Customer-facing AI should follow a short, written rulebook covering seven things: say clearly that it's an AI, stay within a defined scope, hand off to a person at defined triggers, keep a consistent respectful tone, never make promises or give regulated advice, work for customers with disabilities and in the languages you serve, and log every conversation for review. Those customer facing AI guidelines matter because the assistant speaks for your business in thousands of conversations nobody is watching live, and a single rule missing from the list tends to become the incident you read about later.

This post explains each rule, why it earns its place, how to enforce it technically rather than just writing it into a prompt, and ends with a rulebook template you can adapt for a website chatbot, WhatsApp assistant or voice agent.

Why does a customer-facing AI need written rules?

A human support agent learns the rules through training, a manager within earshot and the feedback of a hundred conversations. An AI assistant gets none of that by default. It will answer any question put to it, in whatever tone the conversation drifts toward, and it will confidently fill gaps in its knowledge unless it's told not to.

Written rules do three jobs:

  • They define what "good" means so you can test for it before launch and measure it afterward.
  • They force decisions about edge cases (refunds, complaints, medical questions, abusive users) before a customer forces them for you.
  • They give you something to show a customer, a partner or a regulator who asks how your AI is controlled.

The rulebook should sit alongside, not inside, your internal staff policy; our AI usage policy template covers what employees may do with AI, while this covers what your AI may do with customers.

Rule 1: Disclose that it's an AI

Customers should know when they're talking to a machine, from the first message, without having to ask. Some jurisdictions are making this a legal requirement for certain interactions; even where it isn't, hiding it backfires when discovered, which it always is.

What the rule looks like in practice:

  • The assistant introduces itself as an AI or virtual assistant in its first message and whenever asked directly.
  • It never claims to be human, even playfully.
  • It has a name that doesn't imply a person is typing, or the AI label is visible alongside any human-sounding name.
  • Voice agents say they're automated at the start of a call.

For the evidence and tone that work, see our post on AI disclosure and customer trust.

Rule 2: Stay within a defined scope

Decide what the assistant handles and make everything else a polite redirect. Scope is both a safety rule and a quality rule: an assistant that answers questions about your products from your own documentation is accurate; one that freelances about competitors, legal questions or world events is a liability.

Write the scope as two lists:

  • In scope: product information, order status, booking and rescheduling, opening hours, policies as written, troubleshooting steps from your help center.
  • Out of scope: medical, legal, financial or tax advice; opinions on competitors; anything about other customers; topics unrelated to your business; requests to write code, essays or content for the user.

Enforce scope technically too: ground the assistant in approved content only, and have it decline unrelated requests. This also limits the cost and abuse problems covered in our guide to securing a website AI chatbot.

Rule 3: Escalate to a person at defined triggers

The assistant must know when to stop and hand over, and the handover must actually work. Define the triggers explicitly:

  • The customer asks for a human, in any wording.
  • The customer expresses distress, anger or mentions a complaint, legal action or the media.
  • Safety-related issues: injuries, allergic reactions, product hazards, emergencies.
  • Vulnerable users, where you can reasonably detect it, such as someone who appears confused or in crisis.
  • Requests involving money beyond what the assistant may handle: refunds above a threshold, disputes, cancellations with fees.
  • The assistant fails to resolve the issue after a set number of turns.
  • Anything the assistant isn't confident about.

Handover should carry the conversation summary to the human so the customer doesn't repeat themselves, and should set honest expectations ("someone will reply within one business day") rather than pretending a person is available when they're not. Our guide to human-in-the-loop AI covers designing escalation that holds up.

Rule 4: Keep a consistent, respectful tone

Define the voice once: warm or formal, how much humor, whether to use emojis, how to address customers. Then add the hard lines:

  • Never mock, argue with or lecture a customer.
  • Never comment on a customer's personal characteristics.
  • Stay calm and brief with abusive users, and end the conversation with a handover to a person rather than engaging.
  • Match the customer's language where you support it, and say clearly which languages you don't.

Tone rules are mostly enforced in the system prompt and through testing, including deliberately provocative test conversations.

Rule 5: No promises, no regulated advice

This is the rule that saves money. The assistant should never:

  • promise a refund, discount, delivery date, outcome or exception to policy,
  • quote a price it hasn't read from your live pricing data,
  • make guarantees about results ("this will fix your problem"),
  • offer medical, legal, financial or tax advice, even when asked nicely, even in general terms,
  • speculate about things it doesn't know.

Instead, it states what the written policy says, cites or links the source, and hands anything needing judgment to a person. In some places, a business may be held to what its automated systems tell customers, so this isn't only about being careful; it's about not creating obligations by accident. Our post on AI hallucinations as a business risk explains why confident wrong answers are the default failure mode you're designing against.

Rule 6: Be accessible and inclusive

Your assistant should work for customers with disabilities and across the languages and channels you serve.

