Before automating customer support, you need to understand how each request currently moves from the first customer message to the final resolution. A clear workflow map helps you identify what AI can handle, where approval is required and when a human agent should take over.
Automation works best when it follows a reliable process. Without clear rules, an AI agent may ask unnecessary questions, perform the wrong action or escalate conversations that could have been resolved immediately.
Mapping the workflow first gives your business control over how each request is handled.
Start with the requests customers send most often
Begin by reviewing recent customer conversations and identifying the questions or requests that appear repeatedly.
Common ecommerce support workflows include:
- Order tracking
- Delivery questions
- Product availability
- Product recommendations
- Address changes
- Order cancellations
- Returns and exchanges
- Refund requests
- Damaged or incorrect items
- Discount and payment questions
Group similar conversations together, even when customers use different wording.
For example, “Where is my parcel?”, “Has my order shipped?” and “When will my delivery arrive?” can all belong to the same order-tracking workflow.
Document how each request is handled today
For every common request, write down the steps your team currently follows.
An order-cancellation workflow may look like this:
- Identify the customer
- Verify the order
- Check whether fulfilment has started
- Confirm which items should be cancelled
- Cancel the order if allowed
- Confirm the outcome with the customer
- Escalate if the order can no longer be cancelled
This exposes the decisions, information and actions involved in reaching a resolution.
Do not map only the ideal path. Include what happens when information is missing, the customer changes their request or the action cannot be completed.
Separate information from actions
Not every support request requires the system to change something.
Some conversations only require information, such as:
- Explaining a delivery timeframe
- Sharing a return policy
- Confirming product availability
- Providing an order status
- Answering a sizing question
These are usually easier and safer to automate.
Other requests involve actions, such as:
- Changing a delivery address
- Cancelling an order
- Creating a return
- Exchanging an item
- Issuing a refund
- Applying a discount
Actions need clearer permissions, verification and approval rules.
“A useful workflow map does not only show what happens. It shows what information is required, who has authority and what should happen when the normal path fails.”
Identify the information required at each step
Every workflow depends on specific information.
For an order-related request, the AI agent or support team may need:
- The customer’s email address
- An order number
- The order status
- Fulfilment information
- Delivery or tracking details
- The items included in the order
- The store’s cancellation or return policy
Mark where each piece of information comes from.
It may be available in the ecommerce store, the customer profile, a previous conversation or a policy document. Connecting these sources allows the workflow to move forward without forcing agents to search across multiple tools.
Define decision points
Decision points determine which path the workflow should follow.
For example, a return workflow may ask:
- Is the order verified?
- Is the item eligible for return?
- Is the request within the return period?
- Has the item already been refunded?
- Is the item damaged or simply unwanted?
- Does the request require manual approval?
Each decision should lead to a clear next step.
Avoid vague instructions such as “handle the return appropriately.” Replace them with rules the AI agent and support team can consistently follow.
Decide what can be automated
Review each step and place it into one of four categories.
Automated
The AI agent can complete the step without human involvement. Examples include answering policy questions, checking order status or collecting required details.
Automated with confirmation
The AI can prepare the action but should ask the customer to confirm before completing it. This may apply to cancelling an order or changing a delivery address.
Approval required
The AI gathers the necessary information and requests approval from an authorised team member. This may apply to refunds above a certain value or exceptions outside the normal policy.
Human required
The conversation should be transferred to a person. This may include sensitive complaints, unusual circumstances, legal threats or requests that fall outside the available information.
Add verification before private or sensitive actions
A workflow should verify the customer before showing private information or making changes to an order.
Verification may involve:
- Matching the customer’s email address
- Confirming an order number
- Using a secure account session
- Requesting another approved identifier
- Asking the customer to complete verification through the store
The level of verification should match the risk of the action.
Explaining a general return policy requires little or no verification. Changing an address or issuing a refund requires stronger controls.
Define permissions and approval limits
Your workflow map should show exactly what the AI agent is allowed to do.
For every action, decide:
- Whether the AI can perform it
- Whether customer confirmation is required
- Whether team approval is required
- Which team members can approve it
- Whether a value or time limit applies
- What happens if permission is denied
For example, the AI may be allowed to create return requests but not issue refunds. It may suggest a discount but require a human agent to approve it.
Clear permissions prevent automation from operating beyond your business rules.
Plan for exceptions
Real conversations rarely follow one perfect path.
Customers may provide the wrong order number, ask about multiple orders, change their mind or request something outside the normal policy.
Your workflow should define what happens when:
- Required information is missing
- Store data conflicts with the customer’s message
- An integration is unavailable
- The customer disputes the available information
- The requested action cannot be completed
- The customer becomes frustrated
- The customer asks to speak with a person
A safe workflow should never invent an outcome or claim that an action was completed when it was not.
Design the human handoff
A workflow is incomplete until the handoff path is defined.
When a conversation moves to your team, the agent should receive:
- The customer’s original request
- A summary of the conversation
- The verified customer and order details
- The information already collected
- Any actions already attempted
- The reason for escalation
- The approval or decision required
This prevents the customer from repeating everything and helps the human agent continue from the correct point.
Test the workflow with real scenarios
Before activating automation, test each workflow using realistic conversations.
Include:
- Clear and complete requests
- Short or unclear messages
- Incorrect order information
- Multiple requests in one message
- Customers changing their mind
- Requests outside the policy
- Integration errors
- Requests for a human agent
Testing only the ideal path creates workflows that fail as soon as a real customer behaves differently.
Your tests should confirm that the AI asks only necessary questions, uses verified information, follows permissions and hands off correctly.
Measure and improve after launch
Workflow mapping does not end when automation goes live.
Track:
- How often each workflow is triggered
- How many requests are resolved automatically
- Where customers abandon the conversation
- Which steps cause repeated questions
- Why conversations are escalated
- Which actions fail
- Where human agents correct the AI
- How long each request takes to resolve
Use this information to remove unnecessary steps, improve instructions and update the knowledge available to the AI agent.
A clear workflow creates reliable automation
Customer-support automation should not begin with a list of chatbot replies. It should begin with a clear understanding of how your business makes decisions and resolves requests.
By mapping the information, actions, permissions, exceptions and handoffs involved in each workflow, you create a system that can automate routine support without losing accuracy or control.
The result is not simply fewer tickets. It is a more consistent support operation where customers receive faster help and your team becomes involved at the moments where their judgement matters most.