Actions
AI agent actions: a refund, not a link to a form.
An AI agent action is a task the agent performs during the conversation — it looks up the order, books the demo, moves the refund — instead of telling the customer where to go and do it. Outlearn runs actions from a trigger written in plain language, against Slack, Gmail, Calendly or your own API, and hands off with the transcript when a person has to take over.
Free plan, no card. Actions are on the plans, never per seat.
Last updated 13 September 2026.
One conversation, one action
Sample workspaceWebsite chat
Signed in · PriyaI was charged twice for September. Can one of them be refunded?
I can see both charges. The duplicate from 3 September is refundable under the 30-day policy, so I'll issue it now.
Done. The refund is on its way and should reach your card within 5 to 10 business days.
Refund issued
- Charge
- INV-2291
- Amount
- $99.00
- Status
- Refund in progress
Issue refund
POST /refundsLook up the account
Your site passed a signed identity token, so the agent already knows who is asking.
Check the refund policy
Read from a connected source: duplicate charges within 30 days are refundable.
Run the action
A POST to your billing endpoint, with the auth headers you configured.
Confirm in the chat
The result comes back as a card, the response format chosen for this action.
A custom action against your own API. The rule came from a source; the refund came from the action.
Website chat
Signed in · TomI've moved. Can you send order 48213 to 12 Harbour St, Leeds instead?
Order 48213 hasn't been dispatched yet, so the address can still be changed. Updating it now.
Updated. Order 48213 will ship to 12 Harbour St, Leeds. Here is what I changed:
Order 48213
- Ships to
- 12 Harbour St, Leeds
- Dispatch
- Tomorrow
- Changed by
- Your agent, just now
Update shipping address
PATCH /orders/{id}Look up the order
A GET on your orders endpoint: order 48213, not yet dispatched.
Check the rule
The trigger says: only change an address before dispatch. This order qualifies.
Run the action
A PATCH with the new address, to your endpoint, with your auth.
Show it back
The new address is returned as a card so the customer can check it.
One trigger, written in a sentence, decides when the action may run. The agent reads every message against it.
Website chat
Signed in · AnaHow do I upgrade to Pro?
I can do that for you now. Pro is billed from today and prorated for the rest of your cycle. Shall I go ahead?
Yes please.
Done. I've moved you to Pro. Your next invoice is prorated.
Subscription
- Plan
- Pro
- Effective
- Today
- Next invoice
- Prorated
Change plan
POST /subscriptions/{id}/planLook up the subscription
Verified customer, current plan Starter, monthly billing.
Check the pricing rule
Read from your pricing documentation: upgrades apply today and prorate.
Run the action
The trigger says to ask first. The customer said yes, so the POST goes out.
Confirm in the chat
Plan, effective date and the invoice note come back as a card.
The homepage widget shows this exact exchange. Sources answer the pricing question; the action does the upgrade.
Website chat
Signed in · MarcusI'd like a refund on the annual plan I bought in March.
Annual refunds after 30 days need a person to approve them. I've passed this to the billing team with our conversation so far, so you won't have to repeat any of it. They'll reply here.
Handed to billing · transcript and summary attachedIssue refund · blocked by policy
Handoff · Billing teamLook up the account
Annual plan, purchased six months ago, verified through the identity token.
Check the refund policy
The policy says annual refunds after 30 days are approved by a person. The action does not run.
Hand off with the transcript
The billing team receives the full conversation and an AI summary, in the tool they already use.
Tell the customer
What happened and what comes next, in the chat. Not a ticket number and a wait.
An action that cannot complete is a handoff, not a dead end. The person who picks it up reads the whole story first.
A sample workspace, to show the shape of a run. Our own production numbers are below, with their denominators attached.
Measured, not promised
Resolved means resolved, and we publish the denominator
A conversation that ends in a link to a form is not resolved, whatever the dashboard says. These are production numbers with the arithmetic attached, from the AI support benchmark.
- 29,003
- conversations measured, Oct 2025 to Aug 2026
- 45.9%
- resolved by the agent alone, no human involved
- 50.8%
- resolved alone in the last 90 days of the snapshot
- 5.7%
- handed to a person, with the transcript attached
What an action does
Four things a support lead stops doing by hand
The documentation groups actions by what they touch. Each of these is a task that used to end with “I'll pass this on”.
Alert the team
Slack · Google Chat
Post to a channel when a trigger is met: a high-priority issue, a cancellation risk, or a summary when a chat ends.
Send the email
Gmail · Outlook
Send a conversation summary to the customer when the issue is resolved, or an alert to the support inbox when it is urgent.
Book the meeting
Calendly · Google Calendar
Check availability and book the demo or the call inside the chat, with the slots shown as a card the customer picks from.
Call your own system
Any API
A custom action runs against your own endpoint with your own auth headers: orders, refunds, plan changes, whatever your API can do.
Configure an action
A trigger written like a note to a new hire
Every action has a trigger in plain English, a response format, the channels it may run in and a limit per session. The agent reads each message against the trigger and decides. Test it in the Preview panel with the phrases customers actually use, then switch it on.
Actions · Create Action
ActiveName
Endpoint
Trigger
Use when a signed-in customer asks for a refund on a charge from the last 30 days. Confirm the charge with them first. Do not use for annual plans or charges older than 30 days — hand those to the billing team.
Response handling
Channels
Trigger limitation
Where it connects
The catalogue, and any API you already run
Pick an integration from the catalogue and authorise it, or add a custom action that calls your own endpoint with your own auth headers. The plan comparison shows how many actions each tier allows.
