Everyone runs the same models.
Yours knows your business.

A model out of the box is a stranger who writes well. What makes an agent worth having is everything it knows that nobody else's does — your policies, your wording, the exceptions you would never write down for a stranger. The more of that it has, the more of the work it takes off your desk, and correcting it is how it gets there.

From your sent mail to sending on its own.
Four phases, in order.

Four phases, in the order they happen. No phase is skipped, and none of them needs you to be technical.

  1. Phase 01 · Before day one

    It reads the mail you already sent. Before anything goes live, the agent reads an export of your own mailbox and writes what it would have replied to each message. Held against what you actually sent back then, that gives two things at once: the examples and the facts are already in place on day one, and you get a number — in how many of those cases the draft matched your own answer.

    That number is the test we hold ourselves to before a category is handed over at all. It says the agent knows the right answer; whether you trust it to send is the next thing, and that is earned separately.

  2. Phase 02 · Learning period

    It writes, you send — and every correction sticks. A new category begins as drafts: the agent writes, you send. That is where it starts; it is not where it stays. While it lasts, sending is not on the agent's list at all — it is handed its actions by name, and for mail those are two: write a draft, put a label on the thread.

    The draft sits in your own screen, in the thread it belongs to, with the order details it looked up already in it. Approving takes a second. Correcting is the valuable one: what you change, and why, gets written down. Each correction is one thing it will not ask you again.

  3. Phase 03 · The count

    One number decides, and you can see it move. The number is already there: how many of a category's drafts you approved without changing a word. It stands on your own dashboard, per category, next to what it was last week.

    No gut feeling, no "trust us" — a count you watched build up, on mail that was really yours.

  4. Phase 04 · On its own

    The category will hand itself over. That is where the first three phases are heading. Once the count is there and you switch the category on, that mail stops arriving at your desk. Switching a category over will be yours, and switching it back will be one action on the same screen, with no call needed.

    What still reaches you, then as now: the cases the agent is unsure about, and anything you named as yours. Those keep coming, on purpose.

This is the screen you work in.
Not your mailbox with a plugin.

Not your mailbox with a plugin bolted on — your own helpdesk. The customer's mail on the left, the agent's draft underneath it, the order and history on the right. One screen, everything in it.

The real helpdesk, shown with an example shop. From a new mail to a sent reply, and how you steer it: edit a line, give it a rule, decide what it may send on its own.
The draft is already writtenOrder looked up, tracking checked, policy applied — before you open the thread.
You approve or you correctApproving takes a second. A correction is remembered, so it does not come back.
The count builds itselfEvery approval without an edit moves that category closer to running on its own.

What stops reaching you.
The ordinary mail, mostly.

The aim is not a second screen to check every morning. It is that the ordinary mail stops arriving at your desk at all, and what is left is the part where somebody actually has to think.

How much that is depends on two things. How much of your business goes into the agent — your policies, your past answers, the exceptions you know by heart and have never written down. And how many categories it has earned, because that happens one category at a time.

Everything it learns goes into one file.
On your machine, plain text.

Your corrections land in brand.json — on your machine, in plain text, and you can open it.

brands/your-shop/brand.json
{
  "shipping_windows": { "NL": "1-2 working days", "DE": "3-5 working days" },
  "returns":          { "window_days": 14, "condition": "unworn, tag attached" },
  "tone":             "warm, direct, no corporate padding",
  "always_escalate":  [ "complaint", "press", "legal", "refund > €150" ]
}{
  "warranty":          { "standard_months": 24, "proof": "order number is enough" },
  "returns":           { "window_days": 30, "condition": "original packaging" },
  "tone":              "precise, technical when asked, never condescending",
  "always_escalate":   [ "safety issue", "press", "legal", "replacement > €400" ]
}{
  "shipping_windows": { "NL": "1 working day", "BE": "2 working days" },
  "allergy_flags":     [ "fragrance", "retinol", "nut oils" ],
  "tone":              "warm, reassuring, never medical advice",
  "always_escalate":   [ "reaction reported", "pregnancy question", "press", "legal" ]
}

And beyond the inbox: custom builds.
Whatever else eats your hours.

Customer service is the service. But the mechanism above — an agent does the work, it lands where you already look, your reaction makes the next run better — is not limited to mail. If something else eats your hours, we build the system that takes it over, scoped on the call with its own figure agreed up front.

Reporting

The numbers you now collect by hand every Monday, standing ready before you ask.

Orders & stock

The checks and chasing between your shop, your carrier and your supplier.

Whatever it is

Bring it to the call. If it can be automated well, we scope it. If not, we say so.

One rule holds for everything custom: nothing goes to a paying customer before it has run for real. We would rather turn work down than sell a promise.

Every week you wait,the competition pulls further ahead.

Seth Desaunois Book your audit call30 minutes

No sales pitch. The person who builds it, running the numbers on your operation.