Answering questions on three channels
The same question arrives on WhatsApp, on the website and through a form. Two of the answers come from one agent wearing different clothes; the third has no agent in it at all.
A person who wants to know what documents they need for a loan against property will ask wherever they happen to be — the WhatsApp thread they already have open, the chat box on the website, or by tapping through a form somebody sent them. Three doors into the same business, and the answer had better be the same at all three.
One brain, two doors
Open the two agents side by side and compare what they are pointed at. They read the same knowledge base, through the same model connection. Nothing about the WhatsApp one knows it is on WhatsApp except its trigger and its reply.
| Trigger | A WhatsApp message arrives | A visitor types in chat |
| Knowledge base | The customer handbook | The same one |
| Reply | A WhatsApp Output node | The chat window itself |
| Live? | Draft | Active |
Trigger
- A WhatsApp message arrives
- A visitor types in chat
Knowledge base
- The customer handbook
- The same one
Reply
- A
WhatsApp Outputnode - The chat window itself
Live?
- Draft
- Active
This is what to say when a client asks for “an AI on WhatsApp too”. They are not buying a second assistant, they are opening a second door onto the one they have. It costs a trigger, a reply node and a conversation about what that channel is allowed to say — which is the next section, and it is the part that actually matters.
The same business, different rules per door
Both agents are told what they may not do. The lists are not the same, and the difference is not sloppiness — it is the channel changing what is safe.
## Constraints
- Never promise approval, never quote a rate for them personally, and never say an
amount will be sanctioned.
- Never ask for a PAN, an Aadhaar number or an account number over WhatsApp.
- Never say insurance is required to obtain a loan. It is not, ever.
- Say you are an assistant if you are asked. Do not claim to be a person.## Constraints
- Never say approval is assured, never quote a specific rate as a promise, and
never say insurance is required to get a loan. It is not.
- You may give indicative bands and realistic timelines, described as
indicative.
- Do not ask for documents, PAN or Aadhaar numbers in chat.Read the two together and the reasoning shows through. The web agent may give indicative bands, described as indicative; the WhatsApp one may not quote a rate at all. WhatsApp forbids asking for a PAN or an Aadhaar number outright, because a WhatsApp thread is a place people forward things and screenshots live forever.
A partner should be able to open this screen in a compliance conversation. The rules a client's channels have to follow are written in English, per channel, in a box their own team can read and argue with. Nothing here is a property of the model.
A tool the agent fills in itself
Both agents have a second tool hanging off them besides the knowledge base — a CRM Manager node. Look at what it is configured to write:
4.3 · Answer website chat questions with AI
Four boxes, and two of them are not steps at all — they hang off the agent's tool handle, and run only if the agent decides to use them.
Scroll for all 4 steps →
Its payload is {{ai}}, which is the whole point. The workflow fixes what kind of thing may be created and leaves what goes in it to the agent — so the visitor's name and number arrive as a contact without anybody writing a form. The agent decides whether to call it at all, and a run where it never did is a normal run, not a failure.
On the WhatsApp side the same handle carries a hand-over task instead, used when the question is about somebody's own case, when the answer would be a rate, or when they are cross. Read its Reading a node's configuration instructions and you can see exactly when it fires — it is a sentence, not a setting.
The third door is not a conversation
The last of the three has no agent in it anywhere, which is worth noticing, because it arrives on the same channel as the first one:
4.2 · Qualify and book inside WhatsApp
A structured WhatsApp form comes back answered. From there it is a lookup, a fork, and writes — nothing decides anything.
Scroll for all 8 steps →
You have read this shape already. Look them up, fork on whether they were found, create or update, and converge — the same skeleton as Capturing a lead from any other system. The person filled in a form rather than typing a question, so nothing has to be interpreted and nothing should be.
“Is it an AI workflow?” is the wrong question. Two of these three answer on WhatsApp and only one of them uses a model. Reach for an agent when the input is a sentence somebody wrote; reach for a form and a fork when it is a field somebody filled.
Two of these are drafts, and that is not unfinished
The web one is running. The two WhatsApp ones are drafts, and they are drafts for the same reason: nobody has connected a WhatsApp business number to this org. Everything else about them is done.
This is the distinction from Draft and active: which workflows are actually running, now with a face on it. A partner who says “the WhatsApp automation still needs building” is wrong and undersells the work. It is built. It is waiting on a number the client owns, and a client who wants it live this month can have it live this month.
Change one thing — read what a draft is waiting for
- Open
4.1 · Answer WhatsApp questions with AIand click the node that sends the reply. - Read what it tells you it needs. Write that sentence down — it is the entire gap between this and a live channel.
- Now open
4.3 · Answer website chat questions with AIand look at the equivalent point in the flow. There is no such node: the chat window is the reply. - Open both agents' prompts and find one instruction that exists in one and not the other. Decide whether you agree with it.
- Change nothing else, and leave both as you found them.
Try it
- Ask the live web agent something the handbook does not cover, and watch it decline rather than guess.
- Write the two-sentence answer you would give a client who asks “can it also answer on Instagram?”. The shape of your answer is the point of this lesson.
Next: Booking the meeting, and chasing the no-show — the conversation has gone well, and now somebody has to be in a room at a particular time.
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