What each channel demands of you
The reference page: what you need before you can start, what the customer can do with your reply, and what each channel actually costs.
Every channel looks the same from inside the workflow builder: a trigger on the left, an output on the right. They are not the same. They differ in what you must have before you can send a single message, in what the customer can physically do with your reply, and in what each conversation costs you.
This is a reference page. You are meant to come back to it before you promise a customer a channel, not after.
What you need before you can start
| Channel | You must have | Realistic lead time |
|---|---|---|
| Web chat | Nothing. Paste the widget on your site. | Minutes. |
| A sending domain, verified (SPF/DKIM). | Hours to a day, mostly DNS propagation. | |
| A BSP account, a number that has never been on consumer WhatsApp, and an approved message template for anything you send first. | Days. Template review is a queue you do not control. | |
| Voice | A phone number, a telephony provider, and a voice agent configured with a speech model. | A day or two, longer if the number is being ported. |
Web chat
- You must have
- Nothing. Paste the widget on your site.
- Realistic lead time
- Minutes.
- You must have
- A sending domain, verified (SPF/DKIM).
- Realistic lead time
- Hours to a day, mostly DNS propagation.
- You must have
- A BSP account, a number that has never been on consumer WhatsApp, and an approved message template for anything you send first.
- Realistic lead time
- Days. Template review is a queue you do not control.
Voice
- You must have
- A phone number, a telephony provider, and a voice agent configured with a speech model.
- Realistic lead time
- A day or two, longer if the number is being ported.
The WhatsApp template trap. You may reply freely for 24 hours after a customer messages you. Outside that window every message must use a pre-approved template. Teams design a beautiful proactive onboarding sequence, then discover on launch day that every message in it needs approval, and that one of them was rejected for being promotional.
What the customer can actually do with your reply
| Channel | Links | Formatting | Length | The real constraint |
|---|---|---|---|---|
| Web chat | Yes | Rich | Generous | None. It is the forgiving one. |
| Yes | Rich | Generous | It may not be read for hours. Never use it for anything time-critical. | |
| Yes | Minimal | Short | Read in seconds. A wrong message is seen before you can correct it. | |
| Voice | No | None | One breath | Nothing can be re-read. No link, no price table, no URL. |
Web chat
- Links
- Yes
- Formatting
- Rich
- Length
- Generous
- The real constraint
- None. It is the forgiving one.
- Links
- Yes
- Formatting
- Rich
- Length
- Generous
- The real constraint
- It may not be read for hours. Never use it for anything time-critical.
- Links
- Yes
- Formatting
- Minimal
- Length
- Short
- The real constraint
- Read in seconds. A wrong message is seen before you can correct it.
Voice
- Links
- No
- Formatting
- None
- Length
- One breath
- The real constraint
- Nothing can be re-read. No link, no price table, no URL.
Voice is not chat with a speaker on it. "I've sent you the link" is meaningless mid-call. A number the customer must remember cannot be four digits long. And latency is a behaviour, not a metric — a two-second pause reads as the line dropping, and people start talking over your agent.
What each one costs
Rates from the platform's own billing table. Everything is in credits, and one credit is ₹1 before tax. What matters here is not the individual numbers but the shape:
| Channel | Charged per | Rate | Plus |
|---|---|---|---|
| Web chat | the model call | LLM rates only | nothing |
| email sent | 0.05 – 0.10 credits | the model call | |
| message | 0.10 credits | the model call | |
| Voice | minute, and then some | four line items | transcription 0.15–3.00/min, speech 0.50–24.00/1k chars, the model, and the minutes |
| Any workflow | execution | 0.10 – 1.00 credits | whatever its nodes do |
Web chat
- Charged per
- the model call
- Rate
- LLM rates only
- Plus
- nothing
- Charged per
- email sent
- Rate
- 0.05 – 0.10 credits
- Plus
- the model call
- Charged per
- message
- Rate
- 0.10 credits
- Plus
- the model call
Voice
- Charged per
- minute, and then some
- Rate
- four line items
- Plus
- transcription
0.15–3.00/min, speech0.50–24.00/1k chars, the model, and the minutes
Any workflow
- Charged per
- execution
- Rate
- 0.10 – 1.00 credits
- Plus
- whatever its nodes do
Look at the voice row. It is the only channel billed on four axes at once — you pay to turn speech into text, to think, to turn text back into speech, and for the call itself. A voice conversation is not slightly dearer than a chat one. It is a different order of magnitude, and it is the reason a chatty voice agent is an expensive voice agent.
The design consequence. On chat, a long thorough answer is good service. On voice, the same answer costs more, takes longer, and is harder to follow. Brevity on voice is not a style preference — it is the cheapest and the kindest option simultaneously.
Choosing, honestly
- Start on web chat. No gate, no cost floor, most forgiving. Get the agent right here first.
- Add WhatsApp when customers are already there — and start the template approvals a week before you think you need them.
- Add voice when the conversation genuinely cannot wait, and budget for it deliberately.
- Use email for the paper trail, never for the urgent thing.
Next: whichever door they came through, the agent has to remember what was said two messages ago — and know what it should deliberately forget.
Related lessons
Base rates — what a piece of evidence is actually worth
A face-recognition system that is 99.9% accurate and almost entirely wrong, and a number that sent an innocent woman to prison. Both are the same arithmetic, and it is the arithmetic that decides what any piece of evidence is worth.
ReadConfirmation and survivorship — what you never looked for
Two questions about evidence you did not go looking for. One is a rule you have to discover, and one is a pattern in five famous people — and in both, the thing that would have told you the truth is the thing nobody checks.
ReadLoss aversion, sunk cost and regression — what it costs you
Four questions you answer about yourself rather than about a scenario, and your own answers are the finding. Then the pattern that makes praise look useless and criticism look like it works, whatever you actually do.
Read