AI Assistants

Your AI team members

An Assistant is the AI that talks to your customers. Think of it as a team member with a specific role — a support agent, a sales rep, a receptionist — that's always available and never needs a break.

You can create as many assistants as you need. A common setup is one assistant per purpose:

• A Support Assistant that answers product questions and troubleshoots issues.
• A Sales Assistant that qualifies leads and books demos.
• A General Assistant that handles everything else and routes to the right place.

Each assistant has its own persona, channel assignments, and attached playbooks. They operate independently, so you can tune each one for its specific job without affecting the others.

Creating an assistant

Go to AI Assistant > Assistants and click Create Assistant to get started.

Name and description

Give your assistant a name that makes sense internally — your team will see this in the inbox and in reports. The description is optional but helpful if you have multiple assistants and want to remember what each one does.

The persona (system prompt)

The persona is the most important part of your assistant's configuration. It's the instruction set that tells the AI who it is, how to behave, and what it's responsible for.

A good persona is specific. Instead of "Be helpful and professional," try something like:

"You are a friendly and concise support agent for Acme Co. Your job is to help customers troubleshoot issues with our software products. Always ask one clarifying question before suggesting a solution. If you can't resolve an issue within two exchanges, offer to connect the customer with a human agent."

The more specific you are about tone, responsibilities, and edge cases, the more consistently your assistant will behave.

Channel assignment

You need to assign your assistant to at least one channel before it can respond to customers. Available channels include:

If you assign multiple assistants to the same channel, use Routing Rules to control which one handles a given conversation.

Advanced settings

Once the basics are in place, you can fine-tune how your assistant behaves.

Model selection

Autoch.at lets you choose the underlying AI model that powers your assistant. The right choice depends on what your assistant needs to do:

• Standard models are faster and more cost-effective. They work well for straightforward Q&A, simple routing, and first-response handling.
• Advanced models handle more complex reasoning, multi-step playbooks, and nuanced conversations better. Use these when accuracy and depth matter more than speed.

Temperature

Temperature controls how creative or predictable the assistant's responses are.

A low temperature (closer to 0) makes responses more consistent and factual — good for support and compliance scenarios where you want the AI to stick closely to what it knows. A high temperature (closer to 1) makes responses more varied and conversational — better for sales or marketing contexts where a bit of personality helps.

Most support assistants work best somewhere in the middle.

Playbook attachment

Playbooks define what your assistant is trying to achieve in a conversation. You can attach one or more playbooks to an assistant, and the AI will determine which one applies based on the customer's intent.

See Playbooks for how to create and structure them.

Testing your assistant

Before sending your assistant live, test it in the Widget Playground at Conversations > Messaging > Widget Playground. Select the assistant you want to test and interact with it exactly as a customer would.

Check that it:

• Follows the persona you wrote
• Uses the right training context to answer questions
• Executes its assigned playbooks correctly
• Triggers routing rules when it should

Monitoring performance

Once your assistant is live, keep an eye on how it's doing in Insights > Analytics. The metrics to watch are:

• Total conversations handled — Is it taking load off your team?
• Handoff rate — How often is it escalating to humans? A very high rate might mean the persona or training context needs work.
• Resolution rate — How often does it fully resolve a conversation without human help?

If something isn't working, the most common fixes are updating the persona, adding more training context, or refining the playbook instructions.