Auto Closer

What the Auto Closer does

Support inboxes fill up with mail that is not support: out-of-office replies, delivery failure notices, SEO agencies, "thanks!". The Auto Closer closes those silently. No reply is sent, the customer hears nothing, and the conversation never reaches your queue.

You configure it under Settings, then Auto Closer.

This is the single biggest lever on how your inbox feels. On a typical email setup it is a meaningful share of everything that arrives, and none of it needed a human.

Two layers, and they work differently

There are two independent mechanisms, and knowing which one is doing the work saves a lot of confusion.

Blacklist rules 🚧 run first, on the way in, before any AI touches the conversation. They are plain conditions you write yourself: if any one of them matches, the conversation is closed immediately. They are exact, instant, and free.

The classifier 🤖 is an AI step that reads the message and decides between "this can be closed" and "this needs a response". It runs ahead of everything else in the pipeline, so mail you were going to throw away does not get enriched, analysed and drafted first.

Use blacklist rules for anything you can describe by a fact. Use the classifier for anything that depends on what the message actually says.

Blacklist rules

A rule is a condition on the conversation. The classic example is From Email contains fedex.com, which closes every carrier notification without thinking about it.

Rules are evaluated with "any of them" logic. One match is enough, so each rule can be narrow and specific rather than trying to cover everything.

This is the right place for senders you recognise: carriers, your own order confirmation system, a monitoring tool that mails your support address, a newsletter platform. They are predictable, and a rule that never has to guess never gets it wrong.

The categories

Each is a separate toggle, so you can introduce them one at a time.

• Out of office 🏖 Vacation and absence replies. Detected across languages, so Danish "ferie" and German "Urlaub" are caught as well as English.
• SPAM 🚫 Bulk promotion, phishing, impersonation attempts, suspicious links, romance scams. "We can improve your Google rankings in 7 days".
• Unsolicited marketing and sales 📣 A company trying to sell you something or reach someone internally to pitch. "Can we schedule a demo?"
• Auto-replies ↩️ Machine-generated acknowledgments. "Your message has been received", ticket confirmations from another system, bot receipts.
• Courtesy 🙏 A closing thank you with nothing in it. "Thanks for your help", "Appreciate it".
• Bounce notifications 📭 Delivery failure notices. "Undelivered Mail Returned to Sender". These are detected directly rather than by the classifier, so they are caught reliably.

Custom examples

Below the categories you can add your own examples of messages that should be closed without a reply. These are for the things that are noise in your business specifically and would not be noise in anyone else's.

The important constraint: the AI may only use a custom close when the message closely matches one of the examples you wrote. It is deliberately not allowed to treat "custom" as a catch-all for anything it cannot categorise. When nothing fits, the ticket goes to your team.

Write examples as the kind of message you actually receive, not as a description of a category. "Order received via our B2B portal, no action needed" is useful. "Irrelevant stuff" is not.

How careful it is

Two behaviours are built in to protect you from over-closing.

When in doubt, it does not close. The classifier is instructed to choose "needs response" whenever the case is unclear, and never to invent a reason to close something that does not clearly match one of your enabled categories.

Attachments push towards keeping it. A message with something attached is treated as a strong signal that a human should look. A photo of a damaged item is not spam, even if the words around it are thin.

What it does to your numbers

Auto-closed conversations are flagged as noise, so they can be kept out of your volume and automation figures. This matters: without it, turning on spam closing would look like your automation rate suddenly improved, when really you just stopped counting junk.

Noise closes also get no satisfaction survey. Nobody is asked to rate the spam they sent you. See CSAT.

Note: Courtesy is the exception. A "thank you" close means the case really was resolved, so a survey is still sent.

Seeing what it did

Use the View Runs button on the settings page, or open Auto Closer runs directly. You get every auto-close over a date range, filterable by market, with the reason each one was closed for.

This is the page that makes the feature safe to trust. Reading a week of it tells you more about whether your categories are right than any amount of reasoning about them.

The order to switch things on

Start with bounces and auto-replies. These are unambiguous machine traffic. The risk of closing a real customer is close to zero, and they are often the largest single category.

Add out of office and blacklist rules for your known senders next. Also very safe, and between them these four usually account for most of the noise.

Then spam and marketing. Read the runs page for a week before and after. The thing to watch for is a genuine B2B enquiry being read as a sales pitch, which happens if you sell to businesses.

Leave courtesy until last, and think about it. Closing "thank you" is tidy, and for most teams it is right. But some customers write "thanks, that worked, though one more thing" and the decision is made on the last message. If your customers tend to write long, hold off.

Good habits

Turn on one category at a time. If something goes wrong you want to know which switch did it.

Prefer a blacklist rule to an example. Anything you can describe with a sender address or a domain should be a rule. It is faster, cheaper and it cannot be misread.

Revisit the runs page monthly. Your inbound mix changes as you add channels, markets and partners, and a category that was safe in January can start catching real customers in June.