
1. Capture delivery
Collect each output at the point of handoff and attach the reviewer’s decision.
The AI Workforce method
Define the work, put a person in charge of approval and measure what the team actually uses.

Our delivery KPI
How much work people adopted for the token budget spent.
Illustrative target
45 adopted deliverables ÷ 90 quota points. Calibrate the target to the job and its starting point.

Collect each output at the point of handoff and attach the reviewer’s decision.
Mark used, revise or park. Count each adopted delivery once.
Compare accepted work with quota spend. Check rework and critical runs before changing a role’s scope.
Start with one recurring task, one standard of done and one person who can approve the output.
Map recurring work and identify where it gets stuck.
AI Opportunity MapDefine each role, its manager and acceptance criteria.
AI Team BlueprintUse test data and grant only the access the role needs.
Connected WorkflowReview the outputs and track when people step in.
Pilot Performance ReportTrain managers and document how to review and escalate work.
AI Workforce Operating Playbook
Define the role, onboarding, standard of done, boundaries and review.

Observe the role, review drafts, require approval and expand within written limits when the work supports it.
Start with a task your team already does often and can check quickly, such as preparing a meeting brief or reviewing product listings. Bring one example of good work and one that needs revision. Give the role those examples, a clear deadline and a named reviewer. Run the pilot on a narrow set of tasks before adding more work.
Use five fields: the role, its onboarding material, the standard of done, its boundaries and the review process. Write down what the role receives, what it hands back, and who checks it. Add a short example of work outside its remit so it knows where to hand off instead of guessing.
Give it only the information and permissions its current task needs. A research role can work with approved sources and sample documents before it sees any private files. Start with read-only access where possible and record who can authorise changes. When an output involves sensitive information, stop and ask the named reviewer before sharing it.
Begin in Shadow, where the agent observes and explains what it would do. Then move to Draft, Ask first and eventually Within limits if real work supports each change. Give it more responsibility in small steps. If quality slips or the task changes, reduce access and review again. Consequential external actions remain with the human approver.
Return on Token looks at the work a human actually used and the token budget spent to produce it. Think of two roles that use the same budget: one delivers work people adopt, while the other delivers work that sits unread. That difference is what the KPI makes visible. A busy role can still have weak ROT.
Give each deliverable an identifier and ask the reviewer to record one outcome: used, revise or park. Used counts toward adoption once. Revise signals that more work is needed, and park means the output has not been adopted. Keep the same review period for spend and adoption, and leave the KPI pending when either side is missing.
Record the missed or poor output, route it to the named human and pause the affected routine if continuing could cause errors. Check the role instructions, source material and timing before changing permissions. Run counts show whether work fired; timestamps are needed to establish punctuality. The reviewer can re-scope, test and restore the routine when ready.
Start with the tools your team already uses and test the workflow with sample data. The access model, integration limits and review effort determine what can be connected and how long a pilot should last. Agree on one deliverable and a baseline first. Expand when the team is adopting the work and the review process is dependable.
Pick the first task you want to hand off. Little Liang will help you build the team around it.
