Skip to main content

​
Overview

“Automation” in Minimal AI is not a single switch. Whether your agent sends a reply on its own is decided by several layers that all have to line up:
  • Global automation — the account-wide master switch that decides whether your agent may send replies automatically at all. Configured in your automation settings.
  • Automation coverage — for a specific reply, something has to say “this is safe to send on its own”. There are two independent ways to get coverage: automated protocols and the automation prompt.
  • Guardrails — a final safety check that can still hold a reply back even when everything above says go.
Turning on global automation does not automate every reply. It only unlocks automation — each reply still needs coverage (from an automated protocol or the automation prompt) and has to pass guardrails.

​
The master switch: global automation

Global automation lives in your automation settings and applies to the whole agent. When it’s off, every reply is left as a suggestion for your team, no matter how anything else is configured. When it’s on, automation is unlocked — but a reply is only sent automatically if it also gets coverage and passes guardrails.

​
Automation prompt vs. automated protocols

This is the distinction that causes the most confusion. Both grant coverage — permission for a specific reply to be auto-sent — but they work very differently, and a reply is eligible if either one is satisfied.

​
Automated protocols

Each protocol has an Automated toggle. When a message is handled by protocols, coverage passes only if every matched protocol is set to Automated. If a message spans multiple topics and even one matched protocol is not automated, the reply is drafted instead of sent. This is precise, per-scenario control: you decide exactly which scenarios the agent may handle on its own by toggling them one at a time.

​
The automation prompt

The automation prompt is a plain-language instruction in your agent’s configuration that describes which kinds of replies are safe to send automatically — for example, “Automate responses about order status, delivery times, and return policies.” When the drafted reply fits what the prompt describes, coverage passes. The important catch: the automation prompt is evaluated from only the conversation and the reply the agent drafted. It does not see which protocols matched, what your integrations returned, or which actions were involved. So:
  • ✅ Good automation-prompt instructions describe topics or types of replies: “Automate answers to questions about shipping times and our return window.”
  • ❌ Poor instructions rely on things the prompt can’t see: “Automate when a Shopify order lookup was performed” — it has no visibility into tool calls or protocols.
This makes the automation prompt a broad, topic-level allowance you can set once, rather than toggling every protocol individually.
Coverage is an OR: a reply is eligible to auto-send if its matched protocols are all automated or the automation prompt allows it. You don’t need both.

​
The layers of automation

For a reply to actually be sent automatically, all of the following must be true at once:
1

Global automation is on

Automated mode is enabled in your automation settings. If it’s off, everything stays a suggestion.
2

The reply has coverage

Either every matched protocol is Automated, or the automation prompt allows this kind of reply. If neither is satisfied, the agent drafts a suggestion.
3

Any actions used are automated

Actions have their own automation toggle per protocol. An action only runs on its own when both the protocol and the action are automated. Otherwise the agent proposes the action and waits for a human.
4

Guardrails pass

Safety checks (toxic language, prompt-injection attempts, unfilled placeholders, a customer asking for a human, rate limits, human takeover) must not trip. If one does, the agent leaves a draft.
The logic is an AND across the layers: global automation on + coverage (automated protocols or automation prompt) + any actions automated + guardrails passed → the reply is sent automatically. If any layer fails, the agent falls back to a suggestion.
On any ticket’s insights page, the Automation Coverage section shows both paths side by side — automated-protocol coverage and automation-prompt coverage — so you can see which one (if either) allowed a reply to be sent, or why it was drafted.

​
Active vs. Automated

Don’t confuse a protocol’s two toggles:
  • Active — whether the agent may use the protocol at all. When paused, the agent ignores the protocol even if its base is connected.
  • Automated — whether the agent may auto-send replies for the protocol. When off, the agent still uses the protocol, but leaves its reply as a suggestion.
A protocol must be Active to do anything. Automated only matters once the protocol is active and global automation is on.
1

Start in suggestion mode

Keep global automation off at first. The agent drafts replies your team reviews and sends — the safest way to build trust in its answers.
2

Turn on global automation

Once you’re confident, enable automated mode in your automation settings. This unlocks automation but doesn’t auto-send anything on its own yet.
3

Grant coverage gradually

Start by automating the protocols you trust most (simple, high-volume, low-risk scenarios). As you gain confidence, you can broaden with an automation prompt that allows whole categories of replies.
Not sure which protocols or topics are ready to automate? Ask the AI Manager — it can learn from your ticket history and point out where automation is safe to enable.
Coverage and global automation both have to agree before anything is sent on its own. Automating a protocol (or writing an automation prompt) has no effect while global automation is off, and global automation alone sends nothing until a reply also gets coverage.