Review an email bot's replies before customers get them

When this is done, your email bot writes its replies as drafts instead of sending them. An agent checks each draft and sends it as it is, edits it, or writes a different reply. Supervisors compare the bot's drafts with what agents sent, and you improve the bot until it's ready to answer customers on its own.

Customer sends an email
  → the Conversational bot writes a draft reply – nothing is sent yet
  → Email Draft Created sends the email to a queue
  → an agent sends the draft, or edits it and sends it
  → a supervisor compares the draft with the reply that was sent
  → when the drafts are good enough, draft mode is turned off

It takes about 30 minutes. Roles involved: Administrator (Part 1), Agent (Part 2), Supervisor (Part 3).

Draft mode is a human review of the bot's replies while you test it. It doesn't replace Agent Assist.

Before you start

What	Who	Where
Access to the Email section of the portal	Administrator	Roles
An email workflow with a Conversational bot node, its instructions and a knowledge base	Administrator	Steps 1–6 of AI bot with handoff to an agent, on the Email tab
The email template the bot uses, with the #BotResponse field	Administrator	Email conversational bot response
A queue in the workflow, and agents who can take its emails	Administrator	Build your first email workflow
A supervisor with access to Historical Reports	Administrator	Roles
A mailbox to send test emails from	Tester	–

Part 1 – Set it up (Administrator)

1. Put your signature in the bot's email template

The draft is built from the bot's email template only – the agent sends it as it is. Anything a reply needs besides the bot's answer, such as your greeting, signature or branding, goes in the template.

1. Go to Email › Templates and open the template your Conversational botConversational Bot – A Bot powered by natural language processing that can understand and respond to customer inquiries in a conversational manner, often using LLM technology. node uses.
2. Around the #BotResponse field, add your signature and anything else every reply should have. Do it for each language of the template.
3. Save the template.

Expected result: the template still contains #BotResponse, with your signature after it.

See Configure templates.

2. Turn on draft mode

1. Go to Email › Workflows and open the workflow.
2. Open the Conversational bot node and turn on Enable draft mode.

Expected result: the node's Email Sent output is replaced by Email Draft Created, in red until it's connected.

See Conversational bot (email).

3. Send the drafts to a queue and save

1. Drag from Email Draft Created and choose Queue – or connect it to the queue the bot already hands over to. Set the Routing mode and add an Attribute requirement that your reviewing agents have.
2. Make sure Low confidence (if you set a confidence thresholdConfidence Threshold – A configurable setting that determines the minimum confidence level for automated bot decisions or actions.) and Failed also lead to a queue, so no email is left without an answer.
3. Connect the queue's Agent Disconnected output, for example to End conversation.
4. Click Save.

Expected result: Workflow successfully updated! If Save is disabled, an output isn't connected – look for a warning triangle on a node. See Saving a workflow.

Part 2 – Review the drafts (Agent)

4. Send two test emails

1. From your test mailbox, send an email to the channel's address with a question the knowledge baseKnowledge Base / QnA Bot – A knowledge management system with AI bot capabilities to automatically answer common customer questions. answers.
2. Send a second one with another question from the knowledge base.

Expected result: no reply reaches your test mailbox yet – the bot only writes drafts.

5. Send a draft as it is

1. Log in to the Agent applicationAgent Application – The Agent facing website of Graia providing the UI elements to consume its feature set. as an agent with the queue's attribute, and set your state to Available.
2. Accept the first email task.
3. Above the bot's reply, check the notice The following email draft was generated for your review. The customer's email is quoted below the draft.
4. Read the draft and click Send.

Expected result: your test mailbox gets the reply, with the signature from the template.

See Review a reply drafted by a bot.

6. Edit a draft and send it

1. Accept the second email task.
2. Click Edit. The draft opens in the reply editor, with the customer's email quoted below it.
3. Change some text, then click Discard and confirm.
4. Click Edit again, change a sentence, and click Send.

Expected result: after Discard, the bot's draft is back as it was – only your edits were thrown away. Your test mailbox gets the edited reply.

Part 3 – Compare and improve (Supervisor)

7. Compare the drafts with what was sent

1. In the portal, go to Historical Reports › Conversations and filter by the email media type.
2. Click View and download conversation history (eye icon) on the second test email.
3. Find the bot's draft: it's labeled AI and has a dashed border. It was never sent to the customer.
4. Below it, find the reply the agent sent, labeled with the agent icon.
5. Click the AI draft.

Expected result: you can see what the agent changed. The Event Flow panel opens with the steps behind the draft – for example the knowledge base search and its results.

See Compare bot drafts with the agent's reply and Conversational bot decision trace.

8. Improve the bot, then turn draft mode off

Keep draft mode on while you collect real customer emails. Look for patterns in what agents change: missing facts point to the knowledge base, wrong tone or wrong topics point to the instructions.

1. Update the knowledge base or the node's Instructions – see Controlling AI processes.
2. When agents mostly send the drafts unchanged, open the workflow and turn off Enable draft mode.
3. Connect the Email Sent output that comes back, for example to End conversation, and click Save.

Expected result: the bot's replies now go to customers directly. Emails it can't answer still leave through Low confidence or Failed to your queue.

Check that it works

#	Do this	You should see
1	Turn on Enable draft mode on the node	Email Draft Created replaces Email Sent
2	Send an email to the channel	No reply yet; an email task with the bot's draft reaches an agent
3	Open the task as the agent	The notice The following email draft was generated for your review., the draft, and the customer's email below it
4	Click Send	The customer gets the draft, with the template's signature
5	Click Edit, change the text, Discard	The bot's draft is back as it was
6	Open the conversation in Historical Reports › Conversations	The AI draft with a dashed border, and the agent's reply below it

If it doesn't work

Symptom	Likely cause	Fix
The customer gets the bot's reply directly	Draft mode is off, or a different workflow handles the channel	Step 2
Save is disabled in the workflow	Email Draft Created or another output isn't connected	Step 3
No agent gets the draft	No available agent has the queue's attribute	Step 3
The reply has no signature	The signature isn't in the bot's email template, or not in that language	Step 1
The bot's answer is missing from the draft	The template doesn't contain the #BotResponse field	Step 1
An email reached the queue without a draft	It left through Low confidence or Failed – the agent answers it as a normal email	Step 7, to see why
After turning draft mode off, Save is disabled	Email Sent isn't connected	Step 8

Related

• Email workflows: Conversational bot
• Email templates
• Reply to customer emails (Agent)
• Conversation History
• AI bot with handoff to an agent
• Controlling AI processes