Slack
Alerts and summaries
Google Chat
Alerts and summaries
Gmail
Email
Outlook
Email
Calendly
Booking
Google Calendar
Booking
Shopify
Coming soon
Salesforce
Coming soon
Stripe
Coming soon
Your own API
Custom HTTP action
Status as the documentation lists it. Setup for each integration is in the actions documentation, starting with “What Are Actions and What Can They Do?”.
Sources and actions
A source gives the agent something to say. An action, something to do.
The two work together. The refund policy lives in the documentation you already keep; the refund itself is an action. The documentation's advice is to start with sources, get the agent answering well, then add actions where a conversation still ends in “I'll pass this on”.
Knowledge source
Action
What it gives the agent
Something to say
Something to do
When it is used
When a question comes up
When the trigger's condition is met
The data
Written content, kept in sync
Live, from the system
Examples
Refund policy, setup guide, pricing page
Issue the refund, book the slot, send the email
The difference
Most bots send a link. This one does the thing.
A typical chatbot
Outlearn
Answers the refund question with a link to the refund form.
Checks the policy from your docs, issues the refund through your API, and confirms it in the same chat.
A rules tree you build screen by screen, then rebuild when the wording changes.
One trigger written as a sentence. The agent reads every message against it and decides.
Asks the customer for their email, order number and the last four digits, again.
Your site passes a signed identity token, so the agent knows who is asking before the first reply.
Fails quietly when the system says no, and the customer opens a ticket.
Hands off to a person with the full transcript and a summary when the action cannot complete.
Actions sold as an add-on, priced per seat like the rest.
Actions are on the plans, priced by credits for the AI's work. The free plan needs no card.
Pricing, for the record: credits for the AI's work, never per seat. Free $0 with 100 credits, Starter $99/month with 2,000 credits, Growth $199/month with 4,500 credits, Scale $599/month with 15,000 credits. 14-day trial on paid plans, no card. Actions per plan and full detail on the pricing page. When an action needs a person, the handoff carries the transcript.
Connect one action and ask the agent to book a demo.
Authorise Calendly, write the trigger in a sentence, try it in the Preview panel. A few minutes on the free plan, no card.
What support leads ask about AI agent actions
What are AI agent actions?
An action is a task the AI agent performs during a conversation, as opposed to information it gives.
Where a knowledge source gives the agent something to say, an action gives it something to do: look up an order, book a meeting, send an email, post an alert to a channel, or call your own API to issue a refund or change a plan. In Outlearn each action has a plain-language trigger, a response format and a per-session limit, and the agent decides when to run it by reading the conversation.
How does the agent know when to run an action?
Every action has a Trigger: a plain-language description of when the agent should run it, such as 'use when the customer asks to book a demo, schedule a call or check availability'.
The agent evaluates every message against the triggers of the actions you have created, so there is no rules tree to build. The documentation's advice is to describe the intent and the phrasings people use, and to say when the action should not run. A Trigger Limitation caps how many times it runs in one session, and Available Channels decides whether it runs in chat, in email or both.
Which systems can an Outlearn action connect to?
The documented integrations are Slack and Google Chat for messages and alerts, Gmail and Outlook for email, and Calendly and Google Calendar for availability and bookings.
Shopify, Salesforce and Stripe are listed as coming soon. Beyond the catalogue, a custom action runs an HTTP request against your own endpoint with your own auth headers, which is how refunds, address changes, plan changes or anything else your API exposes get done inside the conversation.
What does the customer see when an action runs?
Whatever Response Handling you chose for the action: plain text, a card with the structured result, or a link.
The documentation recommends cards for time slots and confirmations, so a booking or a refund comes back as a small card the customer can check, inside the same chat. You can add Custom Response Instructions to shape how the result is described, or leave the agent to phrase it.
What happens when an action cannot complete?
The agent does not guess and does not go quiet.
When the policy needs a person, the system says no, or the trigger's conditions are not met, the conversation is handed to your team with the full transcript and an AI summary, into the helpdesk or channel you already use. The customer is told what happened and what comes next, in the chat, and never repeats themselves. The handoff page describes the mechanism in full.
How do I set up an action, and can I test it first?
In the Actions tab choose Create Action, pick an integration or a custom action, authorise the account if the integration needs one, then write the trigger, choose the response format, name the action, pick its channels and set the trigger limitation. The Preview panel lets you try the phrases customers actually use before the action is live. The documentation's advice is to start with sources, get the agent answering well, and add actions once you can see where they would resolve a conversation end to end.
Do actions cost extra?
Actions are not a paid add-on and are not priced per seat.
Every plan is priced by credits for the AI work the agent does, the free plan needs no card, and the number of actions each plan allows is shown in the plan comparison on the pricing page alongside the current prices and credit allowances.
Keep reading
The AI customer service agent
Reads your live docs, answers with the source, takes the action mid-chat, and hands the rest to a person.
Chatbot to human handoff
The full transcript and an AI summary go with the conversation; the customer never repeats themselves.
Knowledge base chatbot
Reads Helpjuice, Zendesk, Confluence, Drive, SharePoint, Slack and your website in place, kept in sync, with the source shown.
AI chatbot for your website, inbox, Slack and Teams
One snippet installs the widget; the same agent answers email, Slack, Microsoft Teams, Google Chat and your app via API.
Chatbot analytics
Auto-solve rate, handoff reasons and the questions the agent could not answer, on one screen from the first conversation.
Zendesk AI Chatbot
Reads your Guide articles and closed tickets, opens tickets mid-conversation.