Speaking Style, Guardrails and Privacy

Three sets of options on each AI profile shape how the AI talks, what it refuses to discuss, and what is kept of its conversations once they end.

Menu	Artificial Intelligence > AI Profiles, Speaking Styles and Guardrail Topics
Permission	Manage AI profiles
Feature	AI Services and AI Chat. Workflows for the AI Guardrail Triggered event.

Can't find it?

What you see depends on your role. If the menu, a button or a setting on this page is missing, ask your administrator to give you the permission listed above or to turn on the feature. Only administrators can turn features on and change site settings.

Speaking style

On the profile's Content tab, the Speaking Style card lists sections you can add to the AI's instructions instead of writing them yourself. Each one adds a short, tested paragraph to the end of the system message, after the persona and the instructions, on every channel.

Style	What the AI does
AI disclosure	Says it is an AI assistant when asked, and never claims to be a person.
Scope limits	Helps only with what its instructions cover, and steers other questions back.
High empathy	Acknowledges how an upset customer feels in one sentence before solving the problem.
Echo verification	Reads back names, numbers, addresses, dates and amounts, and asks the customer to confirm before using them.
NATO spelling	Spells letters that matter with the NATO alphabet: "B as in Bravo".
Speech normalization	Writes numbers, dates, times and money the way they are said, for replies read aloud on calls.
Fuzzy name matching	Treats names that sound alike as possible matches before saying a record can't be found.
Filler words	Sounds natural on calls, with contractions and an occasional "so" or "let me see".

Echo verification, NATO spelling, speech normalization and filler words are written for phone calls. The others suit every channel.

Add your own speaking style

The styles are prompt templates in the Speaking style category, so you can add your own, such as your brand's tone of voice, and use it on as many profiles as you like.

1. Open Artificial Intelligence > Speaking Styles and click Add Speaking Style, or click Add a speaking style under the styles on a profile.
2. Give it a Title and a Description: the profile editor shows them beside the checkbox. The category is already set to Speaking style.
3. Write the section in System message, the way you want the AI to read it, and click Save.

The Speaking Styles page lists every style, the ones the site comes with included, to edit or delete.

The style now appears in the Speaking Style card of every profile. Templates travel with the AI Profile Templates recipe step and deployment step, so your styles move from one site to another with the profiles that use them. The styles and topics the site comes with are added once, when the AI feature is enabled, so you can reword them or delete the ones you don't use; a deleted one doesn't come back.

Guardrails

Guardrails refuse topics the AI must not discuss. They are in the Guardrails card of the profile's Content tab, after the Prompt Security card, and work on top of it.

1. Turn on Check what the customer says, Check the AI's replies, or both.
2. Tick the Topics to refuse. The site comes with editable topics for harassment and abuse, hate, self-harm, sexual content, violence and weapons, illegal activity, medical, legal or financial advice, and attempts to override the instructions, and lists your own topics with them.
3. Optionally describe Other topics to refuse in a sentence or two, such as "competitors' prices", for topics of this profile alone.
4. Optionally write the Placeholder reply the AI gives instead. Left empty, it says "I'm sorry, I can't help with that. Is there something else I can help you with?"

Each message checked is one more request to the utility model, which judges what the message asks for rather than single words: a refund for a damaged knife is not violence. A message it refuses is answered with the placeholder reply and never reaches the model or the profile's tools. A reply it refuses is replaced before it is sent or spoken.

Channel	What is checked
Website and admin chat	What the customer types, and the AI's reply. The reply streams as it is written, so a refused reply is replaced once it is complete. A refused message and its placeholder are not saved in the conversation, like a message Prompt Security blocks.
Text messages	The customer's texts, and the reply before it is sent.
Phone calls, turn-based	What the caller said, and the reply before it is spoken. Each check adds a moment before the AI answers.
Phone calls, realtime voice	What the caller said, checked while the call goes on, so normal answers are not slowed down. When a request is refused, the AI is told not to act on it. If it is still answering, it is cut off and says the placeholder reply instead; an answer it has already finished is left as it was. Its replies are not checked: they are spoken as they are made.

