Built against a real mailbox, not a demo.

What actually hits a support inbox

Most of it was never a customer.

A deterministic census bucketed every email by rule; the surviving candidates were each classified once by claude-haiku-4-5. Only 8.3% of everything that arrived was a real customer writing in.

Why customers actually write in

For apparel, sizing beats shipping.

Across the 2,344 real customer emails, the top three reasons - sizing exchanges, where-is-my-order, and quality claims - are 64.8% of everything. This is the work Podley was tuned for.

11.2% of real customer emails read anxious and under 1% read angry. There were 2 chargeback threats in 20.5 months - rare, but catastrophic when missed, which is the case for escalation guardrails.

The safety gate

The model proposes. Code disposes.

Everything Podley proposes passes a deterministic gate before anything happens. Only 33 specific support actions can run; anything else is refused by name. Money never moves without your approval by default, and hard dollar caps stand even after you grant autonomy. A separate check blocks any reply that claims an action is done when it was actually held for your approval. And reaching your plan limit drops Podley to draft-only - never a surprise bill.

The gauntlet

We hired an adversary.

Before merchant #1, we ran 128 scenarios at the engine. 27 of them were built purely to break it - refund tricks, cap-evasion, sympathy pleas engineered to talk it into overpaying, prompt injection. The rest was ordinary support mail, so we would catch a regression too. Every case that must never move money on its own was held for a human. Not most. Every one.