AI agent draft mode: when to enable automatic sending

If you’re concerned that an agent might send an incorrect response, start in draft mode: the agent writes a reply, and a person reviews it before sending. Enable automatic sending only for specific queries whose answers your team has already approved.

In summary:

• In draft mode, the agent prepares a reply, and a person reviews it before sending.
• With automatic sending, the agent replies without prior approval. Use it only for queries and answers your team has approved.
• When reviewing a draft, check that the reply is correct, complete, in keeping with your company’s tone, and does not need human attention.
• Expand automation by query type once your team has confirmed the replies meet its criteria.

The difference is who approves the reply before it is sent. To assess the email channel, visit the UTIO Email page.

What is the difference between draft mode and automatic sending?

In draft mode, the agent writes a reply, but a person reviews it and decides whether to send it. With automatic sending, the agent sends the message without prior approval. The practical difference is who has the final say before the customer receives a reply.

In draft mode, the agent writes the reply while a person retains control over sending it: they can edit, add to or discard the text. With automatic sending, that prior control no longer applies to messages covered by the rule. Just because the agent can prepare a reply does not mean it should be allowed to send it.

If you are considering this workflow for another channel, see the WhatsApp agent. Set approval requirements according to the type of inquiry and the channel your team handles.

Why start with an AI agent in draft mode?

Starting in draft mode lets you check how the agent responds to real inquiries, spot incomplete answers or a tone that doesn’t fit, and fix those issues before giving it permission to send messages to customers. It also helps you identify which questions need human judgment.

To make your review useful, note what was missing, what was unnecessary, and how the agent should respond. This helps you see whether the issue is isolated or whether you need to adjust a rule for similar inquiries.

Look for three signs: the response leaves part of the question unanswered; it uses a tone that doesn’t fit your business; or it requires you to check information or make a decision that hasn’t been defined. In those cases, don’t automate it.

If you prepare reference responses, Studio for creating agents lets you create an agent by reading your website.

What should a person check before approving a draft?

Before approving a draft, check that it answers the customer’s question, that every detail is supported by the information available, that the tone fits your company, and that it does not expose personal or commercial information or handle a case that should be dealt with by a person.

Use this checklist before sending:

• Accuracy: verify names, terms, amounts and any other specific details.
• Context: check that the reply takes the message and previous conversation into account.
• Tone: make sure the text sounds like your company and does not promise anything you cannot deliver.
• Data: confirm that the message is going to the right recipient and does not include unnecessary personal or commercial information.
• Hand-off: if the case falls outside the process, requires judgement or involves an exception, pass it to a person. For support queries, see the support vertical.

If context is missing or you cannot verify a detail, ask for clarification or pass the case to a person. Do not fill in the gaps with assumptions.

When should you enable automatic sending?

You can consider enabling automatic sending for specific inquiries once your team has reviewed real examples and defined what information the agent can share without prior approval. Before expanding automation, check that the responses remain accurate and that there is a clear way to pass uncertain cases to a person.

Start with one type of inquiry that has a clear, limited answer. Decide in advance what the agent should do if an inquiry falls outside that scope.

• Enable sending only for the category your team has approved.
• Review the responses that are sent and note any cases that need changes.
• If an incorrect response or an exception comes up, pause that category and return to human review.

Expand automatic sending when your team confirms that the category meets the agreed criteria. Don’t assume that what works for one type of inquiry will work for others. If you need help preparing the rollout, see the implementation packs.

What messages should still be handled by a person?

Sensitive or ambiguous inquiries, and messages that fall outside defined procedures, should go to a person rather than receive an automatic response. It is also best to hand off a message when the agent cannot verify a detail, finds conflicting information, or the customer asks for an exception. This keeps an unverified answer from being presented as confirmed.

Set clear rules for identifying these cases. For example:

• The response could commit the team to a price, term, or date it has not confirmed.
• The customer makes a complaint or needs someone to consider an exception.
• The message contains conflicting information or does not make clear what the customer needs.
• The agent cannot find an answer supported by the available information.

In these cases, the agent should say that a person will continue handling the request rather than improvise. If some inquiries come in by phone, check what UTIO’s voice agent offers for that channel. The team can review handoffs regularly and update the instructions when a new case comes up.

Frequently asked questions

Does the agent stop handling messages if it only prepares drafts?

Not necessarily. In this mode, it can draft a reply for a person to review, but it will not send it without approval. The team must decide who reviews the drafts and how to handle messages while they wait. If a query cannot wait, assign it to a person using the usual process.

Who should approve the agent’s drafts?

The approver should know the procedure and be able to confirm that the reply is appropriate for the customer. The team can assign reviews based on the type of query, but it should be clear who is responsible and who covers the task when they are unavailable. This helps reduce pending messages and conflicting approvals.

How long should you keep draft mode on?

There is no single timeframe that works for every small business. Keep draft mode on while the team finds errors or questions that could change the reply. Then assess each type of query separately: if the agent responds consistently to one question, that does not mean it can automatically send other replies.

What should the team do if they disagree about a draft?

Do not send it automatically until you have clarified the criteria. Agree on what information the reply should include, what tone to use, and when to consult a person. If the disagreement depends on the customer’s specific situation, treat that query as an exception and review the procedure before automating similar cases.

Learn about UTIO Email and see which features it offers for your customer support workflow.

Sources

• UTIO · Agentes de IA para voz, WhatsApp y email — utio.io