Teardown: the inbox that drafts itself

The discovery numbers were ordinary, and that was the problem. Around 50 customer emails a day in peak season, 10 to 15 direct messages across Shopify and Meta, spread across seven mailboxes and three messaging channels. Nothing converged anywhere except on the team — one person was clearing the backlog between 8pm and 1am.

The first lesson of inbox automation is that the pain is rarely volume. It's attention. Every message, important or not, cost someone a look. So the design goal was never "answer email faster" — it was to make sure a human only ever sees the messages that deserve a human.

The obvious move is to point a model at the inbox and let it sort. We did the opposite. A pre-filter of plain rules — blocklists, allow-lists, pattern matches the client edits themselves from a dashboard — drops newsletters, noreply notifications and system mail before the AI sees anything. On 16 June it removed 81 messages in a single day, each one logged with its reason. Nothing disappears silently.

That decision made everything downstream cheaper and sharper. The triage model only ever reasons over messages a person would actually care about, and the client can tune the noise floor without touching a prompt or calling us.

Triage classifies every surviving message FLAG or DROP under a rule the founder already carried in her head — an order or booking notification with a customer question inside is a flag; the same notification with no question is a drop. Read the body, not the subject line. Read the whole thread, not the latest message.

That last clause was learned the hard way. An early version read only the newest message in a thread and politely re-asked questions the customer had already answered. The fix wasn't a smarter model — it was a sharper rule. This is the pattern that held up across the whole build: the judgement belongs to the founder, written down as rules; the model's job is enforcing them consistently at a scale one person can't.

Drafting replies is where AI genuinely earns its place — a decent first draft, in the brand's voice, is exactly what a model is good at. But we also found its edges fast. It guessed at product pricing on quotes. It once quoted a school discount that didn't exist. The response wasn't better prompting: pricing was removed from every AI reply, and drafts now link to the website instead of quoting numbers.

Anything touching money, legal or safety never reaches the drafter at all. Chargebacks, legal threats, tax-office correspondence, refunds over $100 — a deterministic keyword matrix catches them and pings a named human on WhatsApp immediately. Where being wrong is expensive, rules win. Where being slow is the enemy, the model wins. The skill is knowing which messages are which.

The deliverable was never an empty inbox. It's a drafted reply waiting in each mailbox's own Drafts folder, prefixed [AI DRAFT], never sent. The team reviews, edits what needs editing and presses send themselves. The first live week was deliberately small — a little over 50 messages processed, about ten drafts — because trust is the actual feature, and trust is earned draft by draft.

There's a graduation path, and it's strict on purpose. A category of reply earns auto-send only after shadow mode shows at least 95% of drafts going out unedited for two straight weeks, with sign-off, a capped canary and a per-category kill switch. Some categories never graduate at all: money, legal, dietary-critical, complaints. Auto-send is something the system has to earn, one category at a time — and as of this writing, it hasn't. Everything still waits for a person.

The honest list is long. The model sometimes answered in prose instead of structured data, which silently broke downstream steps — now it returns strict JSON, and a malformed reply surfaces as zero confidence instead of slipping through. A spam subject containing the right keyword got flagged urgent at 99% confidence. Enquiries forwarded between internal mailboxes were dropped as noise until the rules learned what a forward looks like. An escalation rule treated any mention of compensation as a legal threat until real support tickets taught it otherwise.

The founder co-trains the system. She returned a corrected 15-question FAQ set that shipped the same day, sends video feedback on drafts, and her staff forward automatable emails as examples. An inbox system isn't a launch, it's a weekly tuning loop — the workflows that survive are the ones built to be edited.

Strip away the stack and the guarantees are plain. Nothing sends itself. Nothing disappears silently — every filtered message is logged with its reason. Anything touching money, legal or safety reaches a human fast, not eventually. And a ranked morning digest tells the founder what arrived, what was drafted and what still needs her.

That's the real shape of inbox automation done honestly. Not an AI running the inbox — a system that absorbs the attention overhead so the people keeping final approval can spend it where it counts.