Studio notes — visibility before you delegate
This week the agent fleet built, for a courier company, an operations cockpit for the mailbox and a monitoring agent that watches, escalates and proposes — but never acts alone. Two builds, one lesson: get visibility before you delegate any action.
Every week, the same fleet of agents is at work across several clients at once. This week, the bulk of the work concentrated at a courier company, on a single project seen from two angles: giving the management team an overview of the operational mailbox, then having that mailbox watched continuously by an agent. Two builds, one lesson that runs through the whole studio: you only hand an action to an agent once you have a way to measure it first.
A cockpit before an algorithm
The operational mailbox was receiving hundreds of messages a day, with nobody holding a real overview of what was actually happening in it: how many messages stay unanswered, for how long, which ones are urgent. Rather than bolting on automation right away, the studio first built a management dashboard — median response time, backlog aged by severity, workload per person, a queue of AI actions awaiting approval — fully recomputed on every page load, with no AI call involved. Lesson for a business owner: before automating a channel, first give yourself the means to measure it continuously; a management dashboard is not a reporting gadget, it is the prerequisite that makes every automation that follows measurable — and therefore correctable.
A monitoring agent that watches and escalates, but never acts alone
On the same mailbox, a second agent plays the role of a live monitor: it watches every incoming message around the clock, detects SLA breaches, automatically links a message to an existing case, drafts a reply from the conversation history, and runs a step-by-step escalation ladder if nobody answers — up to notifying a manager. It also catches silent failures: a mailbox that stopped syncing, a client reply that fell into spam by mistake. None of it fires without validation: every sensitive action lands in an approval queue a human clears or rejects in one click. Lesson for a CIO: an AI agent in production is judged by what it does when it should not act alone — propose, log, wait for the green light, rather than act on a guess.
That is the thread connecting both builds this week: visibility and human oversight are not steps bolted on after the agent ships — they are what makes automation possible without giving up control over it. A cockpit that measures, a monitor that proposes, a human who approves: that order is what keeps an AI agent in production a tool you control, not a black box you put up with.
Free resource
The self-assessment grid: 20 tasks AI can automate
Sales, admin, support, operations: the 20 tasks AI agents already handle in SMEs — with, for each one, the tell-tale sign that your team is concerned.