On Instagram it answered around 300 messages in the first week but could not reliably send a working booking link. On web chat, most conversations went quiet after the first reply and nothing was booked. On SMS it performed beautifully — on about forty messages a week, because almost nobody texts the studio.

Same agent, same facts file, same model. Three different outcomes. That is not a coincidence, and it is the reason I tell every client to choose the channel before the platform. The channel determines message length, latency expectations, what media you can send, whether a conversation even has an identity attached, and what "normal" looks like to the customer. The model determines none of those.

The channel sets the rules of the conversation

A model behaves the same everywhere. Customers don't. Each channel carries its own unwritten contract:

• DMs feel like texting a person. Replies are expected in seconds, and two sentences is a full paragraph.
• Email is a letter. Nobody expects a reply in two minutes, so an agent answering in seconds buys nothing.
• Web chat is "answer me now or I close the tab."
• SMS is short, urgent, and unwelcome when abused.

An agent that writes email-length paragraphs into SMS, or one-line fragments into email, is behaving correctly by its own lights and badly by the channel's. That mismatch — not model quality — is what most "the bot is bad" complaints actually are.

What each channel is bad at

The honest list. Every channel on this table works for something; what matters is whether its weakness matches your tolerance.

SMS. Messages split at 160 characters, so a long answer arrives as a stack of fragments, often out of order. No reliable buttons or rich media on many devices, links render unpredictably, and every message costs the customer attention — and possibly money. There are also consent rules around texting customers that vary by jurisdiction; check the current local requirements before you build. Best used for short, structured exchanges: confirmations, reminders, one-question-at-a-time replies.

WhatsApp. Rich media works well and customers send photos freely, which makes it excellent for things like "send me a picture of the damage." The trap is mechanical: outside the 24-hour customer-service window, a business can only send approved template messages, not free-form replies. Agents and flows built assuming open-ended outreach break the moment a conversation goes quiet for a day. The window shapes your whole escalation design, not just your marketing.

Web chat. The one channel where the visitor is already on your site with intent — and the one where silent abandonment is worst. A customer gets the first answer, closes the tab, and the session is dead. There is usually no identity attached, so you cannot follow up even when the conversation was five seconds from a booking. Treating "the agent replied" as success is the classic web-chat error; the customer leaving mid-conversation is the failure that never shows up in a containment dashboard.

Instagram and Facebook DMs. This is where discovery happens, and the threading is chaos: comments plus DMs, multiple parallel conversations, people writing in fragments with typos and voice notes. The inbox mixes sales enquiries, spam and support, and the agent answers all of it in the same voice. Sessions have no clean beginning or end, which makes measuring anything honest very hard.

Email. Structured, permanent, good for detail-heavy queries and record-keeping. But it is a slow channel by mutual agreement. Real-time agent pricing bolted onto a channel where the customer expects a reply tomorrow is paying a premium for speed nobody wanted. Where email does shine is as the destination for summaries and confirmations from faster channels.

The comparison table

Start with one channel, not four

Multi-channel is where agents go to become unmaintainable. Each added channel multiplies your failure modes: different session handling, different length rules, different tone expectations, separate transcripts, separate testing. One channel done properly beats four done approximately, every time.

How to pick the one:

1. Count where enquiries arrive today. Pull last month's enquiries and tag each with the channel it came through. This takes an hour and ends the argument. If 80 percent arrive as Instagram DMs, Instagram is your channel, regardless of which one the team finds interesting.
2. Check that the volume is repeatable questions. A channel with 500 unique one-off questions is worse than a channel with 150 repeats of the same twelve. The repeats are what a facts file eats.
3. Match the weakness to your tolerance. If your bookings die from silent web-chat abandonment and you have no follow-up path, that channel needs more human supervision at launch. If you can live with fragmented SMS answers, SMS is nearly maintenance-free.
4. Write the channel's constraints into the prompt. Maximum answer length, when to send links, when to escalate. The same facts file powers every channel; only these lines change.
5. Read transcripts weekly for a month, then add channel two. Not before. The first channel is where you learn what your customers actually ask; carrying that knowledge into channel two is most of the work.

The four mistakes I see most

1. Launching everywhere at once. Three channels in week one means three sets of bugs, no baseline, and no idea which channel is underperforming or why. The yoga studio from the opening couldn't even tell which channel was working, because all three had changed at once.
2. Choosing the channel the team prefers. The owner likes email; the customers live in DMs. The enquiry log settles this — opinions don't.
3. Porting one script across channels unchanged. An SMS-length answer reads curt in email; an email-length answer shatters in SMS. Same facts, different packaging — the packaging lives in the prompt.
4. Measuring all channels with one number. A single containment figure averages away the web-chat abandonment problem entirely. Report per channel, and read transcripts before believing any of it.

What to do today

• Pull last month's enquiries and count them by channel. Put the numbers in front of the team.
• Pick the single channel with the highest volume of repeatable questions.
• Write that channel's constraints — length, media, link behaviour — into the system prompt.
• Build the facts file against that channel's real transcripts, not your imagination.
• Book the channel-two decision for thirty days out, after a month of transcript reading.

The channel is not a delivery detail. It is half of your agent's behaviour, and the half your platform vendor will happily let you discover in production. Choose it first, launch one, and let the transcript log — not enthusiasm — tell you when to add the next.