Article notes
The Agentic AI Product Gap
When an AI system can take action, a good answer is only part of the problem. This essay looks at the product decisions around its permissions, mistakes, recovery, and value.
Read full essay on MediumThe argument
An action has consequences beyond the response.
A system that updates a record, issues a refund, or moves work between teams needs a clearly defined relationship with the people and systems around it. The product question includes whether it took the right action, stayed within its authority, and left a way to recover when something went wrong.
The Agentic Product Stack is a way to examine those decisions together. It connects the customer workflow to the rules for autonomy, evaluation, monitoring, and business value.
The framework
Ten layers, each with a question to answer.
Workflow Fit
Which part of the user’s work would improve, and where does the system enter that workflow?
Risk Shape
What happens when the system is wrong, and who bears the cost?
Action Surface
Which tools, records, and external systems can it read or change?
Autonomy Boundaries
What can it do alone, what needs approval, and what must it never do?
Evaluation Gates
What evidence would justify a release or a change in its permissions?
Human Control Design
Who can inspect, correct, stop, or take over its work?
Runtime Health Signals
Which signals reveal deteriorating quality, unexpected cost, or repeated human corrections?
Damage Control and Rollback
How can an incorrect action be contained, reversed, or escalated?
Governance Evidence
What records make its actions and the responsible owner traceable?
Business Value Loop
What changes for the customer and the business, and how will we know?
Using the framework
Start with one workflow and a clear boundary.
The layers are a way to expose open decisions. A proposed action might be useful but difficult to reverse. A good evaluation result might still leave uncertainty about when a person should approve it. Making those connections visible is the purpose of the framework.
The companion mapper turns these questions into a draft brief. Its suggested answers are starting points for discussion and validation, rather than evidence that a product is ready.