The price of tokens is rarely the deciding number. What matters is what an agent’s mistakes and escalations cost.

Implement costPerTask(t):

modelCost + escalationRate × humanCost + errorRate × errorCost

and decide(t), where t also has volume (tasks per month) and buildCost (fixed, one-off):

savingPerTask = humanCost − costPerTask(t)
monthlySaving = savingPerTask × volume
paybackMonths = buildCost / monthlySaving

Return { automate, costPerTask, monthlySaving, paybackMonths }. automate is true if the build pays for itself within 12 months. With no saving, automate is false and paybackMonths is null.

Exercises in this module

Terminal