Autopilot features in rules

When you have Autopilot enabled for your company, you can use the following features in your workflows:

• Rule action: Reply with Autopilot instructions

• Branching rule step: Branch with Autopilot

• Dynamic variable: Extract with Autopilot

• Dynamic variable: Summarize with Autopilot

These features can be used outside of the default Autopilot rule based on Topics. You can also start building from templates in the rule library. See this article to learn more.

The following instructions assume you’re familiar with creating rules.

Use the Reply with Autopilot instructions rule action to tell AI how you want to automatically reply to and resolve conversations.

Step 1

Navigate to the rule where you’d like to use the Reply with Autopilot instructions rule action.

Step 2

In the text box, enter instructions for the information Autopilot should include in its auto-replies. You can also add dynamic variables to this field.

Best practices:

• This feature has the full context of the conversation but does not have access to knowledge like an auto-reply without instructions. This means you should include as much detail as possible in these instructions, including all factual knowledge you want included in the reply.

• Be very direct with instructions and avoid ambiguity.

• Some examples:

  • Please generate a reply apologizing to the customer for the outage and directing them to this status page for live updates, website.status.com.

  • Based on what the customer outlined in the conversation please write a summary of their feedback and then tell them their feedback is important to us and we will be adding it to our feature planning process.

• See this section below for additional best practices.

Step 3 (optional)

In the Knowledge sources field, select the knowledge sources you’d like Autopilot to use to generate a reply. You must have a knowledge source configured in your Front AI settings to use the dropdown. See this article to learn more.

Important to know:

• You can select up to 10 individual sources from any combination of Front knowledge bases, public websites, or third-party sources.

• We recommend only selecting sources necessary for the use case you want Autopilot to reply to. When a source is selected it will be used to generate a reply, and AI will not evaluate whether the selected knowledge is the most appropriate fit for the situation.

• If you do not select a knowledge source, Autopilot will only use the content entered in the instructions text box.

Step 4

In the Hold reply for review field, select whether you want the rule to automatically reply (No, reply automatically) or if you want a teammate to review the content before sending it (Yes, create draft).

Step 5

Click Save when finished.

Use the Branch with Autopilot step in branching rules to quickly triage conversations using natural language questions.

Step 1

Navigate to the branching rule where you’d like to use this step. Click the plus sign (+), then select Branch with Autopilot.

Step 2

In the Branch with Autopilot panel, fill in the following fields:

• Question: Select whether you want AI to ask a yes/no or multiple choice question.

• [Yes or no / multiple-choice] question: Enter the question you want Front AI to branch on. This works best when you ask a natural language question.

  • You can add dynamic variables to the text field.

  • Check the box next to Review entire conversation to answer if you'd like AI to reference the entire conversation instead of the most recent message.

  • Check the box next to Review latest message signature to answer if you'd like AI to reference the content in the signature of the most recent message.

• Autopilot answers: For multiple choice questions, enter in the answers you’d like to branch on.

In this example, we want AI to detect if the customer’s most recent message refers to canceling a subscription.

Click Save.

Step 3

New action steps will appear in the flow builder. Set up the rest of your rule based on the actions you want the rule to take based on each branch.

Click Create when finished.

Enterprise plan only: Use the Extract with Autopilot transformation in dynamic variables to identify, extract, and transform data from conversations to automate complex processes.

Step 1

In your rule, click the dynamic variables icon, then click Create dynamic variable.

Step 2

Click Create new, then select the text-based field you want to work with. In this example, we’ll select Message body.

Step 3

Click Add a step, then select Extract with Autopilot. Fill in the following fields:

• Dropdown: Select whether you want to extract Text, Date and time, or Number.

• Instructions: Enter the instructions you want AI to follow. This works best when you use natural language.

In this example, we want AI to detect the outbound flight date in the customer’s most recent message.

Step 4

Click Test instructions, add sample content, then click Extract to ensure the results match what you are expecting to see.

Step 5

Click Save. You can now use this dynamic variable in your rule. 

In this example, we’ll have the action step update a conversation custom field with the date extracted from an inbound message.

Enterprise plan only: Use the Summarize with Autopilot transformation in dynamic variables to identify the key points in long conversations.

Step 1

In your rule, click the dynamic variables icon, then click Create dynamic variable.

Step 2

Click Create new, then select Summarize with Autopilot.

Step 3

Select the format for your summary:

• Summary with highlights: Title with key conversation highlights in bullet points

• Condensed summary: Title only, without any additional details or highlights

• Custom instructions: Generate content using your own instructions

Step 4

Click Save. You can now use this dynamic variable in your rule. 

