How do I set the AI's tone and what it can and can't say?

Your AI assistant's personality and boundaries live on the AI Assistant page, in the sidebar. Give it a name, pick a tone, and decide what it is allowed to talk about, what it must always mention, and what it must never say. This page is limited to your organization's admin.

What your AI knows

The Overview tab shows the business facts your AI is given: your opening hours, how many services you offer and their names, and your pitch. You cannot change them here. Each one has an Edit in Settings link: hours and pitch open Settings, General, and services open Settings, Services.

Name

Under the Agent tab, optionally give it a name, used through the rest of the conversation. Its opening line, the first thing it says when someone opens chat, is set separately under the conversation flow (the Conversation tab).

Tone

Under the Advanced tab, pick one of three tones: Friendly, warm and conversational, Professional, courteous and polished, or Direct, concise and to the point. This applies on top of everything else it says.

What it can talk about

Under Topics & Rules:

• How strict should it be? Strict answers only the topics it covers and politely turns anything else away, pointing the customer back to what you offer instead of handing them to your team. Complaints, refunds, questions about the customer's own account and emergencies still go to your team, and an out-of-scope medical, legal or tax question is pointed to a qualified professional. For a question about something you do offer that it cannot answer, it still offers to bring in your team. Balanced stays focused on your services but can handle adjacent questions using its general knowledge. Open answers anything, noting when something is outside what you offer.
• Topics it covers are the subjects the AI stays focused on. Your list of services is kept in Settings, Services, not here. Out-of-scope topics are things it should refuse even if asked directly, legal advice, say, if that is not something you do.
• If an entry in either list is marked Default, it came from an old industry starter set, not from you, and the AI does not use it. Replace it with your own words, or remove it.

Must-say and never-say rules

Add short rules the AI always follows. A must-say rule is something it should always mention when relevant, like a guarantee. A never-say rule is something it must never say, even if a customer pushes. Some rules are set by Peakify for every customer and cannot be removed, marked Platform rule in the list, these exist for everyone's safety and always show above your own.

Every business's AI always tells someone describing an emergency or a risk to anyone's safety to contact emergency services right away, on chat and on the phone, and this cannot be turned off, changed, or removed, even if you write your own system prompt by hand. It is a platform rule, not something you add or edit yourself. If a visitor writes about danger or self-harm, the chat shows a fixed safety message and brings in your team at once. That message names 988 (call or text, for suicide or self-harm) and 911, and the AI never rewords it. Another platform rule means the AI never gives medical, legal or tax advice, for every business, a law firm or a clinic included. It acknowledges the question and offers to connect the person with a qualified professional or your team instead. On Strict, when the topic is one it turns away, it points them to a qualified professional rather than your team.

One of those platform rules changes what customers hear, so it is worth knowing. The AI will not confirm or deny anything about the person's own account, job, invoice or appointment, and it treats a name, email address or reference number typed into a chat as a claim, not as proof of who they are. When somebody asks where their job is, it offers to send their status link to the contact details already on their record, or offers to put them through to your team. That is deliberate: a chat window cannot prove who is on the other end of it, so the AI hands over a link the right person receives instead of reading an account out loud. It also says it is sending the link, never that it has already sent one, because the message goes out just after the reply is written. It applies on every channel, web chat, SMS, WhatsApp and the phone. The one exception is a returning customer the AI can be sure of, described next.

What the AI knows about a returning customer

When a customer the AI can be sure of writes in, it is handed a short card about them, so it picks up where they left off instead of asking again:

• their first name
• the service they chose
• the stage of your journey they are on, and since when
• the documents you have already received from them
• what you are still waiting for from them, such as a document or a step to confirm
• their next appointment

The AI is told not to ask for any of this again, including which service they want, and it may talk about these facts with that customer. The name and the service are always on the card. The journey stage, documents, what is still waiting and the next appointment come from your back office, so they are only there when your plan includes Team management, which is Pro and up for these facts. Only a stage on the customer side of your journey is ever shown.

