Clients of a coaching practice each expect their session recording. Manual delivery had quietly fallen sixty-six emails behind. The fix wasn't more effort — it was an agentic pipeline with a human at the one gate that matters.
These aren't meeting notes. They're recordings of people doing vulnerable inner work — and a recording sent to the wrong person isn't a typo. It's a privacy breach.
Sessions landed across two separate video-call accounts and a phone line with no API at all. Recordings were named inconsistently — sometimes a first name, sometimes a room name, sometimes a caller-ID that was simply wrong. Each one had to be matched to the right client, its link checked, and a personal email written and sent. Done by hand, it consumed hours a week — and the backlog had grown to sixty-six clients waiting.
The obvious automation — match and blast — was never on the table. The whole design question was: how much can the machine do while a human still approves every single send?
A daily pipeline that harvests, matches, validates, and drafts — then stops, and asks.
Review-first isn't a default that shipped with the software. It's a ruling — and the validation rules under it each exist because reality demonstrated the failure once.
Auto-send stays off until trust is earned. The pipeline proposes; the owner disposes — in plain language, from one daily review email.
A truncated link once reached a client. Now links are extracted only from full source content, checked for access barriers that would lock a client out, and paired with per-source download instructions.
Caller-ID names can simply be wrong. Matching runs against calendar records with a maintained alias-correction table, and anything ambiguous is surfaced for a human ruling instead of guessed.
A success response proves the call succeeded — not that it did what you meant. Sent mail is confirmed by checking what actually left, not what was requested.
The daily review job simply stopped producing its email — the quietest kind of failure, the kind that ruins automated systems. Because every send is human-gated, silence produced nothing wrong: no misdirected recording, no stale link, no client contacted in error. The gap was detected, the nine days were swept in a single validated batch — every link checked, every identity confirmed, approval given in one message — and clients were caught up. The failure was absorbed, not amplified.
That's the test that matters for agentic systems. Anyone can demo the happy path. The design question is what happens when a component fails silently on a Tuesday — and the answer here was: nothing a client would ever notice.
This is the architecture every agentic-AI company is trying to figure out: autonomy where it's safe, human judgment exactly where stakes are real.
Bring me the messy problem