An after-hours chatbot workflow should route conversations by intent and urgency, not send every visitor the same away message. At 1:17 a.m., one customer may want a person, another may need an order update, and a third may report an urgent problem. Each conversation should end in one clear state: answered, handed off, captured for follow-up, or marked urgent.
Your after-hours chatbot needs a decision tree
Support managers need overnight coverage without creating a vague queue for the morning team. The workflow must decide what the AI support agent can answer, what information it should collect, and what the team should review first.
Use four decisions:
- What does the customer want? Identify the intent, such as a routine product question, order-status request, booking request, billing concern, or request for a person.
- How urgent is it? Separate questions that can wait from service interruptions, safety concerns, time-sensitive order problems, or other issues your team defines as urgent.
- Does the customer want a human? Treat that request as a routing signal, even when the agent knows the answer.
- Can the agent answer from approved content? If the answer is missing or unclear in the knowledge base, capture the issue instead of guessing.
This follows the basic after-hours routing pattern described in Velaro’s guide to business hours and agent availability and the priority model in Crisp’s after-hours customer service guide. The operating rule is simple: every route must end in a state the morning team can work.
Start with four conversation states
Use these states as the backbone of the workflow.
| State | Use it when | Next step |
|---|---|---|
| Answered | The question is routine and supported by approved content | Give the answer and offer a relevant next step |
| Handed off | The customer requests a person or the request needs human judgment | Preserve the thread for a human response |
| Follow-up needed | The agent cannot confirm the answer or the customer needs a later response | Collect contact and issue details |
| Urgent review | The conversation matches your urgent-issue rules | Mark it for the team’s first review |
The inputs are intent, urgency, a human request, and answer confidence. The state is the output. Do not let the agent end an unresolved conversation without choosing one.
For routine questions, use approved training sources. AssistLoop can use website content, uploaded PDF, DOCX, or TXT files, pasted text, and exact Q&A pairs. Exact Q&A pairs help when wording matters, such as a refund rule or policy that should not be paraphrased. See AssistLoop’s training sources.
A question that needs a human moves to handoff or follow-up. An unanswered question does not become an improvised answer. An urgent issue enters review with the details needed for the team’s first look.
Route each state to a specific customer-facing response
Write the response for each path before you put the workflow live.
Routine answer: Answer from approved content and offer the next useful step, such as a relevant page, a booking option, or a request for more detail. If the answer is not in the approved content, move to follow-up.
Request for a human: Acknowledge the request directly. State that the support team is offline, then collect the customer’s name, preferred contact method, issue summary, and desired outcome. Use human handoff when the request needs a person.
Unresolved issue: Say what the agent cannot confirm. Ask for the customer’s name, contact method, affected product or order, issue description, and requested action. This gives the morning team something specific to review instead of an unanswered question.
Urgent issue: Ask only for details needed to assess the issue. Capture the description, contact details, affected product or order, and information your team requires for first review. Mark the conversation as urgent, but do not promise an immediate human response when nobody is monitoring the channel.
Define what happens when no human is available
Set the after-hours message around the real support window. State when the team is offline and what the agent can do now. Do not imply that a live agent is watching the conversation if that is not true.
Before an unresolved conversation ends, collect at least:
- The customer’s name and preferred contact method
- A plain-language description of the issue
- The affected product, account, or order
- The action or outcome the customer wants
Use human handoff for requests that need a person. The failure condition is easy to spot: a customer asks for a person, provides contact details, and still receives no clear follow-up state. That chat is neither resolved nor ready for review.
Build the morning follow-up queue
The morning queue should be an ordered worklist, not a transcript dump. Start with urgent issues, human-handoff requests, unresolved conversations, repeated questions, and chats where the agent could not answer confidently.
Order the queue by urgency first. Then consider any follow-up commitment your team made. After that, consider the number of customers affected when that information is available. A possible outage may deserve attention before several unrelated product questions.
When a teammate follows up, use the preserved context. The customer should not need to explain the same issue again. The responder should see what the agent asked, what the customer supplied, and why the conversation entered the queue.
Repeated overnight questions are training signals. Turn them into approved website content, an uploaded source, pasted text, or an exact Q&A pair using AssistLoop’s training sources. Review the next occurrence to see whether the new source answers the question.
Add actions for requests that should finish overnight
Some requests can finish without a morning response. Use Agent Actions for tasks connected to an endpoint, such as booking a meeting or checking order status.
Separate information requests from actions that change customer data, create a commitment, or require human judgment. A support-hours question can use approved content. Booking a meeting changes a calendar. Checking an order returns data from another system.
For a booking request, collect the required details and complete the booking when the connected action supports it. If the action fails, move the conversation to follow-up with the details already collected.
For an order-status request, return only the information the connected system provides. If it returns an exception, an unclear result, or no matching order, send the conversation to the morning queue rather than inventing a status.
Every action needs a clear input, a defined output, and a failure path. If one is missing, keep the request in follow-up.
Test the workflow before you trust it
Run these tests with the same widget and content customers will use. Read the resulting conversations, not only the final state.
- A visitor asks for an agent immediately. Confirm the conversation enters handoff or follow-up with its request and previous context intact.
- An urgent issue arrives outside service hours. Confirm the agent collects the required details and marks the conversation for first review without promising instant human help.
- A customer gives an incomplete contact request. Confirm the agent asks for the missing detail and does not treat the conversation as ready for follow-up.
- The answer is missing from the knowledge base. Confirm the agent says it cannot confirm the answer and captures the issue instead of inventing a policy.
- A routine question, booking request, and order-status request arrive. Confirm each reaches the intended answer or Agent Action path. A failed booking or order lookup should enter follow-up.
A failed test usually points to an overly broad intent rule, a missing required field, or an unwritten failure path. Fix the rule and run the test again.
Put the workflow into AssistLoop
Start by training an AI support agent on your own content. Add the sources customers actually use, then place the agent in your website widget for overnight conversations.
Use human handoff for requests that need a person. When the team returns, work the queue by state instead of opening conversations at random.
Add Agent Actions only when the request has a clear input, output, and failure path. Booking meetings and checking order status can finish overnight when the connected action returns a reliable result. Exceptions belong in follow-up.
Choose a small set of real after-hours questions, test the four states, and inspect what the morning team receives. Then create your AI agent and put the workflow through those tests before expanding its scope.
FAQ
How should a chatbot handle customer support after hours?
It should identify the customer’s intent, urgency, and request for a human before choosing a response. Routine questions can receive approved answers, while unresolved or urgent issues should retain their context for follow-up.
Can a chatbot provide overnight customer support?
Yes. An AI agent can answer questions from approved company content through a website widget outside the support team’s normal hours. It can also collect customer details and prepare conversations for the next review period.
When should an after-hours chatbot hand off to a human?
Offer handoff when the customer asks for an agent, the answer is not supported by approved content, or the request needs human judgment. The conversation history should travel with the handoff so the customer does not have to start over.
What should a support team review the next morning?
Start with urgent issues, human-handoff requests, unresolved conversations, repeated questions, and chats where the agent could not answer confidently. Use those conversations to follow up with customers and improve the agent’s training content.
Written by