Every support team has a voice, whether or not anyone wrote it down. It’s in how the owner replies on WhatsApp, in which words the returns page uses, in whether the apology comes first or last. An AI agent doesn’t inherit any of that by default. It inherits the model’s voice — fluent, general, and eager — and if nobody decides otherwise, that’s what customers hear under your name.
Deciding otherwise is called conversation design, and in a large company it’s a job title. In a small one it’s an afternoon, done once and then kept. This is what it covers and how to do it in Anchor.
What conversation design covers
| Area | The question | Where it lives in Anchor |
|---|---|---|
| Voice | Who is talking, and how do they sound? | Agent name, Greeting, Tone |
| Structure | How long is an answer, and what comes next? | Your documents; Playbooks for multi-step flows |
| Scope | What does it help with, and what does it decline? | Scope topics, Handoff topics |
| Handoff | What does the customer see when a person takes over? | “Waiting for a teammate”; the summary your teammate gets |
| Quality | How do you know it’s still right? | Playground tests; the Review queue |
Most of the effort goes into the second row, which surprises people. The settings take ten minutes. The documents are the script.
Start with the voice: three fields
Settings → Agent has three fields that decide the first impression.
- Agent name.The name in the widget header. Not “Bot”, not “AI Assistant” — the name your team would use if they’d hired someone. It sets the register for everything under it.
- Greeting.The first line a customer reads. Make it a sentence that says what it can do: “Hi — ask me about orders, delivery, or returns and I’ll answer from our help center.” A greeting that promises anything (“How can I help you today?”) invites the questions it isn’t for.
- Tone:Friendly, Professional, Casual, or Formal. Pick the one your team already uses on WhatsApp, not the one that sounds most like a bank. A shop that writes “we’ve got you” shouldn’t have an agent that writes “we regret any inconvenience caused”.
That’s the voice. It’s worth doing well, and it isn’t where the work is.
Your documents are the script
A grounded agent quotes you. That’s its safety property, and it has a consequence people miss: if the returns policy is written in legalese, the answer arrives in legalese, with a source link to prove it. The agent can’t paraphrase its way to a warmth it wasn’t given.
So the highest-leverage move in conversation design is rewriting the ten documents that answer the most questions in the voice you want a customer to hear. Compare:
Before:“Returns are accepted within a period of thirty (30) days from the date of delivery, subject to the item being unused, in its original packaging, and accompanied by proof of purchase.”
After:“You have 30 days from delivery to return anything, as long as it’s unused and still has its packaging. Keep your receipt or order number handy.”
Same policy, same facts, and the second one is what you’d say across a counter. Written that way, the agent says it that way. Categories in the knowledge base make the top ten easy to find, and the Content gaps list under Analytics → Worklists tells you which documents get asked about most — which is the order to rewrite them in.
Design the handoff, because it’s the moment customers remember
The most-quoted line in any support transcript is the one where the bot gives up. Design it.
In Anchor, the customer sees a single line — “Waiting for a teammate” — inside the same thread, and the conversation keeps its history. Your teammate sees more: an AI handoff summary of what was said, and Collected details — the order number, the amount, the transaction reference — that the customer already gave, so nobody asks for them twice. Whether the agent keeps answering what it can during the wait is a switch (“Keep assisting while waiting for a human”). For a small team with no night shift: on.
What you design is what leads up to it: the Handoff topics field, written as decisions (“refund requests outside 30 days”) so the agent hands over judgment calls and answers policy, and a greeting that doesn’t promise a person is standing by when one isn’t.
Don’t over-explain
The model’s instinct is to caveat. Yours should be to cut. Three places padding creeps in:
- The greeting.It doesn’t need a disclaimer about being an AI plus a privacy paragraph plus a menu. Anchor has a separate Privacy notice field, shown before a customer shares details, so the greeting can just greet.
- The documents. A paragraph that explains why the policy exists before stating it will be quoted in that order. Fact first, reason after, if at all.
- Scope topics.A long list of everything it may discuss reads as everything is in scope. Name the four or five things it’s for.
Multi-step conversations get a playbook
Some conversations aren’t an answer, they’re a sequence: a return needs the order number, then the item, then the reason, then the instructions. Left to itself the agent will do that in whatever order the customer happens to supply things. A playbook fixes the order: a “use when” condition and a numbered list of steps, which the agent follows. Anchor can draft one from a plain-English description, and you edit the steps until they read like your team would do it.
Getting started without hiring anyone
- Name an owner. One person decides how the agent talks. Usually the person who already answers the WhatsApp.
- Read twenty transcripts.Conversations in the dashboard, newest first. Note every reply you’d have phrased differently, and which document it came from.
- Rewrite those documents. Voice first, then length. Publish.
- Save the questions as tests.Playground → Tests, so the next content change can’t quietly undo the voice you just set.
Small steps compound
None of this is a launch. It’s a document a week in a better voice, a handoff line you’re not embarrassed by, and a tone setting that matches the people behind it. The customer who asked about delivery at 11 PM doesn’t know any of that happened. They just notice that the answer sounded like your shop.
Common questions
How do I change the tone of an AI support agent?
In Anchor, Settings → Agent has a Tone setting with four options — Friendly, Professional, Casual, Formal — plus an Agent name and a Greeting. Those set the register. The bigger lever is the documents themselves: a grounded agent quotes your content, so rewriting your most-asked policies in the voice you want is what actually changes how it sounds.
Why does my AI agent give such long answers?
Usually because the source is long. A grounded agent answers from your documents, and a document that explains the reasoning before the rule gets quoted in that order. Put the fact first and cut the preamble. Keep the greeting to one sentence and use a separate privacy notice rather than stacking disclaimers on it.
What should an AI agent say when it hands a customer to a human?
One clear line inside the same conversation, not a form or a ticket number. Anchor shows "Waiting for a teammate" and keeps the thread open, so the teammate replies in the same place with a summary and any details the customer already gave. Don't promise a person is standing by if nobody is; do let the agent keep answering what it can while they wait.
Do I need a conversation designer to set up an AI support agent?
Not for a small team. Name one owner, read twenty recent transcripts, rewrite the documents that produced the replies you'd have phrased differently, and save the questions as tests so future changes don't undo the work. It's an afternoon, then a document a week.