An agent is not “AI in a feature.” It is execution capacity. It performs real work that either (a) wasn’t done before, (b) was done by internal teams, or (c) was outsourced — and that work has performance expectations: cost, turnaround time, measurable quality, and volume limits.
I’ve seen a good amount of “agent” projects start model-first: pick a strong model, wrap it in a chat UI, and hope useful work falls out. That can produce impressive demos, but production value comes from systems people trust — because they’re observable, auditable, and controllable.
The right starting point is not the model. It is the org chart.
From org chart to map
Take the functions that exist today, map the workflows they own, and make the work explicit: inputs, decisions, handoffs, outputs, and the metrics those workflows are already held to.
Put each workflow on one page — trigger, key decisions, system actions, handoffs, and what “good” looks like in cost, time, quality, and throughput. Then ask a concrete question: which workflows (or slices of workflows) can move under agent ownership without breaking safety, controllability, or accountability.
The clean cut
The only question that matters: where is the clean cut between “agent-owned” and “human-owned.” Not by job title, but by decision type. An agent can own the parts that are repeatable and testable. Humans keep ownership of policy changes, high-blast-radius decisions, and ambiguous exceptions.
Autonomy should ramp in levels, not as a switch flip. A practical sequence is:
- Draft only — human executes
- Pre-fill with approval — human confirms
- Execute inside hard limits — human handles exceptions
- Broaden scope — once the system proves stable
This progression forces discipline. Each step up requires evidence in metrics and clean rollback paths, not simply confidence in the underlying model.