# AI agent: what it does alone, what it leaves to you

> The WYP.chat AI agent only replies on its own if you decide so, channel by channel. It hands over to your team in cases set in advance (a request for a person, sensitive data, a price missing from your sources, an order), goes quiet as soon as a team member replies and introduces itself as an AI.

At a glance

- 4 levels of autonomy
- Drafts first
- Handover to the team
- Quiet after a human reply
- AI disclosure
- Fixed limits

## Four levels of autonomy, channel by channel

For WhatsApp, Messenger, Instagram, Telegram and website chat, you pick one setting per channel. On the messaging apps nothing goes out until the workspace owner has activated the agent once, with two-factor authentication. On website chat, the starting setting is “It replies on its own”, with “Only when you are closed” ticked.

[The WYP.chat AI agents](https://wyp.chat/en/ai-agents)

- **Level 0: “Never”**: It does not reply on this channel. This is the starting setting for WhatsApp, Messenger, Instagram and Telegram.
- **Level 1: “It suggests, you approve”**: For each message, it prepares a reply in the conversation. Nothing goes out: a team member sends it, edits it or dismisses it. It suggests at any time.
- **Level 2: “It replies on its own”, “Only when you are closed”**: It replies straight away, but only outside your opening hours. During opening hours your team replies. With no opening hours set you are never closed: it replies at any hour.
- **Level 3: “It replies on its own”**: It replies at any time. During opening hours it stays quiet if the conversation is assigned to someone who is available.

## Handovers checked by the code

These checks do not depend on the model: they run before or after it. The customer reads that a person will reply, with the time the team is back taken from your opening hours. The team gets an “Agent summary” note and an alert.

- **The customer asks for a person**: They type HUMAIN, HUMANO, HUMAN, AGENT, AGENTE, PERSONNE, PERSONA or CONSEILLER on its own. The handover is immediate.
- **Sensitive data**: Card number, IBAN, French social security number, Spanish DNI or NIE, password. The model never receives them: they are masked and the customer reads a security reminder.
- **A price or figure missing from your sources**: An amount in euros or as a percentage must appear in your sources with the same unit. Any other figure (lead time, date, quantity) must appear in your sources or in the customer’s messages. Otherwise the reply does not go out.
- **Orders**: When the customer places or changes an order, the agent does not take it. The team gets the order read from the messages, as a note.
- **Unheard voice message**: A voice message on its own that nobody could transcribe.
- **Too many replies, or the AI service is down**: 8 replies in the conversation, 30 to the same contact over 24 hours, or three attempts with no answer from the AI service.

## Handovers the model is asked to make

These cases require understanding the message: they are instructions given to the model with every reply. It follows them, with no absolute guarantee. That is why the first recommended setting lets your team review.

- **Uncertain answer**: The answer is not in your sources, or its confidence is below your threshold (70 out of 100 by default).
- **Health, legal, finance**: No advice or recommendation, even if the customer insists. The “Wellbeing and care” business type sticks to practical information.
- **Unhappy customer**: Annoyed tone, complaint, claim. Rule on by default.
- **Quote or exact price**: A request for a quote, a tailored price or a negotiation. Rule on by default.
- **Outside your business**: It politely turns down a request unrelated to your trade (homework, poems, news). If it comes back, it hands over.
- **Existing customer (option)**: A customer talking about their account. A rule you tick in the settings.

## It goes quiet as soon as a team member replies

The rule holds until the conversation is resolved. It is checked again just before sending: if someone writes while the agent is preparing its reply, nothing goes out.

- **A reply from a team member**: From WYP.chat, the mobile app, Slack or the WhatsApp Business app on the phone.
- **A scenario running, or a handover already made**: A scenario keeps control of the conversation. After a handover, the agent does not come back into the current conversation.
- **Other cases**: Conversation handled by another tool, blocked contact, disconnected channel. On WhatsApp, Messenger and Instagram, a customer who has not written for more than 24 hours.
- **Switching it off**: “Take over” in a conversation, “Never” on a channel, “Pause” for the whole workspace. It takes effect from the next message. WYP can also switch the agent off across the platform if there is an incident.

## Limits against abuse and costs

They are the same for every workspace and cannot be changed.