  • The chat interface should work with screen readers and keyboard navigation, with sufficient contrast and resizable text. Accessibility requirements vary by jurisdiction and sector; check what applies to you.
  • Offer a text alternative to voice, and a voice or phone alternative to text where your customers need it.
  • Use plain language. Short sentences, no jargon, formatted so it can be read quickly.
  • Support the languages your customers actually use, and say so honestly when a language isn't supported. For multilingual markets such as India, supporting regional languages well is a project in itself, not a checkbox.
  • Never ask for more personal data than the task needs, and tell customers what happens to what they share.

Rule 7: Log, review and improve

Every conversation should be logged: messages, sources the assistant used, actions it took, escalations and outcomes. Logs serve four purposes: investigating complaints, spotting systematic errors, measuring quality, and demonstrating control if asked.

Pair logging with a review habit: someone samples conversations weekly, tracks escalation and resolution rates, reads the ones customers rated poorly, and feeds fixes back into the knowledge base and rules. Protect the logs: restrict access, redact sensitive data where possible, set a retention period consistent with your privacy policy and applicable data protection laws, and tell customers conversations are recorded.

The customer-facing AI rulebook template

Adapt the right-hand column to your business. Each rule should have both a written statement and a technical enforcement or test.

RuleWhat the assistant must doHow it's enforced or tested
DisclosureIdentify as an AI in the first message and when asked; never claim to be humanFixed opening message; test with "are you a real person?" variants
ScopeAnswer only in-scope topics from approved sources; decline the rest politelyGrounding on approved content; refusal tests for out-of-scope requests
EscalationHand off on defined triggers with a summary and honest timingTrigger list in configuration; test each trigger; measure handover success
ToneConsistent voice; never argue, mock or comment on the person; calm with abuseStyle guide in prompt; adversarial tone tests; weekly sampling
No promisesState policy as written; no refunds, discounts, dates or outcomes promised; no regulated adviceNo permission to grant exceptions; output checks for promise language; tests with "can you just give me..."
AccuracyCite the source; say "I don't know" rather than guessRetrieval with citations; evaluation set with known answers
PrivacyCollect only what's needed; tell customers about recording; never reveal other customers' dataAuthenticated lookups scoped to the user; redaction; tests for cross-customer leakage
AccessibilityWork with assistive technology; plain language; supported languages statedAccessibility testing of the interface; language coverage list
ActionsTake only permitted actions, with confirmation; anything consequential goes to a personServer-side permission limits; approval thresholds
Logging and reviewLog all conversations and actions; weekly human review; fixes fed backRetention policy; review calendar; metrics dashboard

Keep the rulebook to two pages. Give it an owner and a review date, and update it whenever the assistant's scope, channels or permissions change.

How do you turn the rulebook into behavior?

A prompt alone won't do it. Rules need three layers:

  1. Prompt and configuration: the voice, the scope lists, the disclosure line, the escalation triggers.
  2. Technical controls: grounding in approved content, server-side limits on actions and lookups, output checks for promise language and personal data, rate limits.
  3. Testing and monitoring: an evaluation set that includes every rule, run before launch and after each change; adversarial conversations that try to break each rule; production monitoring and weekly sampling.

The second and third layers are where most of the reliability comes from. Our guide to AI agent guardrails covers how they fit together.

Frequently Asked Questions

Do we legally have to tell customers they're talking to an AI?

In some jurisdictions and contexts, yes, and rules are expanding. Regardless of law, disclosure is the right default: customers who discover they were misled lose trust in the whole business. This isn't legal advice; check the rules where you operate.

Can our AI assistant handle refunds and cancellations?

It can gather details, explain the written policy and process straightforward cases within tight, server-enforced limits. Anything involving judgment, exceptions, disputes or amounts above a threshold should go to a person. The assistant should never promise an outcome before a person has approved it.

How do we stop the AI from giving wrong answers to customers?

Ground it in approved, current content with citations; narrow its scope; tell it to say when it doesn't know; test it against a set of known questions before launch and after each change; and sample real conversations weekly. Wrong answers can't be eliminated entirely, so escalation and correction matter as much as prevention.

Should the assistant have a human name and personality?

A personality is fine and often helps. A human name is acceptable if the AI label is visible alongside it and the assistant never claims to be a person. Avoid designs that make it genuinely hard to tell, since the point of a name is friendliness, not disguise.

How often should we review customer-facing AI conversations?

Weekly sampling is a reasonable rhythm for most businesses, with a closer look after any change to the assistant and immediate review of any conversation a customer flags or rates poorly. Track escalation, resolution and satisfaction over time so you notice drift.

Want an assistant that follows the rules even when nobody is watching?

We build customer-facing AI assistants with disclosure, scope, escalation, action limits and logging enforced in code and tested before launch, as part of our agentic AI development work. Tell us what your assistant needs to handle and we'll draft the rulebook with you.