An AI agent that sorts your email: anatomy of a real case

When we listed the AI agent use cases actually deployed in SMBs, sorting incoming email came first, and it is almost always the first custom AI agent a company puts into production. The reason is simple: in a business of 10 to 50 employees, the shared inbox is the company's real operating system. Everything lands there: quote requests, claims, invoices, follow-ups, solicitations. And someone, every morning, pays in hours to put that pile in order.

Still, "AI sorts your email" is an abstract sentence, and abstraction is a poor advisor when a contract is on the table. This article opens the hood: the six organs of a triage agent, where the difficulty actually hides (hint: not the AI), where errors happen and what they cost, and the accuracy you are entitled to demand before go-live. No fictional client; stated assumptions and the arithmetic that follows.

The short answer, for the busy

An AI email triage agent connects to your existing inbox (Gmail or Microsoft 365, through their official APIs), reads each incoming message, classifies it against a taxonomy specific to your business, extracts the useful data (file number, amount, deadline), then routes it to the right person according to explicit business rules. Ambiguous cases are not guessed: they are escalated to a human, and every correction improves the agent. Inference cost is marginal: with Claude Haiku 4.5 at $1 US per million input tokens and $5 per million output tokens, a few thousand emails a month cost under $15. The real investment is designing the taxonomy and the routing rules. In Quebec, such an agent on subscription starts around $947 per month and pays for itself once manual triage exceeds roughly an hour and a quarter a day. Accuracy is measured on a labeled sample before go-live, never taken on promise.

Six organs, one circuit

A triage agent is not a glorified spam filter. It is a six-stage circuit, and each stage is designed separately.

Intake. The agent connects to the existing inbox through the platforms' official interfaces: the Gmail API can read messages and modify their labels; the Microsoft Graph API can move a message between folders and assign categories. Nobody switches email software: the agent works inside the tool your team already uses.

Classification. The model reads the message and assigns it one of the categories you defined. This is the "intelligent" step, and paradoxically the least problematic in 2026: current models classify text remarkably well when the categories are clear.

Extraction. For each category, the agent pulls out structured fields: file number, client name, amount at stake, deadline. This is what separates "classifying" from "preparing the work": the colleague who receives the message also receives its summary and key data.

Routing. Explicit business rules decide the destination: this category goes to this person, this combination (claim plus large amount) goes up to management. These rules live outside the model, in readable code, for a precise reason: you can audit them and change them on a Tuesday afternoon without touching the AI.

Escalation. When the agent is unsure, it does not guess: it places the message in a human review queue, with its best hypothesis and the reason for its doubt.

The feedback loop. Every human correction (a reclassified message, a changed route) is used to adjust the agent's instructions. A triage agent is never "finished": it converges. We have described elsewhere how an AI agent gets built, without code; email triage is the purest illustration of that mechanic.

The hard part is not the AI, it's your taxonomy

Here is what surprises every client: the longest phase of a triage project is the workshop where the company defines its categories. Not the technical integration. The workshop.

A good taxonomy has three properties. Categories are mutually exclusive: a given email belongs to exactly one. They are actionable: each category maps to a different decision (different recipient, different response time, different reply template). If there is no distinct decision at the end, the category does not deserve to exist. And the taxonomy includes a relief valve: an explicit "other / to review" category, because a taxonomy with no escape hatch forces the agent to guess.

In practice, the exercise reveals something else: the company discovers its own employees do not sort the same way. Person A files renewal requests under sales, person B under customer service. That disagreement existed before the agent; the agent merely makes it visible and forces a decision. The taxonomy that comes out of this workshop is a specification of your operations, and it has value even if the project stopped there.

Where errors happen, and what they cost

A triage agent makes mistakes. A vendor who claims otherwise is wasting your time. The useful question is twofold: where, and at what cost.

Errors cluster around known profiles. The email that contains two subjects at once (a billing question and a complaint). The implicit urgency, never named, that only context betrays. The chain forwarded three times with the real subject buried at the bottom. The attachment that carries the substance while the body says "see attached". And absolute novelty: the request type never seen before, born of a service you just launched.

The cost of an error is asymmetric. A solicitation filed under "suppliers" costs the colleague thirty seconds to re-forward. An urgent claim buried in the wrong queue can cost a client. The design must reflect that asymmetry: on critical categories, the agent is tuned to prefer escalating wrongly over classifying wrongly, even if that means more human review. Missing an urgency is forbidden; bothering a human for nothing is tolerated.

And the honest comparison point must be named: human triage already makes errors. The shared inbox where two people grab messages depending on their workload has a real error rate; it is simply measured by no one. The agent logs every decision. Its error rate is at least known, category by category, and that is the precondition for improving it. It is the same guardrail logic as for an AI agent in customer service: the value comes as much from the guardrails as from the model.

Accuracy is measured before go-live, not after

How do you know the agent is ready? Not by watching it run for a week "to see". The serious method has three steps.

First, the golden sample: a few hundred real emails from recent months, hand-labeled by your team using the agreed taxonomy. Tedious and irreplaceable: this sample is the operational definition of "correctly sorted" at your company.

Second, per-category measurement. A global score ("94% accuracy!") hides what matters. What you want to see: performance on each category, especially the critical ones, and the escalation rate (what share of messages the agent will send to human review). An agent that classifies 85% of messages with high reliability and escalates the rest is worth more than one that classifies 100% of messages with diffuse doubt everywhere.

Third, shadow mode: before handing the agent the controls, it runs for a few weeks in observation. It classifies, humans keep sorting, you compare. On switchover day, nobody discovers anything.

If a vendor does not offer this process unprompted, demand it. It turns "trust us" into "here are the numbers, on your own email".

The worked case, with numbers

Let's state the assumptions openly. A service SMB receives 150 emails per business day, about 3,300 a month. Triage takes the equivalent of two hours a day, 44 hours a month, at a fully loaded hourly cost of $35: roughly $1,540 a month of manual sorting.

Now the agent. Inference first: at about 2,000 tokens per email (message, context, instructions), 3,300 emails represent a few million tokens a month. With Claude Haiku 4.5 at $1 US per million input tokens and $5 per million output (Anthropic's official rates, checked July 2026), the model bill stays under $15 a month. Anthropic's own documentation gives the same order of magnitude: about $37 US to process 10,000 support tickets. The cost that matters is the agent subscription itself, design, integration, monitoring, and evolution included: with us, from $947 per month.

Suppose the agent classifies 90% of messages without intervention and escalates 10%, about 330 messages a month that still need human eyes: say 8 hours of review, $280. The agent total ($947 + $15 + $280) comes to about $1,240 against $1,540: a net saving of roughly $300 a month. Honestly? That number is not what justifies the project. What justifies it is what money does not measure directly: the urgent message that surfaces in thirty seconds instead of three hours, triage that keeps running through vacations and sick days, and 36 hours a month returned to work that requires a brain.

The flip side of that honesty: at 40 emails a day, the arithmetic does not hold. Keep sorting by hand, and come back when the volume starts to grind.

Where to start at your company

A free and revealing exercise: take the last 200 emails from your shared inbox and try, as a team, to sort them into ten categories. Where you hesitate, the agent will hesitate. Where you decide without thinking, the agent will excel. That one afternoon will tell you whether the case exists at your company.

And if it does, the next conversation is about your taxonomy, your volumes, and your routing rules, not about accuracy promises. We do it in 30 minutes, no commitment.

→ First conversation, no commitment