Inbox Triage
An agent that reads the inbox first, so the founder doesn't have to.
I was spending the first ninety minutes of every day just reading and sorting email before I could do a single thing that actually moved the company forward.Founder, YC-backed B2B SaaS
A confidence gate sits between Draft and Send — nothing leaves without a human glance.
We started by mapping the real inbox, not an idealised one. A week of shadowing showed that the founder’s mail fell into five recurring shapes — investor threads, customer escalations, sales replies, internal noise, and the long tail of things that only a human should touch. Those five became the priority buckets the agent sorts against.
The classifier was built against a month of the founder’s actual email, labelled by hand so the model learned this inbox rather than email in general. We tuned it until its five-way call matched the founder’s own sorting closely enough to trust, then held the routine categories back for drafting and pushed everything ambiguous straight to the top of the queue.
For the replies that repeat — scheduling, intros, “here’s the deck,” polite declines — the agent writes a first draft in the founder’s voice and stages it. The review surface is deliberately plain: a single morning digest where each item is one glance to approve, edit, or skip. No new app, no dashboard to learn.
We shipped it with the off-switch built in from day one. The agent never sends on its own, drafts sit until a human clears them, and a single toggle pauses drafting entirely if the week goes sideways. That restraint is what turned a clever demo into something the founder leaves running.
Whether to ever let it send on its own. We decided it never auto-sends — every draft waits for one human glance. Slower on paper, but it's the reason the founder actually trusts it.
On retainer we watch for classifier drift as the business changes, retune the priority buckets each month, and add new reply templates as new patterns emerge in the inbox.