In this example, we’ll have the action step escalate the conversation to another inbox, then add a comment with a summary of the conversation to help the next team quickly gather context.

In this example rule, we’ll create an Autopilot rule that will take different actions depending on if a conversation has a Topic, if the conversation is urgent, and when the message is received based on business hours.

Step 1

In the Topics page, enable auto-replies for the Topics you’d like to include in your rule.

Step 2

Create a new branching rule and select the shared inboxes the rule applies to. In the Triggers step, add the Topic is identified trigger.

Click Save.

Step 3

Click the plus icon (+), then add a new Branch with Autopilot step to check if a message tagged with those Topics is urgent.

Configure a yes/no question with the question “Is this message an urgent matter?”, then click Save. You’ll see three new branch results appear, one for the TRUE state, one for the FALSE state, and one for No answer matched.

Click Save.

Step 4

Set up the actions or branches following the Branch with Autopilot step:

• If the AI determines a message is TRUE then you can add an action. In this example, if a conversation is urgent we want to assign it to a specific teammate who handles urgent communications.

• If the AI determines a message is FALSE, you can add another branch. In this example, we want to add a branch to check if the message was received within or outside of business hours.

• If AI cannot determine if the conversation is urgent or not, you can add another branch or action. In this example, if AI cannot determine the urgency of the conversation, we want to auto-reply with a message template.

Step 5

Set up the actions for the business hours step.

• Message is received within business hours: Assign the conversation to a specific team via round robin assignment

• Message is received outside of business hours: Have Autopilot send an auto-reply

Step 6

Click Save to finish. The completed rule will look like the following:

You can preview how your Autopilot rule will behave using the test conversation feature. This allows you to:

• Test branching logic across multiple rule paths in one go.

• Preview how Autopilot steps will behave, making it easier to iterate on AI-powered workflows by knowing exactly how they'll perform with real conversation data.

• Catch edge cases and fine-tune AI behavior before it impacts customers.

When using the Reply with Autopilot instructions rule action, follow these best practices to optimize its effectiveness.

All routing logic should be handled before the reply action — in your branching rule's classification steps and conditions. Don't write instructions like "If the customer asks for a refund, do X." That decision should already be codified in a Branch with Autopilot step upstream. By the time the reply action fires, the conversation has already been routed to the correct branch, and the prompt should simply tell Autopilot how to respond in this specific scenario.

Good example — This prompt assumes the conversation has already been classified as an invoice timing question from a VIP customer by an upstream Branch with Autopilot and Branch by condition step. The prompt doesn't re-decide what the message is about; it just tells Autopilot exactly how to respond:

Tell the customer that they will receive an invoice at the latest 48 hours after their order has been placed. Also share a link to the Help Center article on invoices. Then let them know they should reach back out if they haven't received the invoice after 48 hours.

Sample verbiage:

Thanks for reaching out. Sending invoices in a timely manner is really important to us and that's why we always guarantee we will send them to you within 48 hours.

More information on our invoice policy can be found here: [link] 

If 48 hours have passed and you still haven't received your invoice, please reach out to us and we will be happy to assist. Thanks!

Notice the structure: the prompt tells Autopilot how to respond (with the 48 hour invoice expectation), where to point the customer (a Help Center link), and provides sample verbiage that models the tone and format. The sample verbiage gives Autopilot a strong template to work from while still allowing it to adapt to the specifics of the conversation.

Bad example — Trying to make the prompt do the decision-making:

If the customer is asking about billing, answer their billing question.

If the customer is asking about a bug, ask them for reproduction steps.

If the customer is asking for a feature, let them know we'll pass it along.

This belongs in a Branch with Autopilot step, not in the reply prompt. Each of those scenarios should be its own branch with its own tailored reply action.

Use this as a starting point and fill in the specifics for your scenario:

[What should the reply do — e.g., answer the customer's question, confirm receipt, etc.]

Information to include:

1. [First thing to mention, ask for, or address]

2. [Second thing]

3. [Third thing, if applicable]

   (Link to relevant Help Center article if one exists: [URL])

Sample verbiage:

[Write out a sample reply in the tone and style you want Autopilot

to follow. Include the sign-off you want used.]

The most effective instruction prompts follow a pattern:

• What to do — A clear directive: ask for information, answer a question, confirm receipt, etc.

• Specific items — A numbered list of exactly what to include.

• Links and resources — Point to specific Help Center articles or documentation where relevant.

• Sample verbiage — A model reply that sets the tone, structure, and sign-off you want. This is the highest-leverage element — it gives Autopilot a concrete example to work from rather than abstract instructions.

• What to avoid — What the reply should never contain (internal team names, timeline commitments, speculation about unreleased features, etc.).

