Where AI assists, where humans decide.

A single diagram that has settled more executive debates than any other artifact we ship. One workflow, split into the tasks AI does and the tasks people own.

When a board asks whether we are "replacing the team with AI," I stop talking and draw this. It ends the debate faster than any deck. Here it is, cleaned up.

AI · ASSISTS HUMAN · DECIDES 01 / Draft Decompose work Tickets, options, tests. AI · ASSIST 02 / Generate Code & risk note Diff, summary, signals. AI · ASSIST 03 / Watch Telemetry & evals Flag anomalies fast. AI · ASSIST Scope go-ahead What is in, what waits. HUMAN · DECIDE Deploy go / no-go Read note, click ship. HUMAN · DECIDE rollback / re-scope loop
One workflow, two lanes — AI assists on the top row, humans decide on the bottom row, and nothing ships without crossing the line.

The line in the middle.

The dashed line is the whole point. Everything above it is a task we are happy to hand to a model — drafting, generating, watching. Everything below it is a decision, and decisions stay with people. The work crosses the line twice, and both crossings are human.

AI does the typing. Humans do the deciding. The line between them is not a phase of the project — it is the product.

Why this settles the debate.

Executives are not actually arguing about AI. They are arguing about accountability — who is responsible when something ships. The diagram answers that without a single sentence about model architecture.

  • Every assist task feeds a decide task. AI never closes a loop on its own.
  • Every decision has a named owner. The bottom row is a person, not a pipeline stage.
  • The loop is dashed and orange because it is rare. Rollbacks happen; they are not the plan.

What it does not say.

It does not say AI is unimportant. The top row is most of the keystrokes. It says the keystrokes are not the judgment, and we price and staff the engagement so a person always owns the judgment. That is workforce-first, drawn in one figure.

— Kelvin Tran. Walnut Creek, CA. June 2026.

All posts Read next: Hard hats and a Responsible AI rollout

Comments