In 2025, a developer reported that Cursor's AI support agent had told users the company enforced a three-device limit - a policy that didn't exist. The message spread through developer forums. People cancelled subscriptions. Cursor's cofounder had to post a public response. The company quietly added a label to all AI support replies: "This response was generated by AI."
The agent was doing exactly what most AI email responders are built to do by default: reply to customer emails autonomously, without any human check before send. That architecture - the AI reads, drafts, and fires - is how Zendesk's agentic email AI works out of the box. It's the default in Lindy's email automation too. When the AI gets it right, nobody notices. When it doesn't, the wrong email is already in your customer's inbox.
An AI email responder that keeps a human in the loop works differently. The AI drafts. A human reviews. Then it sends. That single change - adding a review step before delivery - is the difference between an AI agent that accelerates your team and one that creates incidents you spend a week repairing.
What does "human in the loop" mean for an AI email responder?
A human-in-the-loop AI email responder means the AI generates a draft reply, and a human approves it before it reaches the recipient. The AI does the reading and writing. The human is the final decision-maker on what gets sent.
The alternative is autonomous send: the AI reads the inbound email, writes a reply, and delivers it without any review. Both modes use the same AI drafting capability. The only difference is whether a person sees the output before the customer does.
The two architectures side by side
In autonomous mode, the sequence is: email arrives - AI drafts - email sends. No human step. In human-in-the-loop mode: email arrives - AI drafts - human reviews - email sends. The review typically takes 10 to 30 seconds per email.
Those 30 seconds are what stand between your AI's draft and your customer's inbox. For emails that touch pricing, policy, account status, or anything that could be legally binding, that checkpoint is the entire safety margin you have.
Why sequencing matters more than you'd expect
A 2024 study published in Cognitive Research: Principles and Implications tested what happens when humans review AI-generated outputs. The finding matters for anyone designing an approval workflow: when participants saw the AI's answer before forming their own judgment, decision accuracy dropped to 36.8%. When they judged independently first and then reviewed the AI output, accuracy held at 66.2%.
The implication for email: if your "human in the loop" means a reviewer reads whatever the AI drafted and clicks approve, you're getting 36.8% accuracy on the errors that matter - not 66.2%. The review has to be substantive to catch errors. Design for that or the checkpoint is a formality.
What goes wrong when an AI email responder sends without human approval?
Cursor's situation was bad but recoverable - a fabricated policy claim that spread through forums before someone caught it. The Air Canada case went further. The airline's chatbot told a passenger he qualified for a bereavement fare discount that didn't exist. Air Canada argued the chatbot was a "separate entity" from the company and therefore not bound by what it said. The British Columbia Civil Resolution Tribunal rejected that argument and ruled Air Canada liable for the misinformation. What the chatbot said became the airline's legal obligation.
Both incidents share the same root: an AI reply reached a customer without anyone on the team seeing it first.
Three failure modes that show up in practice
Factual errors are the most visible. The AI states something that isn't accurate - a wrong policy, a deadline that moved, a feature that doesn't exist. Customers cite these when they dispute charges or file complaints, and they're hard to walk back once the email is sent.
Tone errors are subtler but costly. A technically accurate reply that reads as cold or dismissive is still a bad reply. The AI has no relationship with the customer - it doesn't know this person just had a bad experience, or that they're a high-value account, or that this is the third time they've reported the same issue. A human reviewer catches that in three seconds.
Context errors are the easiest to miss. The AI doesn't know about the pending resolution, the recent escalation, the credit already applied to the account. These don't look wrong to the AI, so it doesn't flag them. They're immediately obvious to anyone on your team who knows the customer.
All three are avoidable with a 30-second human review. None are avoidable without one.
Why do Lindy and Zendesk default to autonomous send?
Their product positioning is built around full automation. Zendesk's agentic AI for email, released June 2025, is marketed as the ability to "handle customer emails end to end, delivering a resolution without human intervention." Lindy's email automation defaults to autonomous send, with human-in-the-loop available as an option you configure yourself.
This is a product choice, not an oversight. Full autonomy is easier to sell and easier to demo. The risks are real but they're downstream of the purchase decision.