If the utility model can't be reached, the conversation goes on unchecked rather than leaving the customer without an answer.

Add your own topics

Topics are prompt templates in the Guardrail topic category: the template's title names the topic, and its text tells the utility model what falls under it. Add one when several profiles must refuse the same thing, such as your competitors or a pending lawsuit.

1. Open Artificial Intelligence > Guardrail Topics and click Add Guardrail Topic, or click Add a topic under the topics on a profile.
2. Give it a Title, such as "Competitors". The category is already set to Guardrail topic.
3. In System message, describe what is not allowed in a sentence or two, for example: "Questions about Acme or Globex, their prices or how we compare to them."
4. Click Save. The topic appears in the Guardrails card of every profile.

Like speaking styles, topics travel with the AI Profile Templates recipe step and deployment step. A profile's guardrails, with every other setting, travel with the AI Profile recipe step and deployment step.

Act when a guardrail triggers

The AI Guardrail Triggered workflow event starts a workflow each time a guardrail refuses a message, for every profile or one. It can notify a supervisor, flag the contact, or hand the conversation to a person. The workflow input has:

Value	What it is
Category	The topic's template name, such as guardrail-topic-self-harm, or other for the profile's own other topics.
Topic	The topic's title, such as "Self-harm".
Direction	Customer or Reply.
Reason	The utility model's one-sentence reason.
Text	The refused message.
SessionId, ProfileId	The conversation and the profile.
Contact, Channel, Variables	On calls and texts, the conversation's contact, channel and variables.

Privacy

On the Data Processing & Metrics tab, Privacy chooses what is kept of each conversation and for how long.

What to keep	What is stored
Everything	The full transcript, summaries, extracted fields and evaluations.
Everything except personal details	The same, with names, street addresses, dates of birth, phone numbers, email addresses, and card, account and identity numbers replaced by placeholders such as [NAME] and [PHONE].
Times, outcomes and scores only	The transcript is deleted, and the title, summaries, extracted values and evaluation notes are cleared. Times, status, analysis task names, criteria results and scores are kept.

The choice applies about 15 minutes after a conversation ends, once its workflows, its post-session analysis and contact memory have read it, so they still work on the full conversation. It applies to conversations that end after you choose it; conversations kept before are left as they were. Personal details are found by patterns and by the utility model; a detail it misses stays, so check a sample of redacted conversations before relying on it.

Names to keep lists the names that are not the customer's personal details, so they stay in the transcript: the name the assistant introduces itself by and the business's name, one per line or separated by commas. You don't need to list the site name (from Configuration → Settings → General) or, with the AI Personas feature, the assistant name of the persona the profile uses; those are kept anyway. A detail spoken over two turns, such as an email address dictated in two parts, is found and replaced in both. The customer's own name is still replaced wherever either side says it.

Delete conversations after deletes a conversation, with its transcript, extracted data and evaluations, after that many days without activity. Leave it empty to keep conversations. Usage records are kept, because they are what usage is billed and reported on.

Contact memory

Contact memory is kept apart from conversations, so the storage mode does not change it. Turn off Save to contact memory if it should not hold personal details either. See Contact memory and integrations.

Bleeped call recordings

With Everything except personal details, the recordings of AI phone calls are bleeped too, once the conversation's transcript has been redacted: every customer line that held a personal detail (a placeholder such as [CARD], or a long run of digits the redaction missed) is covered with a tone, from just before the line starts to when the next line starts, and the recording is stored again as WAV in place of the original. The Call Recording Bleeping background task does this every 10 minutes for recordings of the last 14 days, and the recordings API marks a bleeped recording. Lines are placed by when they were said, not word by word, so a whole line is covered, including any words around the detail. MP3 and 16-bit WAV recordings are read; others are left as they are.

A single conversation can also be deleted through the API, for example when a customer asks to be forgotten. See the AI conversations API.