Auto-Send vs. Approval Mode: Choosing the Right AI Email Workflow

When you set up an AI-powered email support system, one of the first decisions you face is how much autonomy to give the AI. Should every AI-drafted response be reviewed by a human before it reaches the customer? Or should the AI send responses on its own for certain types of questions? The answer is not binary, and getting it right has significant implications for response speed, quality, customer trust, and team workload.

This guide examines both approaches in depth, helps you understand the tradeoffs, and provides a framework for deciding which mode to use and when.

Understanding the Two Modes

Approval Mode (Human-in-the-Loop)

In approval mode, the AI drafts a response to every incoming customer email, but that draft goes into a review queue rather than being sent directly. A human agent reviews the draft, makes any necessary edits, and then approves it for sending.

The workflow looks like this:

1. Customer email arrives
2. AI analyzes the email and drafts a response using your knowledge base
3. The draft appears in the agent's review queue
4. The agent reviews the draft for accuracy, tone, and completeness
5. The agent approves, edits and approves, or discards the draft
6. The approved (and possibly edited) response is sent to the customer

Auto-Send Mode

In auto-send mode, the AI drafts a response and sends it directly to the customer without human review. The agent can still see what was sent in the conversation history and can follow up if needed, but there is no pre-send review step.

The workflow is simpler:

1. Customer email arrives
2. AI analyzes the email and drafts a response using your knowledge base
3. The response is sent to the customer automatically
4. The agent can review what was sent and follow up if necessary

The Case for Approval Mode

Quality Control

The strongest argument for approval mode is quality control. No AI system is perfect, and the consequences of an incorrect response range from mild embarrassment to genuine business harm. Approval mode gives you a safety net that catches errors before they reach the customer.

Common issues that human review catches:

• Factual inaccuracies: The AI may misinterpret a knowledge base article or apply information from the wrong context
• Tone mismatches: The AI might be too casual for a frustrated customer or too formal for a friendly inquiry
• Missing context: The AI might not account for information from previous conversations or account-specific details
• Hallucination: In rare cases, the AI may generate information that is not in the knowledge base at all
• Sensitive situations: Billing disputes, complaints, and churn risks require human judgment that AI cannot reliably provide

Building Trust Incrementally

When your team is new to AI-assisted support, approval mode lets everyone build confidence gradually. Agents see how the AI performs across many interactions, learn where it excels and where it struggles, and develop an intuition for when to trust the drafts and when to rewrite them.

This trust-building period is valuable. Teams that skip it and go straight to auto-send often pull back after the first significant mistake, and the resulting loss of confidence is harder to recover from than if they had started with approval mode from the beginning.

Training Data Collection

Every interaction in approval mode generates valuable feedback. When agents edit a draft, those edits tell you what the AI got wrong. When they approve a draft without changes, that confirms the AI got it right. Over time, this data helps you understand which topics, question types, and customer segments the AI handles well and which need more knowledge base investment.

The Case for Auto-Send

Speed

The most obvious advantage of auto-send is speed. Responses go out in seconds rather than minutes or hours. For customers waiting for answers to straightforward questions, this is a genuine improvement in their experience.

Research consistently shows that response time is one of the strongest predictors of customer satisfaction in support interactions. An accurate response that arrives in 30 seconds is almost always preferred over an equally accurate response that arrives in 30 minutes.

Agent Efficiency

When routine questions are handled automatically, your support agents can focus their time and energy on the complex issues that genuinely benefit from human attention. Instead of spending 60% of their time reviewing and approving AI drafts for simple questions, they can dedicate that time to nuanced problems, relationship-building with important customers, or improving the knowledge base.

Scalability

Auto-send scales linearly with email volume without requiring proportional growth in headcount. If your ticket volume doubles, auto-send handles the additional routine questions without any change in staffing. This makes it particularly valuable for teams experiencing rapid growth or seasonal volume spikes.

Finding the Right Balance

Most teams find that the answer is not "all approval" or "all auto-send" but a thoughtful combination of both. The key is matching the automation level to the risk and complexity of each interaction.

