To enable the AI Agent to auto-reply to tickets and execute actions, you'll need to create a Procedure for each type of question. Procedures are the most important piece of guidance that you can provide to your AI Agent. The AI Agent will follow the Procedure to handle the ticket automatically by auto-replying and if required, performing actions such as approving a cancellation.

This help file explains Procedures in eDesk and how they are used by an AI Agent.

01 What's a Procedure?

A Procedure is a set of rules that you create to tell the AI Agent how to handle a specific types of messages. For example, "Where is my order?" or "I want to return this item." Procedures are used only by an AI Agent, and are not used by AI Assist.

Think of a Procedure as a flowchart. The AI Agent follows the Procedure to walk through the customer's message and the details of their order and then decide what to do, for example, send a reply, ask for more information, or hand the ticket over to the human Care Team. The example below shows a Procedure that tells the AI Agent how to handle a cancellation request.
A grahical representation of a Procedure

When a customer message arrives in eDesk, eDesk's AI classifies it into a topic, and if a Procedure exists for that topic, then the AI Agent can follow the Procedure to generate an auto-reply or execute an action.

02 What's in a Procedure?

A Procedure is built as a flowchart using a small set of steps, which are:
  • Conditions
  • Routing Conditions
  • Gates
  • Guidance
  • Outcomes
Read on to learn more.

Conditions

A Condition is a checkpoint that branches into 2 paths in the flowchart: a Yes path and a No path. The AI Agent will check the information it has and use it to decide to follow either the Yes path or the No path. 

Routing Conditions

Where as a Condition branches to two paths, a  Routing Condition branches to three or more paths. 

Examples:
●    Use it to identify a Channel type – Amazon vs eBay vs your storefront - and perform different actions for that Channel.
●    Identify a Region – EU vs UK vs US, where policies differ
●    Identify a Customer segment – B2B vs Consumer vs VIP
●    Identify a Order value – e.g. under £20 vs £20–£100 vs over £100

Each branch coming off a Routing Condition has its own path, and can lead to entirely different sub-flows, outcomes, or actions.

How is it different from a Yes/No Condition?
●    A condition is a yes/no check. Use it when there are two possible answers.
●    A routing condition is for three or more clear-cut, mutually exclusive cases, and it's typically placed near the top of a procedure, so it can steer the whole rest of the flow (e.g. send Amazon tickets down one path, eBay down another).

A good use case
Controlling exactly where automation is allowed. For example, you might route refunds under £20 to an automatic outcome, refunds between £20–£100 to a path that replies but leaves the action for a human, and refunds over £100 straight to a team member. This gives you granular control over which tickets the AI Agent can act on, without touching the rest.

Gates

A Gate is a building block that instructs the AI Agent to pause and get something before it can continue to the next building block in the flowchart. Instead of making a decision based on information it already has (which is what happens in Conditions and Routing Conditions), i it asks a question, waits for a reply, or fetches extra information first. If the customer has already shared the information, the AI Agent won't ask for it again, instead it moves onto the next stage of the flow.

Add a Gate when the AI Agent needs to:
  • Ask the customer for something - E.g., a photo of damage, an order number, or a screenshot of a charge.
  • Confirm before taking action - E.g. checking "You'd like to cancel this order, is that right?" before going ahead.
  • Wait for the customer to do something - E.g, waiting for a returned item to be sent back before processing a refund.
  • Look something up - E.g. checking the latest delivery status.
  • Sort a reply into categories - E.g. asking why a customer wants to return an item, then routing their answer into dedicated paths for broken item, wrong item, or changed mind categories.
A Gate has two parts:
  1. What it asks or looks up – the message sent to the customer, or the information it's retrieving
  2. What counts as a match – how the AI Agent decides whether the customer's response satisfies the Gate (or which category it falls into)
If the customer doesn't reply, the Gate simply stays open and waits, your AI Agent won't guess or move forward without the information it needs.
A Gate can also be set up to trigger an action once it's satisfied, for example, automatically sending a return label once a customer confirms they'd like to proceed.

Guidance

Guidance is a short note you write inside a Procedure to steer how your AI Agent thinks or responds at a particular point without creating a new branch. Unlike a Condition or Gate, Guidance doesn't ask a question or wait for anything. It simply gives your AI Agent extra context or direction to keep in mind as it continues down the path.
If a customer has already provided evidence of a damaged item, you might add Guidance like "The customer has provided evidence of a defective product, so accept the return unless the item shows signs of misuse or damage is just to the packaging."

This doesn't split the conversation into different paths it just shapes how the AI reasons and replies from that point onward.

Add Guidance when you want to:
  • Steer tone or wording for a specific situation.
  • Give the AI extra context that isn't a yes/no decision.
Keep your Guidance general rather than tied to one specific customer's exact wording or product. Describe the situation (e.g. "the customer is expressing frustration about a delay") rather than quoting an actual message, this helps your AI Agent apply the Guidance consistently across similar tickets, not just the one you were thinking of when you wrote it.

Outcomes

An Outcome is the final step in a Procedure. Ultimately, Outcomes decide what happens for the customer and the ticket. Each path in the flowchart will end with an Outcome. 

The available Outcomes:
  • Delegate to AI Agent
    The AI Agent will handle this conversation and reply to the customer.
  • Escalate to Human Agent
    Hand the ticket over to a member of your team.
  • Accept Return
    Approve the customer's return request.
  • Refuse Return
    Decline the return request.
  • Accept Cancellation
    Approve the customer's cancellation request.
  • Refuse Cancellation
    Decline the cancellation request.

Outcomes can trigger actions

Some Outcomes can be set up to do more than just reply, they can also carry an action, like issuing a refund or sending a return label. You can even set different actions depending on whether a reply was sent or not, for example:
  • If the AI Agent replies through this Outcome → send a return label.
  • If the Outcome is reached but no reply is sent (e.g. the ticket is simply tagged and handed off) → notify a specific team.

Important

  • Reaching an Outcome doesn't always mean a reply gets sent, sometimes the right result is to quietly tag a ticket, escalate it, or close it without a response.
  • Actions attached to an Outcome (like a refund) will only actually happen if your AI Agent auto-sends the reply. If it steps back and a team member sends a draft instead, the action won't fire, nothing happens automatically without your team member choosing to do it manually.
  • That last point is worth building into your procedures: Always think about what should happen when the AI Agent successfully replies through an Outcome, and what should happen if it doesn't.

Further Reading

To learn how to create a Procedure, see here.

To learn more about the AI Agent and AI Assist, see here.