KoltrixHow AI reads long email threads, and where it slips
Quoted history, signatures, context limits and who-said-what confusion: how AI models process long email threads, why summaries go wrong, and how to help them.
On this page(6 sections)
- What the model actually sees
- Where the model slips
- 1. Who said what
- 2. Old information treated as current
- 3. The opposite: recency bias
- 4. Hitting the context limit
- 5. Attachments that weren't read
- 6. Signatures and disclaimers mistaken for content
- How to spot a slip quickly
- Habits that make threads easier to read, for AI and humans
- What to expect from good tooling
- Key takeaways
A thirty-message email thread looks like a conversation to you. To an AI model, it often looks like the same paragraphs repeated a dozen times, wrapped in signatures, legal disclaimers and "Sent from my phone" lines, with the names of who said what scattered unpredictably through the text.
Understanding what the model actually receives explains most of the odd things AI summaries and drafts do with long threads. It also points to simple habits that make threads easier for both machines and people to read.
What the model actually sees
When an AI tool summarizes or drafts a reply to a thread, it typically receives the thread as text. Depending on the tool, that might be each message separately with its sender and date, or the raw body of the latest message with all the quoted history underneath. Either way, a long thread contains several kinds of material that don't help:
- Quoted history. Most email clients include the entire previous conversation below each reply. A thread of twenty messages can contain the first message twenty times.
- Signatures. Name, title, phone, address, pronouns, a booking link, a logo's alt text, sometimes a quote.
- Legal disclaimers. "This email and any attachments are confidential..." paragraphs, sometimes longer than the message itself.
- Client artifacts. "On Tue, Sep 3, at 10:14, Sam wrote:", "From: / Sent: / To: / Subject:" blocks, forwarded-message headers, mobile footers.
- Formatting debris. HTML turned into text can leave stray characters, broken tables and repeated link URLs.
Good tools strip a lot of this before the model sees it. But stripping is imperfect, because every email client formats quotes slightly differently, and people reply inline, top-post, bottom-post and forward in ways that break neat patterns.
Where the model slips
1. Who said what
This is the most common and most damaging error. When quoted text is interleaved, or when a person replies inline inside someone else's message, attribution gets murky. The model might attribute the customer's complaint to your teammate, or summarize "we agreed to a refund" when actually the customer asked for one and nobody agreed.
Why it happens: the sender of each message is clear, but the authorship of each line inside a message isn't. Inline replies ("see my comments in red below") are especially hard, because the color that marked them is often lost when HTML becomes text.
2. Old information treated as current
A thread about pricing might mention three different numbers over two weeks as the negotiation moves. If the model weighs everything equally, the summary may quote the first offer rather than the final one.
Why it happens: repeated quoting means early messages appear many times, which can give them more weight than the single latest message where the decision was made.
3. The opposite: recency bias
Some models lean heavily on the most recent message and treat earlier context lightly. If the last message is "Sounds good, thanks!", the summary may conclude everything is settled, when there's an unanswered question three messages up.
4. Hitting the context limit
Every model has a limit on how much text it can consider at once. Very long threads, especially with heavy quoting or pasted logs, can exceed it. Tools handle this by truncating (often dropping the oldest or middle content) or by summarizing in chunks. Either way, something gets less attention, and it's not always obvious what.
5. Attachments that weren't read
A message might say "see the attached contract, section 4 is the problem." If the tool doesn't read attachments, or can only read some file types, the model is summarizing a conversation about a document it never saw. A careful tool will say so. A less careful one may guess.
6. Signatures and disclaimers mistaken for content
Occasionally a disclaimer leaks into a summary ("the sender notes this email is confidential") or a signature's tagline gets treated as a statement. Mostly harmless, but it shows the model couldn't tell boilerplate from substance.
How to spot a slip quickly
When you read an AI summary of a long thread, check the parts most likely to be wrong:
| Check | Why |
|---|---|
| Any numbers, dates or prices | Most likely to be outdated or mixed up between messages |
| Any claim that something was agreed | Attribution errors turn requests into agreements |
| Open questions | Recency bias can hide unanswered ones |
| References to attachments | The model may not have read them |
| Who is asking for what | Inline replies scramble authorship |
If the thread involves money, commitments or anything legal, read the relevant messages yourself. Summaries are for orientation, not for decisions.
Habits that make threads easier to read, for AI and humans
Most of these are just good email practice. The AI benefit is a side effect.
Start a new thread for a new topic. A thread that began as a bug report and drifted into a contract renewal will confuse any reader. Split it, and reference the old thread in one line.
Trim quoted text when replying. Many clients have a setting to quote only the selection you highlight. Fewer repeated paragraphs means clearer history.
Avoid inline replies in colored text. If you need to answer point by point, quote each point on its own line and put your answer underneath, so structure survives as plain text.
State decisions explicitly. "To confirm: we'll refund the March invoice, and the plan changes on April 1." A clear sentence like this gives summaries something unambiguous to find.
Put key facts in the body, not only in attachments. One line summarizing what's in the attached file helps anyone who reads the thread later.
Keep signatures short. Name, role, maybe one link. The thread gets lighter for everyone.
What to expect from good tooling
A well-built email AI should separate messages, label senders and dates, strip quoted history and boilerplate where it can, tell you when content was truncated or attachments weren't read, and make it easy to jump from a summary to the original message. When you're evaluating tools, test them on your messiest real thread, not a clean demo.
Key takeaways
- Long threads are full of repeated quotes, signatures and disclaimers that make AI's job harder.
- The most common errors are attribution mix-ups, outdated numbers, missed open questions and unread attachments.
- Check numbers, agreements and open questions in any AI summary before acting on it.
- Splitting topics, trimming quotes and stating decisions plainly help both AI and human readers.
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.