A Framework for Deciding

Consider these factors for each type of customer interaction:

Favor auto-send when:

• The question has a clear, factual answer that is well-documented in your knowledge base
• The stakes of an incorrect response are low (the customer can easily follow up or the error is minor)
• The topic does not involve money, account access, or sensitive personal information
• Your AI's accuracy for this type of question has been consistently high during the approval mode period
• The question is common enough that you have seen many examples of the AI handling it correctly

Favor approval mode when:

• The question involves billing, refunds, or financial decisions
• The customer is upset, frustrated, or threatening to leave
• The topic requires nuance, empathy, or judgment that goes beyond factual answers
• The information is complex or has important caveats and exceptions
• The AI has historically struggled with this type of question
• The customer is a high-value account where getting it wrong has outsized consequences

The Gradual Transition Approach

The safest way to introduce auto-send is incrementally:

Phase 1: Full Approval Mode Start with every draft going through human review. Use this period to establish baseline metrics and understand AI accuracy across different question types. This phase typically lasts 2-4 weeks.

Phase 2: Identify Auto-Send Candidates Analyze your data from Phase 1 to identify question categories where the AI consistently drafts accurate responses that agents approve without significant edits. These are your candidates for auto-send.

Phase 3: Limited Auto-Send Enable auto-send for your highest-confidence categories only. Continue using approval mode for everything else. Monitor closely for any quality issues.

Phase 4: Expand Auto-Send As confidence grows and your knowledge base improves, gradually expand the categories eligible for auto-send. Continue to keep sensitive and complex topics in approval mode.

Phase 5: Steady State Most mature implementations settle into a pattern where 40-60% of emails are handled via auto-send and the remainder go through approval mode. The exact split depends on your industry, customer base, and risk tolerance.

Configuring the Right Workflow in Relay

Relay gives you granular control over which emails go through approval mode and which are auto-sent. You can configure these settings at the mailbox level, adjusting the automation level based on your comfort and the results you are seeing.

Starting with Approval Mode

When you first set up a mailbox in Relay, approval mode is the default. Every AI-drafted response appears in your review queue. This is intentional: it lets you validate the AI's performance against your specific knowledge base and customer base before trusting it to send on its own.

Transitioning to Auto-Send

When you are ready to enable auto-send, Relay lets you do so at the agent configuration level. You can set confidence thresholds and category rules to control which types of responses are eligible for automatic sending.

Monitoring Auto-Send Quality

Even with auto-send enabled, Relay provides visibility into what is being sent. You can review auto-sent responses in your conversation history, track customer satisfaction metrics, and quickly disable auto-send if you notice quality issues.

Common Questions

What if the AI makes a mistake in auto-send mode?

When an auto-sent response is incorrect, the customer will typically reply with clarification or a correction. Your team can see the full conversation, identify the error, and respond with the correct information. Treat it the same way you would treat a mistake made by a human agent: acknowledge it, correct it, and investigate the root cause.

How accurate does the AI need to be before enabling auto-send?

There is no universal threshold, but most teams feel comfortable enabling auto-send when the AI's drafts are approved without changes at least 90% of the time for a given question category. For high-stakes categories, you might want 95% or higher accuracy before considering auto-send.

Should we tell customers they are receiving AI-generated responses?

This is a business decision that depends on your brand, industry, and customer expectations. Some companies are transparent about AI involvement and find that customers are fine with it as long as the answers are helpful. Others prefer to keep the AI in the background. Either approach can work.

Can we switch back to approval mode after enabling auto-send?

Yes. Switching between modes is a configuration change, not a one-way door. If you enable auto-send and notice quality issues, you can switch back to approval mode immediately. This flexibility is one of the advantages of the gradual transition approach.

The Bottom Line

The choice between auto-send and approval mode is not permanent, and it does not need to be all-or-nothing. Start with approval mode to build confidence and collect data. Identify the categories where the AI performs reliably. Gradually introduce auto-send where it makes sense. Keep approval mode for sensitive and complex interactions.

This measured approach gives you the speed benefits of automation for routine questions while maintaining human oversight where it matters most. Over time, as your knowledge base improves and the AI handles more categories reliably, you can expand automation at whatever pace feels right for your team and your customers.