Overview
With Minimal AI, you set up AI customer support agents that resolve tickets in your helpdesk just like a human would. An agent works best when it’s connected to your systems, trained on your business processes through protocols, and given a clear tone of voice so every reply is on-brand.How an agent handles a ticket
When a new message comes in, your agent works through a series of steps. Understanding them makes it much easier to tune behaviour, because each step is driven by different settings.1
Reply decision
First the agent decides whether to engage at all. Driven by your Reply settings and Rules — for example, it can skip automated notifications, spam, or messages from business partners.
2
Understand the request
It identifies what the customer is actually asking, including when a single message contains multiple questions.
3
Match protocols
It finds the protocol(s) that cover each topic in the message.
4
Gather data
It calls your connected integrations to look up whatever the protocol needs — an order status, a subscription, tracking info, and so on.
5
Write the reply
It drafts the answer, applying your Rules, Tone of Voice, and language settings.
6
Safety & automation checks
Finally it decides whether to send the reply automatically or leave it as a suggestion for your team, based on your automation settings and guardrails.
These steps are independent. Your Tone of Voice and Protocols shape what the reply says, but they do not influence whether the agent replies — that’s decided only by your Reply settings and Rules. Keeping this separation in mind makes tuning much more predictable.
Setting up an AI Agent
Create an agent from Dashboard > Agents and clicking Add agent. From there you’ll see the settings described below. An agent only processes tickets while it’s set to active, and it only replies to messages that arrive during that time. After a message comes in, a reply typically appears within a minute or two — this depends on the complexity of the ticket, your configured response delay, and how quickly your helpdesk notifies our platform.Agent Name
The name for this agent. It’s only visible to you, never to your customers.Channels
Channels appear once you’ve connected a helpdesk through integrations. Once a channel is connected and the agent is active, the agent starts writing suggestions or automated messages in that channel based on your automation settings.Stores
Select the stores this agent can access while answering tickets — used for order, customer, and product lookups.Training Center Base
Select the Training Center base this agent can access. The base determines which business processes (protocols) the agent follows while resolving tickets.Website
Add webpages you want the agent to reference. More on connecting website content here.Language settings
Control the language your agent replies in. Select Use same language as incoming ticket to have the agent reply in the customer’s language. For custom logic — for example, reply in Dutch to Dutch customers and English to everyone else — select Other and describe the logic in plain language. The agent can also generate translations of incoming messages (and of suggestions, when not in automation mode) for your human team. Once enabled, pick the language to translate into. If a message is already in your preferred language, no translation is made.Tone of Voice
Describe how you want the agent to sound. Tone of Voice only affects the style of a reply — phrasing, openings, and sign-offs — never what the agent does or which steps it follows. Be specific and give examples; the agent follows these instructions literally. For example, if you write “Be inspirational and motivating”, replies will be very inspirational and motivating. Expect to refine tone of voice over time — it’s normal to iterate before you find the right balance. Frequently asked questions:- Can Tone of Voice be empty?
Yes. The agent already has sensible default tone instructions and behaves as a helpful support assistant unless you tell it otherwise. - Can I set a greeting / sign-off in Tone of Voice?
Yes. For example: “Always open with Hi [name], and close with Best, [name]”. The agent will then consistently use those openings and sign-offs.
Rules
Rules are always-on do’s and don’ts that apply to every reply, regardless of topic. They shape how the agent behaves once it has decided to reply. Rules can include guidance on when to escalate to a human, but they should not contain integration steps or automation logic — keep those in protocols and your automation settings. Examples of rules:- “Never discuss ingredients unless a protocol says otherwise.”
- “Always write our company name as SitRight, regardless of how the customer spells it.”
- “If the customer says they’re happy, invite them to leave a review and share our review link.”
Tone of Voice and Rules are applied to every reply, no matter the question. That makes them powerful levers for controlling your agent’s overall style and behaviour.
Reply settings (when to stay silent)
Reply settings let you tell the agent which messages it should not reply to. When a message matches, the agent does nothing and leaves the ticket open for your team — think of it as a spam filter for your agent. Be as specific as possible for reliable results. Common examples:- Messages from marketplaces like bol.com
- Messages from business partners or suppliers
- Sales pitches from people trying to sell you products or services
Automation vs. suggestions
Every reply is either sent automatically or left as a suggestion (draft) for a human to review and send. You control this in your automation settings.- Suggestion mode — the agent drafts a reply and leaves it in your helpdesk for a human to review. This is the safest way to start.
- Automated mode — the agent sends replies on its own. A reply is only sent automatically when automation is enabled and the relevant protocol is set to automate and it passes your guardrails.
- Guardrails block automatic sending when something looks off — for example toxic language, prompt-injection attempts, unfilled placeholders, or a customer explicitly asking for a human. When a guardrail trips, the agent leaves a draft instead of sending.
- Rate limits cap how many messages the agent will auto-send per ticket and per hour, so it can’t run away in a loop.
- Human takeover — once a human replies in a ticket, the agent stops automating on that ticket (unless you’ve enabled continuing after takeover).
- Response delay — a configurable delay (30 seconds by default) before a reply is sent or drafted, so replies don’t feel instant and your team has a moment to intervene.