What "configurable" HITL costs you in practice
Lindy's HITL documentation describes adding an approval step inside the workflow builder. You decide when it triggers. You build the routing. You maintain it as the product evolves. If you skip it or set it up wrong, the default behavior kicks in: autonomous send.
For teams that have the time and skill to build this correctly, it works. For teams that can't or don't - the default is what your customers experience.
Zendesk's architecture handles escalation differently: human intervention in agentic email happens when the AI can't resolve a ticket, not before the AI replies. An unresolvable thread that escalates to a human has already sent at least one AI-generated reply without review. Escalation isn't the same as approval.
Neither product makes human review the built-in default for all outgoing customer email. Both require something from your team to get there.
How does a human-in-the-loop email workflow actually run?
The well-designed version looks like a lightweight review queue, not a second inbox. Here's the sequence:
- Email arrives. The AI reads it, pulls relevant customer context, and writes a draft reply.
- The draft enters a review queue visible to the right team member.
- The reviewer reads the draft - typically 10 to 30 seconds - edits if needed, and approves.
- The email sends.
The review step is only a bottleneck if the queue is poorly designed or if every draft needs significant editing. For a well-set-up workflow handling a defined set of email types, most drafts need minor adjustments or none at all.
What makes a HITL review fast enough to be real
The review step only works if it's fast enough to do properly. A few things determine whether it is.
Context has to be on one screen. The reviewer should see the inbound email, the draft, and relevant customer history without switching tabs. If finding the context takes longer than reading the draft, reviews get skipped.
The AI should be drafting well enough that reviewers are adjusting the last sentence - not rebuilding from the second line. If most drafts need significant editing, the AI isn't configured well enough for the email types you're automating. Fix the training, not the review process.
There also has to be a clear escalation path. Some emails should never go through AI at all - active disputes, legal language, anything sensitive enough that a draft reply is worse than no reply. The workflow needs to route these for full human handling rather than generating something that'll be scrapped anyway.
Get those things right and a team handling 60 emails per day can run real review coverage without it becoming a separate job.
What the Uplift approach looks like
Uplift builds and runs the complete workflow - the AI drafting layer, the review queue, routing logic, and escalation rules. You describe your email volume and the types of emails your team handles. Uplift scopes it, builds it, and keeps it running as the tools and APIs underneath it change.
Human-in-the-loop is the default, not a configuration option. Your team doesn't build the approval step. They use the queue.
The agentic workflows overview covers the spectrum from manual to fully automated and where HITL sits in that taxonomy. For concrete examples of what AI agents handle across sales, ops, and support, see the agent examples. If your team also runs Slack alongside email - same autonomous-vs-HITL choice, different channel - the Slack agents breakdown covers how that plays out.
Frequently asked questions
What is an AI email responder with human in the loop?
A human-in-the-loop AI email responder is one where the AI drafts replies but a human reviews and approves each one before it sends. This differs from autonomous AI email, where the AI sends without review. The review step typically takes 10-30 seconds and lets reviewers catch factual errors, tone problems, or missing context before they reach customers.
What are the risks of letting AI send customer emails without human review?
The main risks are factual errors (AI stating incorrect policies, prices, or deadlines), tone errors (accurate replies that read as dismissive or cold), and context errors (replies that miss important customer history). Companies are legally liable for what their AI says to customers - a principle established by the Air Canada chatbot ruling in 2024, where the airline was held responsible for a chatbot's fabricated discount policy.
How does Lindy handle human in the loop for email automation?
Lindy supports human-in-the-loop email but as a configurable option you add yourself in the workflow builder. The default is autonomous send. This means your team needs to correctly build, test, and maintain the HITL step for every email use case you automate - and if you miss one, the default kicks in.
Is a human-in-the-loop email workflow slower than fully autonomous AI?
Slightly, by the time it takes a person to read a draft - typically 10-30 seconds per email. For a 30-email queue, that's roughly five to fifteen minutes of active review across the team. Compared to composing emails from scratch, HITL AI is still significantly faster. Most teams using this approach handle two to three times the email volume with the same headcount.
Can Uplift set up a human-in-the-loop AI email responder for my team?
Yes. Uplift builds the complete workflow - the AI drafting layer, the review queue, routing logic, and escalation rules - based on your team's specific email types and policies. Human-in-the-loop is built in by default, not an add-on you configure. You describe what your team handles; Uplift builds and runs it.