The method I actually work from to take operational work from “where can we use AI?” to an Autonomous Operational System a business will hand real work to. It runs in two halves: discovery, to find where autonomy creates outsized value, then design, to make that autonomy earn trust. Below the method sits the deeper build, launch, and scale checklist I pull from. It is long on purpose, so cherry-pick what fits your situation rather than following it like a recipe.
Stop asking where AI fits. Start redesigning the work around autonomy, then design it to earn trust. The rule underneath all of it: delegate where the expected loss is lower than the value released, and match autonomy to how reversible a decision is.
Discovery
Start from an outcome and a constraint, never a tool.
Design
Decide, per decision, how much the agent may do.
Rollout
Autonomy is earned stage by stage, not configured up front.
When every team can call the same models and ship in a weekend, building stops being the moat. What compounds is judgment: which operation to hand over, where work actually waits, and what a customer would never forgive.
Everyone gets the same models. Not everyone gets the same judgment, built into data, workflows, and trust.
Autonomy is not a toggle you flip. It is earned with evidence: evals, logs, visible reasoning, and undo are product features, not compliance add-ons.
Your agent will be wrong. Design what happens next before launch, and treat every agent like a new employee with a badge, a role, and a budget.
Always start with customer needs, validate continuously, and iterate based on feedback. In AI products, this means understanding not just what customers want, but how AI can fundamentally change their workflows and outcomes.
Define metrics early, measure everything, let data guide your strategy. AI products require sophisticated measurement frameworks that capture both technical performance and business impact.
Ship fast, learn quickly, and iterate. Perfect is the enemy of good. In AI development, rapid iteration is crucial for model improvement and feature validation.
Break down silos, align teams, and work together toward shared goals. AI products require deep collaboration between product, engineering, data science, and business teams.
I refine this as the agentic stack changes. Drop your email and I will send the occasional update, plus new deep-dive posts. No spam.
If you want a second set of eyes on what to hand over, how much autonomy it should earn, or how to prove it is working, let us talk it through.