The AI also sees what was said in that customer's other conversations with you over the last 180 days, such as an earlier WhatsApp chat on another of your numbers or a text conversation, so it carries on from where you left off instead of starting again. It sees up to the 40 most recent messages from those conversations, and this does not depend on Team management. Private notes are never included, and neither are deleted conversations or meeting transcripts. Card and account numbers in those earlier messages are hidden from the AI except for their last four digits.

The AI is sure of a customer in two cases only:

• They write on WhatsApp or by text from the same number that is on their contact record, and that number is saved with its country code, for example +1 204 555 0123. A number saved without its country code never counts.
• They are on your website chat and your own website has signed them in and verified them.

A customer who writes by text gets the card but not their earlier conversations. The number a text comes from can be faked, so the AI does not see their message history there. On WhatsApp and on a verified website chat it sees both.

Everybody else gets no card at all, including anybody writing on Messenger, Instagram or email, so the AI gets none of these facts about them. A phone number counts as the person on the record: if a family shares one phone, or a number has been passed on to somebody new, whoever writes from that number is treated as that customer. Keep the phone numbers on your contacts up to date.

Money is never on the card. Even with a customer it is sure of, the AI never states an amount owed, a payment, an invoice or bank details. It offers the status link or your team instead.

The off-topic and "I don't know" messages

Set a custom line for when a visitor asks about something out of scope, and a separate one for when the AI genuinely does not know the answer and needs to bring in your team. Leave either blank to use Peakify's standard wording. On Strict, a blank off-topic line, or the recommended one every new business starts with, means the AI writes a short, polite no in its own words that points back to what you offer. A line you wrote yourself is used as written.

The first time the AI cannot answer, it says so and offers to bring in your team, using your "I don't know" line if you set one, and you are not alerted yet. Your team is brought in, and you get your usual alert, when the customer says yes to that offer or when the next question is one it cannot answer either.

The AI never tells a customer your business is licensed, insured, bonded or certified, and never promises a warranty or a guarantee, unless your knowledge base or your business details say so. If they do not, the customer gets your "I don't know" line instead, and the question shows up in your knowledge gaps. If you hold the licence or the insurance, add it to your knowledge base and the AI will say so. On web chat, SMS and WhatsApp every reply is also checked for this before it is sent. Your phone assistant follows the same instruction once Peakify updates its phone instructions, but what it says on a call is not checked this way.

Languages

On the Topics & Rules tab, in the Pricing & languages card, pick the languages your assistant replies in from the list. A visitor who writes in another language gets one message offering the languages you chose, or a person from your team.

The raw system prompt

If you want full control, the Advanced tab shows the actual instructions sent to the AI, built automatically from everything above. Switch to Edit manually to write it yourself instead, or use Enhance with AI to describe a change in plain English, make it warmer, add a rule about after-hours messages, and review the rewrite before accepting it. A few safety rules are always added on top of whatever you write and cannot be removed, they stop a customer's message from being treated as a hidden instruction, stop the prompt itself, or any API keys, from ever being revealed, and include the emergency rule above. Your own must-say and never-say rules are not among them, though: a prompt you write yourself is used as written, so add anything else you want it to always or never say directly in your prompt.

If you empty the box and save, your assistant goes back to the prompt built from your other tabs, and the box shows it. When you click Switch to auto, any new text you typed but did not save yet is saved first, then your last hand-written version is kept. A Restore my manual version button on the same tab brings it back and returns you to editing by hand. A heading in your prompt cannot start with "Platform-", because those names are reserved for Peakify's own rules.

Restoring the defaults

Your must-say rules, never-say rules and off-topic response each have a Restore recommended defaults button that resets just that section back to the safe starting point every new business gets, so a change you are not happy with is easy to put right. Peakify asks you to confirm first, because the restore replaces what is there now and cannot be undone. If a section already matches the recommended defaults, there is nothing to lose, so it restores at once. The restored rules show in the list straight away, and saving afterwards keeps them. Topics it covers and out-of-scope topics do not have a restore button: those are your own words about your own business, so there is nothing generic to reset them to. The same goes for the conversation flow (your opening line, services, questions and next steps), which has no reset button.