What happens when you copy your planner into the thread
Published on
A booking link is a fine invention for the moment someone has already decided to come and see you. Inside a running email exchange with a client it feels different: a little too much “sort it out yourself”. This article is about the alternative we built, and mostly about the rule that makes it usable: the language model may write the sentence, but never the time.
Where the booking link stops
A booking link works beautifully once someone has decided they need to see you. A patient making an appointment, a lead booking an intro call from your website: one link, done.
There is a second kind of conversation where it fits less well. You are in an email exchange with a client, there is already plenty in it about the work itself, and then a time needs to be found. Sending a link feels a little too much like “sort it out yourself”, so you type three times instead. After which two of them do not work, you look at your calendar again, and four emails go by for a half-hour call.
That back and forth is the gap. Not the booking, but the arranging around it.
The idea: cc instead of link
You get your own planner address. Instead of sending a link, you put that address in the cc of the email you were writing anyway. From that moment it reads along.
From: you
To: your client
Cc: your plannerShall we have a call next week about the project? My planner is in the cc.
The planner replies in the same thread with three times that are genuinely free in your calendar at that moment. If the other party picks one, the appointment is set, they get the ordinary confirmation email with a link to move or cancel it, and it appears in your calendar. They installed nothing, created no account and clicked no link.
The rule that decides everything
Here is the heart of it, and it is a design choice rather than a technical detail: the language model writes the sentence, the code picks the time.
The model reads the email and answers one question: what kind of message is this, and what wish does it contain? Nothing more. Which times exist is decided by the same calculation the booking page uses: working hours, exceptions, buffers, minimum notice, existing appointments. The model then gets those times back as numbered placeholders:
Anna is free at {{1}}, {{2}} or {{3}}. Which one suits you best?
The code then fills in the real times and checks what is left. If there is a date, a time, a weekday or a duration anywhere that did not come from a placeholder, that email does not go out and the agent falls back on a dull template.
The effect is that the quality of the model decides how pleasant the email reads, and never whether the appointment is right. That difference is exactly what makes this kind of feature usable or not.
What went wrong when it met real mail
We let it loose on our own mailbox for an afternoon. Four times the model wrote something that was not true, and all four times it stayed in a draft instead of reaching a client. They are instructive enough to write down.
“This week”, when it was next week. The three times were right, the sentence around them said “this week” and the appointments were six days later. There is no date in that sentence, so the check saw nothing. Relative periods are now part of it, in three languages.
A reference to times that were not there. To “I was thinking in about a month?” the model wrote: “Could we move this time to one of these moments?” No invented date, and nonsense all the same: there were no moments underneath. Text without times may no longer point at times either.
The signature read along. Below one email sat phone numbers, an address, two logos and the line “available on Monday, Tuesday and Thursday”. That last one is not a wish about this appointment, but the model reads it anyway. It also meant we were storing data we do not need. The snippet now stops at the signature.
Gmail wraps its attribution line. “On Tue, 4 Aug 2026 at 12:43, Dennis Tijnagel <…>” sits on one line and “wrote:” on the next. Our detection looked for the verb on the same line, so the whole quote came along.
What stands out: none of the four were on the calendar side. The times were right every time, because they come from the code. All four were in the clutter that email programs wrap around a message, and in a language model’s urge to finish a sentence.
Why you press send first
By default nothing leaves without you having seen it. The agent writes, the message waits in the app, you press send. Not because we do not trust it, but because the first few times a machine emails your client on your behalf are exactly the times you want to know what it says. See the four mistakes above: all four sat waiting on a screen rather than in someone’s inbox.
Once you have watched it work for a while, you switch that approval off with a checkbox. And there is one exception we do not hang on a button: cancelling and moving take away something that exists, on the strength of an email only a language model has read. Those two only happen at the moment you approve.
What it deliberately does not do
- No negotiating with the other party’s calendar. It proposes from yours. Access to your counterpart’s calendar takes permission nobody grants, and it would turn the product into something your client needs an account for.
- No public address. You put it in the cc; it is not an address for your website. A planner address lying around in the open is a spam target without an owner.
- No reading attachments or calendar invitations. Not yet, and possibly never.
- No endless conversation. Twenty messages in one thread, five within ten minutes: after that it stops and you get a note. A conversation that has not produced an appointment after twenty messages will not get one out of number twenty-one.
Where the mail arrives
We receive on our own mail server in Amsterdam. That was a deliberate choice: the usual solution is an American service that turns inbound mail into a webhook, and that is precisely the kind of party we warn about in our piece on the CLOUD Act. With the mail agent it is not even an email address at stake but the content of a conversation between you and your client.
Of that conversation we keep as little as possible: the first part of the text without the quote and without the signature, who took part, and what the agent understood from it. The email itself leaves the mailbox after processing, and after ninety days the rest is gone too. What you do not keep cannot leak.
Who this is for
For anyone who spends the day coordinating with others: consultants, recruiters, account managers, anyone whose calendar does not feel like their own. If your days fill up through a booking page, little changes for you; a link was already enough.
And it is not an assistant that takes over your work. It takes over one piece, the piece where three people try to guess each other’s calendar, and it does that in the language of the thread. The rest of your email stays yours.
Frequently asked questions
Can the agent do whatever it likes with my calendar?
It only books the appointment type you point it at, on times that are genuinely free according to your working hours, buffers and existing appointments. Moving and cancelling only happen after you approve the message. Booking happens straight away, because between "yes, Tuesday works" and your click someone else can take that slot.
What if someone spots my planner address?
The address is personal and contains a random part, so it cannot be guessed. On top of that, the agent only acts when your own address is in the email, as sender or as recipient. A forwarded thread you are no longer part of gets nowhere. And you can refresh your address with one click; the old one then stops existing.
Does my client see that they are talking to a machine?
Yes. Below every message there is one line: sent by the scheduling assistant of [your name]. That is not modesty but a rule from the EU AI Act, and in practice it is also good advertising: your recipient’s next question is usually what you are using for that.
What happens to the content of those emails?
The mail arrives on our own mail server in Amsterdam, not at an American service. We keep the first part of the text without the quoted part and without the signature, for at most 90 days; the email itself leaves the mailbox after processing. To work out what is being asked, that piece of text goes to our language model at Mistral in France. If you would rather not, take the address out of the cc.
Sources
This article explains regulation in plain language and is not legal advice. For your own situation, consult a lawyer.