KoltrixThe limits of AI email sorting: edge cases to expect
Six situations where AI email sorting predictably gets it wrong, why each one fools the model, and the small habits and rules that cover the gaps.
On this page(8 sections)
- Edge case 1: The customer writing from a personal address
- Edge case 2: The investor update that looks like a newsletter
- Edge case 3: The cold pitch from a real partner
- Edge case 4: The forwarded thread
- Edge case 5: The automated email that's actually urgent
- Edge case 6: Replies to your own marketing email
- The pattern behind all six
- Bottom line
AI sorting fails in patterns, not at random. Once you know the patterns, you can stop being surprised by them and set up a few cheap safeguards instead.
This teardown walks through six edge cases that show up in almost every business inbox. For each one: what happens, why the model gets it wrong, and what to do about it.
Edge case 1: The customer writing from a personal address
What happens. A paying customer emails from [email protected] instead of their work address. The message is short: "Hey, can't log in since this morning, any ideas?" It ends up somewhere other than the main view, or gets treated as low priority.
Why it fools the model. Free webmail addresses are used by everyone: customers, job seekers, cold pitchers, scammers. With no reply history and no company domain to match against, the sender looks like a stranger. The content is brief and informal, which doesn't help.
What to do.
- Make a habit of glancing at lower-priority categories once or twice a day, not once a week.
- When you find one, move it and, if the person is a long-term customer, add a sender rule so it never happens again.
- In your signup or support flows, gently encourage customers to write from the address on their account. It helps you verify them too.
Edge case 2: The investor update that looks like a newsletter
What happens. Your investor, or a founder you advise, sends a monthly update through a newsletter platform. It has a nice template, an unsubscribe link and a hundred recipients. It gets filed with marketing newsletters.
Why it fools the model. Technically, the model is right: it is a bulk email sent through bulk infrastructure, with the List-Unsubscribe header and everything. The model doesn't know that you personally care about this sender.
What to do. This is exactly what sender rules exist for. One rule per important bulk sender, set once, solves it permanently. Keep a short list of these senders in your notes so new teammates understand why the rules exist.
Edge case 3: The cold pitch from a real partner
What happens. A legitimate potential partner, maybe a company you've been hoping to integrate with, sends a first email that reads like every other outreach message: a compliment, a pitch, a request for fifteen minutes. It gets sorted with cold pitches.
Why it fools the model. First contact from a company you want to hear from looks exactly like first contact from one you don't. The model can't know your strategy.
What to do.
- Skim the cold-pitch pile on a schedule. Twice a week is plenty for most teams. You're looking for names, not reading every message.
- If you're actively seeking certain partners, tell your team which company names to watch for.
- Accept that this one can't be fully automated. The cost of a five-minute skim is small compared to missing the one email that mattered.
Edge case 4: The forwarded thread
What happens. A colleague forwards you a long thread from a customer with "Can you handle this?" on top. Or a customer forwards an internal discussion from their own company. The sorting looks oddly off: the sender is familiar, but the content is someone else's.
Why it fools the model. The visible sender and the actual author are different people. Quoted history may include marketing footers, automated messages, or several unrelated topics. The model has to guess which part matters.
What to do. Prefer sharing a link to a thread (or giving mailbox access) over forwarding inside your own team. When customers forward things, read from the bottom up; the actual request is often in the newest message at the top, but the context is at the bottom.
Edge case 5: The automated email that's actually urgent
What happens. A payment failure notice, a domain renewal reminder, a security alert about a new login, a "your account will be suspended" warning from a vendor. All automated, all sent from no-reply addresses, all easy to file with routine notifications.
Why it fools the model. Automated, templated, transactional, sent from bulk infrastructure. Nearly every signal says "routine notification." Only the content says "act now," and that content looks a lot like the hundred routine notices before it.
What to do. This is the edge case where rules beat AI most clearly. Write explicit rules for the handful of senders whose alerts can hurt you:
from: [email protected] → label: Action-needed
from: [email protected] → label: Action-needed
subject contains "payment failed" → label: Action-needed
Then make sure someone looks at that label every day.
Edge case 6: Replies to your own marketing email
What happens. You send a product update to your users. A dozen of them hit reply with real questions, bug reports or feedback. Those replies carry your own newsletter's subject line and quoted template, and some get sorted as newsletters.
Why it fools the model. The quoted content is a marketing email, with all the patterns of one. The new part, a customer's two-line reply at the top, is a small fraction of the message.
What to do.
- Send product updates from an address that can actually receive replies, and route those replies to a shared mailbox.
- After any big send, check where the replies are landing for the next couple of days.
- If the replies matter (they usually do), a rule that catches replies to that sending address avoids the problem.
The pattern behind all six
Every edge case above has the same structure: the model judges by how an email looks, and these emails look like something they aren't. Personal addresses look like strangers. Investor updates look like marketing. Urgent alerts look like routine ones.
That gives you a simple rule of thumb:
| If the signal the model sees is... | ...but you know... | Use |
|---|---|---|
| Unknown sender | They're a customer | A sender rule after first contact |
| Bulk email | You care about this sender | A sender rule |
| Outreach wording | You want this company | A scheduled skim |
| Automated notice | It can cost money | A label rule plus a daily check |
| Quoted marketing | The reply on top is real | A reply-routing rule |
Bottom line
AI sorting does most of the work most of the time, which is exactly why its blind spots are easy to forget. Expect these six cases, cover the high-stakes ones with explicit rules, and keep a light, regular habit of skimming the lower-priority piles. That combination catches nearly everything without putting you back to reading every email yourself.
Start with Koltrix
Your domain, one inbox, and an API that sends.
A team inbox where AI sorts and drafts (nothing is sent without your click), plus the transactional API and SMTP relay your product sends with. 7 days free, no card.