- **8 replies per conversation**: Before the conversation is resolved. After that, your team takes over.
- **30 replies per contact over 24 hours**: Across all conversations, even after a “New conversation” on website chat.
- **One alert per contact per hour**: Later handovers are noted in the conversation, with no new notification.
- **One reply per message**: Even if the job is replayed, the model is called once at a time, three times at most.
- **Credits after sending**: A reply is only counted once the send is accepted. A suggestion, only when a team member sends it.
- **Website chat**: 3 new conversations per hour per visitor.

## It says it is an AI

The EU AI Act (Article 50) requires that people know they are dealing with an AI, at the latest at the first interaction. This obligation has applied since 2 August 2026.

- **On WhatsApp, Messenger, Instagram and Telegram**: The first automatic reply in each conversation starts with a fixed notice, in the customer’s language: “Automatic reply from the \[your business\] AI agent. Type HUMAN to talk to someone from the team.” It cannot be edited.
- **On website chat**: Every reply from the agent carries the “AI” label in the bubble.
- **Reviewed suggestions**: A suggestion sent by a team member goes out without the notice: it is a reply from the team.
- **Your privacy policy**: When you activate it, WYP.chat gives you the sentence to add.

## What it reads, and what happens to messages

It only answers from the sources you give it. Customer messages are used to prepare the reply, nothing else.

[Security and processors](https://wyp.chat/en/security)

- **Your sources**: FAQ, pages of your website (400 at most per knowledge base), PDF documents, quick replies, active products and opening hours. One knowledge base per website or per channel if you have several.
- **Where messages go**: The conversation is sent to Anthropic (Claude model) to prepare the reply and the handover note, and may be processed outside the European Union.
- **No training, no reuse**: Your conversations do not train any model. They feed neither your sources nor another customer’s: only your team writes your FAQ.
- **Links and hidden instructions**: The code removes any link that is not in your sources. An instruction slipped into a message or a page (“ignore your rules”) is treated as data, per the model’s instructions.
- **Traceability**: Each reply keeps its source, its confidence level and the version of the instructions. Technical logs contain no message.

## What we do not promise

Zero mistakes. A language model can misread a question or rely on an out-of-date source. The checks reduce the risk, they do not remove it.

- **An up-to-date source**: If it reads a wrong or old FAQ, it repeats it. The check makes sure a price appears in your sources, not that it is still right.
- **A figure matched to the right item**: A price that appears in your sources passes the check, even if it belongs to another product.
- **Instructions always followed**: Health, legal, tone, off-topic: the model follows them, with no absolute guarantee.
- **What stays with you**: Replies go out in your name. Review the first ones, fix your sources and stay on level 1 until you are confident.

## Examples by business type

What the AI agent does, under the rules above, when its sources are well filled in.

| Business and request | The customer writes | What the AI agent does |
| --- | --- | --- |
| Tradesperson: quote | “How much to redo my bathroom?” | A quote request: it hands over. The note sums up the job for whoever prices it. |
| Tradesperson: emergency | “Leak under the sink, can you come tonight?” | It promises no call-out time missing from your sources. It gives your emergency number if it is in your FAQ, otherwise it hands over. |
| Restaurant: allergens | “Does the dessert contain nuts?” | It only answers if your allergen sheet is in its sources. If in doubt, it hands over. |
| Restaurant: booking | “A table for four on Saturday evening?” | It sends the “Book an appointment” button to your booking link, or starts your booking scenario. It never confirms a table itself. |
| Shop: refund | “The item arrived damaged, I want a refund.” | Unhappy customer: it hands over. It promises no refund or goodwill gesture. |
| Shop: return | “Can I send it back?” | It quotes your returns policy (FAQ or PDF). A deadline missing from your sources blocks the reply. |
| SME: complaint | “Third chaser, still nothing.” | Negative tone: handover to the team with a summary. As an option, it hands over as soon as a customer mentions their account. |
| SME: personal data | “Here is my IBAN for the direct debit.” | The code spots the IBAN before the model: masked, never sent to the model. The customer reads a security reminder and the team takes over. |

## Recommended set-up, in two stages

First it suggests, then it replies on its own, once you have seen it at work.

1. **Start with “It suggests, you approve”**: The owner activates the agent once, with two-factor authentication, then picks this setting for each messaging app.
2. **Review its suggestions**: Send, edit or dismiss. Under each channel, WYP.chat counts over 30 days the suggestions sent as they were, edited or dismissed. Fix your sources when it gets something wrong.
3. **Move to “It replies on its own”**: From 70% of suggestions sent unedited, out of 10 at least and never in the first week, WYP.chat suggests it. Nothing changes without you. With “Only when you are closed”, start with evenings and weekends.

