The principal of a small architecture practice spent one to two hours every day doing one thing: reading each email, keeping the ones that carried a task, a decision or an open question, and saving each as a message file into the right job folder, so the paper trail lived on disk, searchable, immune to the mailbox’s size limit. He had done it for years and he was good at it.
He was also, by his own account, sometimes wrong, and the emails he got wrong had a shape. A project manager juggling four jobs writes “Re: 2419 Oak Street” in the subject and asks a question about the Pine Avenue beam in the body. A structural engineer replies to a thread about one job with a value-engineering option for another. The subject is a lie the sender did not intend to tell.
An agent that files email will meet this email on its first day. What it does next decides whether it is trusted.
Evidence that votes
The filer identifies the job by evidence, and the evidence votes. An earlier email in the same thread is a vote. A job number in the body is a stronger vote than one in the subject. Words from a job’s name in the body are a vote. Where this sender’s email was filed before, learned from the folders already on disk, is a vote. The model’s read of the whole message is a vote.
Most email is unanimous, and the filer files it. The interesting case is the split decision: the subject says Oak Street and the body says Pine Avenue. That is a conflict. So is a tie, a low-confidence read, or a job that has no folder yet.
Ask
A conflict is not resolved by picking the stronger vote. It becomes a quick check. The principal sees the email, the candidate jobs, the reason for the doubt and the exact folder each choice would land in, picks one, corrects the party or topic if needed, and confirms. Nothing is written until he does.
The queue is short, because most email is unambiguous. It also contains, almost exactly, the emails he would have got wrong on a busy afternoon, which is why he described the tool as asking him about the right ones.
And every confirmation teaches. The sender-to-job map updates. Confirming an email from someone new adds them to the list of tracked senders. The next email from that contractor on that job files itself.
The folder is the paper trail
Inside each job, the filer keeps two sides, received and sent, and within each side it files by issue. An issue folder is named by date, party and topic: a date, then a short code for who (client, general contractor, structural engineer, mechanical, the authority having jurisdiction), then the topic. An email about the same party and the same topic joins the existing folder and bumps its date, so the folder sorts to where the conversation currently is. An email about a different topic opens a new one. “RFI 011” and “RFI 014” are different issues even though they share a word, and the rules say so.
Messages are saved numbered inside the folder, as native message files written through the mail client itself. Nothing in the mailbox is moved, flagged or deleted. The mailbox is the inbox; the disk is the record.
Filter first
Before any of this, the filer decides whether an email deserves filing at all. Only tracked senders are considered: clients, consultants, contractors, plan checkers. Among their email, keyword rules and the model look for an ask, a deadline, a request for information or proposal, a submittal, a question. Courtesy replies, FYIs, vendors and newsletters stay in the inbox. An email the principal has highlighted is always filed. A stranger with a strong ask, say a request for proposal, goes to the quick check rather than being ignored.
Why this shape
The temptation with a filing agent is to make it decisive, because decisiveness looks like intelligence. But a paper trail is the one place where a confident wrong answer is worse than no answer: a misfiled email is found a year later, by someone looking for it in the wrong folder, often with a lawyer. The right design is an agent that files the obvious, asks about the ambiguous, shows its reasons for both, and learns from every answer. A test suite that runs without a key, without the mail client and without a network keeps it that way.
About the authors. Neugent Consulting designs and builds AI systems for operators. The examples in this piece come from client work; clients are anonymised and details are simplified.
