No Answer Means No: Why Our AI Never Sends an Email Without Approval

There is an action in TheroAI that cannot be undone: sending an email. A draft gets overwritten. A research result gets discarded. A running workflow gets stopped. An email that has left the server is out. It sits in the inbox of a customer, an applicant or a supplier, and no software in the world will bring it back.

That is why one thing was clear to us early on: an assistant that writes emails is useful. An assistant that sends the wrong email creates a mistake nobody can correct. One of the most important product decisions in TheroAI grew out of this contrast. We deliberately made the product slower at one step: the send. This article tells the story of how that decision looks in the product, why a five minute timeout plays the central role in it, and where we chose to set limits for ourselves.

The One Step Nobody Can Take Back

Language models write good emails. They hit the right tone, structure cleanly and write faster than any human. That is exactly where the risk sits. A model can write a convincing reply and pull a number from the wrong quote while doing so. It can address the wrong person in a long thread. It can quote an internal remark that should have stayed internal. Humans make these mistakes too. The difference: an assistant that prepares dozens of messages a day multiplies the opportunities.

One could hope that better models will solve the problem on their own. We think that is the wrong order. Models keep improving, but even an error rate close to zero remains an error rate. For reversible actions you can live with that. For sending, we did not want to.

So we drew the line where reversibility ends. Everything before the send may be fast and automatic: searching the mailbox, gathering context, writing the draft, revising, shortening. The send itself stops and waits for a human. The time savings live in the writing anyway. The final check before sending costs seconds.

From Text to Draft

If you have TheroAI write an email, you first work in an email canvas next to the chat. The canvas has fields for To, Cc and Subject, with the message body below. The assistant fills the fields from the context of the conversation, and you edit every line directly in the canvas.

Enlarge

The most important button on this screen is called "Create draft". It is deliberately not called "Send". The default path out of the canvas leads into the drafts folder of your own mailbox. The message sits there as a perfectly normal draft that you open in Gmail or Outlook, review and send yourself. For most situations this is the right working mode: the AI does the writing, the send stays entirely with you.

Sometimes, though, the assistant really should send. For example when you ask it to send a meeting confirmation directly to a colleague while you keep working. For this case there is the second mechanism.

The Amber Card

Every direct send, whether through the Gmail API or through Microsoft Graph, stops at an approval card before execution. The card appears in the chat, directly below the request, and it is amber. That makes it stand out visibly from normal replies. The card shows three things: the recipient, the subject and the full message body. Exactly as the email would leave the building. No summary, no shortened preview. The text on the card is the text that will be sent.

Enlarge

Below it sit two buttons: Approve and Decline. Nothing else. We deliberately decided against building approval as a global setting that you flip once and then forget. The decision about a specific email needs a look at exactly that email. Whoever reads the card checks the two questions that matter within a few seconds: Is the message going to the right address? Does it say what it is supposed to say?

Five Minutes of Silence Means No

In our view the most interesting design decision hides in an inconspicuous detail: the timeout. If nobody reacts to the approval card within five minutes, the send counts as declined.

No answer means no.

That sounds trivial. It is the opposite. Every system that waits for a confirmation needs a default for the case where the confirmation never arrives. There are exactly two candidates: execute when in doubt, or stop when in doubt. For an irreversible action, only one of them is defensible in our view.

Executing when in doubt would be more convenient. But it would turn every missed notification into a sent email. One meeting, one lunch, one phone in flight mode, and the message would be out. The costs of the two failure cases are extremely asymmetric. An email stopped by mistake costs you another attempt. An email sent by mistake costs you, in the worst case, a customer relationship or a confidential piece of information. With failure costs this unequal, there is only one defensible default in our view.

Why five minutes exactly? The window has to be long enough for a quick check, even if you are working in another app at that moment. And it has to be short enough that a forgotten card is not still armed hours later. An approval you did not grant in the morning should not be able to trigger anything in the afternoon. By then the context may long since have changed.

Declined Does Not Mean Lost

A strict default creates a follow-up problem. If the system stops when in doubt, what happens to the work? An assistant that throws away a carefully researched reply after five minutes would be a bad tool. Nobody likes having the same text written twice.

That is why there is a pressure valve: if no approval happens, TheroAI still places the finished text as a real draft into the drafts folder of your mailbox. Expired or declined never means that work is lost. You open your mailbox later, find the draft, read it calmly and decide yourself: send, edit or delete. The decline stops the send. The result is preserved.

This detail is what makes the strict rule workable in everyday use. A hard no without a pressure valve creates frustration, and frustration creates workarounds. A no that preserves the work as a draft costs almost nothing.

What We Deliberately Restricted

An honest description includes two limitations, and we set both on purpose.

First: forwarding currently exists for Outlook only. A forward passes someone else's text on under your name, which makes it a sensitive special case of sending. We enable write functions per provider, one by one, and only once we trust the behavior of the respective interface. For Microsoft Graph that is the case for forwarding; for the Gmail API it is not yet. Until then, forwarding stays switched off there.

Enlarge

This separation is visible in the AI tool settings. Reading tools such as searching the mailbox or opening individual messages run without a confirmation step. Everything that acts on the outside world carries the "Schreibend" badge and the hand icon for the mandatory confirmation. Administrators see at a glance which tools run freely and which ones stop at the approval card.

Second: alongside the personal mailboxes, TheroAI has a system mailer for automatic messages from workflows. It sends via SMTP and is restricted to your organization's own domains. Internal notifications to your own team work fine with it. A mass mail to external addresses technically cannot be built this way, regardless of what a prompt demands.

The Rule Holds When Nobody Is Watching

Workflows are the hardest test for a rule like this. A workflow runs while no chat window is open. A form gets submitted, a trigger fires, an agent works through its instruction. If a step in this chain wants to send an email through a personal mailbox, exactly the same mechanic applies: the run pauses mid-flight and creates an approval request. Only the approval resumes the run. Without approval, nothing is sent.

This mattered to us because automation otherwise becomes the back door. An approval requirement that only applies in chat would be window dressing. In TheroAI there is no path on which a model writes to the outside world through your mailbox without a human having seen and confirmed the specific message. Not in chat, not in a workflow, not at two in the morning.

What You Can Take Away

You do not need to use TheroAI to take something away from this story. If you evaluate an AI tool that can send emails or other messages to the outside world, five questions help:

• Which of the tool's actions are irreversible? The list is usually short, and outbound sending is almost always on it.
• Does the approval show the actual data, meaning recipient, subject and full text? A summary is not enough, because mistakes hide in the details.
• What happens when nobody reacts? Does the system execute when in doubt, or does it stop when in doubt?
• Is the work preserved after a stop, for example as a draft in your own mailbox?
• Does the approval requirement also hold in automations, when no human is watching?

Our answer to the third question is in the title of this article. We made TheroAI slower at one step, and we consider that step the best-invested waiting time in the entire product. An email that did not go out can still be sent at any time. An email that went out cannot be brought back by anyone.