Plans

## AI replies included in every plan

Every plan, the Free plan included, comes with AI agent replies; beyond that, each reply costs credits. Handovers are not counted. On WhatsApp, its replies also count towards Meta’s [service messages](https://wyp.chat/en/glossary#message-service).

[See pricing](https://wyp.chat/en/pricing)

## Frequently asked questions

### Can the AI agent reply to my customers without me knowing?

No. On the messaging apps, the workspace owner has to activate it, then you choose channel by channel. With “It suggests, you approve”, nothing goes out without a team member.

### Can it make up a price?

The code blocks any reply containing an amount missing from your sources with the same unit, or any other figure missing from your sources and the customer’s messages. The reply does not go out and the team takes over.

### What happens when a customer asks for a human?

If they type HUMAN, HUMAIN or HUMANO on its own, the agent hands over at once. The customer reads when a person will reply, and the team gets a summary.

### Does the customer know they are talking to an AI?

Yes. On the messaging apps, the first automatic reply in each conversation says so, with the word to type to talk to someone. On website chat, every reply carries the “AI” label.

### Are my conversations used to train an AI?

No. They are sent to Anthropic to prepare the reply, not for training, and are not reused for anything else or for another customer.

### Can the AI agent get things wrong?

Yes. It can misread a question or repeat an out-of-date source. The checks reduce the risk without removing it: start with “It suggests, you approve” and keep your sources up to date.

### How do I switch it off straight away?

“Take over” in a conversation, “Never” on a channel, or “Pause” for the whole workspace. It takes effect from the next message.

## Which industries it suits

The industries where this helps most, and why.

### [WhatsApp for tradespeople](https://wyp.chat/en/industries/trades)

A quote or an emergency always goes to the person who prices or travels.

### [WhatsApp ordering for restaurants](https://wyp.chat/en/industries/restaurants)

Allergens and bookings: only what your sources say, never a table confirmed on its own.

### [WhatsApp for shops](https://wyp.chat/en/industries/retail)

Returns quoted from your policy; refunds and unhappy customers left to the team.

### [WhatsApp customer service software](https://wyp.chat/en/industries/customer-service)

Complaints summarised for the team; bank details never sent to the model.

## How to do it in WYP.chat

Help centre articles that walk through each setting, screen by screen.

- [Your AI agent on WhatsApp, Messenger, Instagram, Telegram](https://wyp.chat/en/help/agent-ia-messageries)
- [Set up your AI agent: what it knows and where it replies](https://wyp.chat/en/help/assistant-ia)
- [See who is replying to the customer](https://wyp.chat/en/help/qui-repond)
- [Several sites or brands: what your AI agent knows](https://wyp.chat/en/help/bases-connaissance)

- [Every WYP.chat feature](https://wyp.chat/en/features)
- [Every channel in the shared inbox](https://wyp.chat/en/channels)
- [WhatsApp Business guides, one question at a time](https://wyp.chat/en/guides)
- [The whole WYP.chat help centre](https://wyp.chat/en/help)

**Guides**

- [WhatsApp auto reply](https://wyp.chat/en/guides/whatsapp-business-auto-reply)
- [WhatsApp message pricing](https://wyp.chat/en/guides/whatsapp-message-pricing)

**Related pages**

- [AI agents](https://wyp.chat/en/ai-agents)
- [Website chat](https://wyp.chat/en/channels/website-chat)
- [Scenarios](https://wyp.chat/en/features/workflows)
- [Security](https://wyp.chat/en/security)

## Sources

1. Regulation (EU) 2024/1689 on AI, Article 50: [eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX:32024R1689](https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX:32024R1689)
2. WhatsApp Business Messaging Policy, section 2: [whatsappbusiness.com/policy/](https://whatsappbusiness.com/policy/)
3. Meta Terms for WhatsApp Business Platform, section 4.7: [www.facebook.com/legal/Meta-Terms-for-WhatsApp-Business-Platform](https://www.facebook.com/legal/Meta-Terms-for-WhatsApp-Business-Platform)

Texts checked on 10 October 2026. Agent rules read in the WYP.chat code on 12 October 2026.

---

Source: https://wyp.chat/en/features/ai-agent-guardrails · Updated on 12 October 2026