Delegation Lab · Orahovo Furniture Group · simulated
Same team, same problem. Redesign how the work gets done.
Every week, Orahovo Furniture Group (a fictional company) produces a performance review: what moved against budget, why, and what to do. Four people and three entity managers make it by hand today. In four screens you decide who does each activity, how an agent's task is specified, how one exception is handled and where a small improvement budget goes. Five to eight minutes.
1 · Baseline: how the review is made today
Six activities, four lanes. The analyst collects approved inputs from the entity managers, calculates variances against the 2026-08 budget, investigates anything unusual and drafts the explanation; the manager approves it; the controller recommends a response; the manager assigns actions back to the entities.
Most of the week goes to waiting and rework rather than the work itself, and only a measured baseline makes that visible.
2 · Who does what
For each activity choose who does it. The consequence line under every option is rendered from the model's parameters for that activity, not written by hand.
Who does an activity decides its speed, its cost and how many claims nobody checked, and the four modes trade these off differently.
3 · Specify the agent task
The agent gets a written task. Each choice maps to a modifier in the model (listed under Assumptions on the last screen).
An agent's result is only as good as the task it was given, and the specification is where the accountability lives.
Optional: let the agent draft the explanation with your task spec
One bounded call to the same explain endpoint the showcase uses: the numbers are computed in code, the model only writes about them, and every number in its draft is checked against the fact pack. Capped per day and cached per question.
Draft for Model output — checked, not simulated
Evidence handed to the model
4 · Change the organisation
You have 10 improvement points. Each lever moves exactly one constraint in the model; the constraint and its value per point are shown on the lever.
Better tools fix tool problems, while ownership, capacity and shared definitions are organisational and decide most of the outcome.
Assumptions: every parameter and the formula of each metric
What this asks of each role
Pick a role. The paragraph is keyed on the choices you actually made in this run; it is written text, not generated.
Further reading. Two references shaped this lab and are cited only here: Anthropic, "Building effective agents" (2024), and Microsoft, Work Trend Index 2026. The simulation model, its parameters and the company are the author's own; all figures are simulated.