Governed Agent Operations
What the bot wrote: "You're booked for Tuesday at 10am."
What happened: Composed by the bot. Never sent.
AI automations narrate success. Production verifies it. Your bots say "booked," "sent," "reported" — and nothing downstream checks whether the world agrees. I audit automation stacks for the gap between what they claim and what they recorded, and I hand you the evidence.
From a public audit of five popular automation templates — all in the top 200 of 11,788 by views. Every claim traces to a committed record in the repo:
books the appointment, tells no one
The booking step returns ACCEPTED; the next node throws before the confirmation sends; the customer hears nothing. The canvas looks intact — the error sits in a log nobody reads.
the knowledge base keeps the policy you replaced
Edits append instead of replace. Both the old and new version come back, and whether the bot tells you depends on the model, not the system.
the percentages in your client report are the model's arithmetic
Inputs computed deterministically, the change column handed to an LLM. Six of eight cells moved across identical runs. The drift was small; the check was absent.
one bad row, and the campaign emails no one — hourly, forever
One row missing a column entirely halts the whole batch. The sheet reads zero sent, truthfully, and nobody checks it.
Three steps. Each one earns the next, and the first shows you the findings before you pay.
No client sweeps have run yet — this is new, and I'd rather you know that than find out. What exists is the method, the public audit, and the checklist, all open for you to check before we talk.
I also don't audit stacks with no automation running. If nothing's deployed, there's nothing to find — come back when there is.