The simplest mental model for agentic workflows is 10/80/10. Humans handle the first ten percent: defining the objective, setting constraints, providing context. Agents handle the middle eighty percent: the execution, the iteration, the pattern matching, the throughput. Humans handle the final ten percent: reviewing output, making judgment calls, approving or redirecting.
This is not a theoretical framework. It is a practical operating model that we apply to every workflow we redesign. The ratio will shift depending on the domain and the stakes involved, but the structure remains consistent. Human intent at the front. Agent execution in the middle. Human judgment at the end.
Why the Bookends Matter
The first ten percent is where most implementations fail. Teams deploy agents without clearly defining what good output looks like, what constraints apply, or what context the agent needs to operate effectively. The result is technically impressive but operationally useless output that requires more human effort to fix than it would have taken to do manually.
The final ten percent is where trust is built. When a human reviews agent output and confirms it meets the standard, that review cycle creates a feedback loop. Over time, the review becomes faster because the agent's output improves. But you cannot skip the review stage. Trust in agentic systems is earned through repeated verification, not assumed through vendor promises.
Applying the Model
Take any workflow in your organization. Map the steps. Identify which steps are setup (defining objectives and constraints), which are execution (repeatable, pattern-based work), and which are review (judgment, approval, quality assurance). That mapping tells you exactly where agents should operate and where humans should remain. The eighty percent in the middle is your automation surface. The bookends are your quality guarantee.