Unlike teammate-sent replies, Autopilot-generated replies don't pull from your configured email signature. If you want a consistent sign-off, include it directly in your instruction prompt. You can hard-code elements like:

Sign off with:

"Thanks for reaching out,

The [Team Name] Team"

When you first set up a prompted reply action, you can use the Hold reply for review option and select Yes, create draft. This generates the reply as a draft that a teammate can review before sending, letting you verify quality before switching to fully automatic replies.

When using the Extract with Autopilot dynamic variable, follow these best practices to optimize its effectiveness.

Extraction is built to pull a single value out of a conversation — not to classify a message or make a series of decisions about it. Any routing or classification logic should be handled before the extraction, in your branching rule's Branch with Autopilot steps and conditions. By the time the extraction runs, the conversation should already be routed so that the instruction can focus on retrieving one value.

Good example — This instruction assumes an upstream Branch with Autopilot step has already confirmed the sender asked for a phone call. The prompt does one thing — pull the value:

Extract the first occurrence of a phone number in the message.

Bad example — Trying to make the instruction do the decision-making:

If the message mentions a call request, and there is a phone number in the email, confirm the number. Ignore phone numbers mentioned in signatures. If no phone number is mentioned anywhere, say “null”.

This tiered, if-then logic belongs in a Branch with Autopilot step, not in the extraction instruction. Layering it into the prompt distracts from the extraction itself and produces inconsistent results. Decide whether the message contains a relevant value upstream, then let the extraction simply pull it. There is no need to account for null or empty values, as the rule node will simply skip any instances where there is no matching value present to extract. 

This template can be used for each of the possible AI Extract dynamic variables, so be sure you’re using the one that most closely aligns with your desired extracted variable. Use this as a starting point and fill in the specifics for your scenario:

Extract [the single value you want — e.g. the delivery date, the order number, the first phone number mentioned].

[If more than one candidate may appear, say which one to take — e.g. “extract the first one mentioned."]

The most effective extraction instructions follow a pattern:

• One value — Name the single piece of data to pull. Keep the instruction free of conditional, if-then logic.

• Which one — When more than one candidate could appear (for example, several phone numbers in a signature), point to the one you want by its position, format, and/or order: "extract the first phone number mentioned that starts with a +49 country code." 

• Output type — Set the dropdown (Text, Date and time, or Number) to match the field or action where you'll use the result.

Keep each dynamic variable focused on a single value. If you need to capture distinct kinds of a value — for example, a phone number or an email — set up a separate variable for each rather than asking one instruction to decide which type it's looking at. Each variable stays simple, and you can use the results independently in your rule.

Extraction returns a value only when that value actually appears in the message — AI will not invent one to satisfy the instruction. As a result, an instruction like "if no value is found, return none" won't reliably hand you a usable placeholder to act on; when nothing matches, the extraction simply comes back empty. If you need to route or act on conversations where the value is absent, make that decision with a Branch with Autopilot step and reserve the extraction for pulling the value when it is present.

Before saving, use Test instructions (Step 4 above). Include a typical message that contains the value, a message where the value is phrased unusually or sits among similar-looking values, and a message where the value is absent — so you can confirm the extraction behaves the way you expect in each.

It depends. The default Autopilot rule that is automatically created does not apply to your company’s rule limit. If you create additional rules, these rules are counted towards your rule limit.

No. Topics are not required to use the Reply with Autopilot instructions rule action and Branch with Autopilot rule step.

Yes. The Reply with Autopilot instructions rule action and Branch with Autopilot rule step can be used in rules outside of the default Autopilot rule set up in your Topics settings. These features don’t need to wait for a Topic to be identified before firing.

The Reply with Autopilot instructions rule action uses information from the last 20 messages in the conversation where the rule triggered.

It will not reference information from the conversation’s comments, other parts of the rule, custom fields, or other types of data in Front. However, this is on the roadmap for a future update.

The Branch with Autopilot step looks at the message content in the last message of a conversation.

No. The Reply with Autopilot instructions rule action is designed to support use cases where you need to resolve a specific issue, and therefore only need to add specific knowledge. Since you are the expert for this workflow, Front cannot make the decision about which knowledge to use.

Similar conversations are numerous and varied, so we want to only offer the most definitive sources of knowledge.

You'll still have access to the Get summary dynamic variable in any existing rules and macros. To use the new Summarize with Autopilot dynamic variable, you must have the Enterprise plan with the Autopilot add-on enabled.

See this FAQ to review which messages trigger errors when using the Test conversation feature on rules using AI steps.

See the Autopilot article for more information.