Chatbot conversation-design template

This is a design template for one chatbot intent at a time — reset password, track order, cancel subscription, update payment. Fill in a fresh copy for each one. A bot that handles ten intents well beats one that gestures at fifty and loops on all of them. Replace everything in [brackets] before it goes near a customer.

The whole job of conversation design is to make three things good: the path when everything goes right, the path when the user says something you didn't expect, and the path to a human. Skip the second two and you've built a demo, not a support channel.

1. Frame the intent

Answer these before you write a single line of dialogue.

Rule: if the honest answer to row 3 is No, don't design a chatbot flow — design a fast, clean handoff to a human and stop here.

2. Script the happy path

Write the shortest sequence of turns that gets a well-behaved user to a resolved outcome. One question per turn. Never ask for something you can look up yourself.

3. Confirm before you act

• The bot echoes its understanding before any state change: "You want to cancel the [Pro] plan on [account] — confirm?"
• Irreversible or costly actions require an explicit yes — never inferred from silence or a vague reply
• After acting, the bot states plainly what changed and what happens next

4. Fallback that moves forward

Do not ship "Sorry, I didn't understand that." Every failed turn hands the user a next step.

• First miss: rephrase the question and show an example of a valid answer
• Offer 2–4 concrete options mapped to intents you actually support — not "ask me anything"
• Cap retries at 2 on the same step; after that, escalate — don't loop
• Never blame the user ("invalid input"); own the miss ("I didn't catch that — try...")

The failure mode you are designing against is the CFPB's "doom loop": a customer trapped in "continuous loops of repetitive, unhelpful jargon or legalese without an offramp to a human." Around 37% of the U.S. population had used a bank chatbot by 2022, and the regulator's headline complaint was bots with no way out.

5. The human offramp (non-negotiable)

• A human is always reachable — no magic word, no fighting the bot to find one
• Escalation triggers are explicit: [2 failed attempts] · [low confidence] · [detected frustration] · [explicit "agent" request] · [any sensitive/regulated intent]
• The full conversation and any actions taken transfer with the handoff — the customer never re-explains from zero
• If no human is available, say so honestly and give a real alternative (callback, ticket, hours) — never fake availability

6. Voice and tone

Follow your brand voice guide; keep turns short and human. Warmth is not decoration: 64% of consumers say they're more likely to trust AI agents that feel friendly and empathetic, and 63% will switch to a competitor after a single bad experience (Zendesk, 2025 CX Trends).

7. Edge cases and safety

• Out-of-scope questions get a graceful redirect, not a hallucinated answer
• Sensitive topics — billing disputes, account security, cancellations, anything regulated — route to a human by default
• The bot never invents policy, prices, or account facts it can't retrieve
• Prompt-injection and jailbreak attempts are tested before launch
• AI disclosure meets your legal obligations for each region you operate in

8. Pre-launch review — the gate

Ship nothing with an unchecked box.

• Walked the happy path end to end on a real [test account]
• Ran every fallback branch and confirmed each has a forward step
• Triggered every escalation path and confirmed context transferred intact
• Tested with messy real utterances — typos, multi-issue threads, non-native phrasing — not just clean scripted ones
• A named owner signed off, and a rollback / kill switch exists

---

Ship one intent, watch real transcripts for a week, fix what breaks, then design the next. Conversation design is a maintenance habit, not a launch event — the transcripts will tell you where the flow lies to you, if you keep reading them.