KotaML Logo
KotaML
Framework/Execution Controls
4

Economic Control Is Decided in the Execution Layer

AI systems have many levers that affect cost, quality, and performance. First-order controls determine the overall margin shape, while second-order mechanisms fine-tune behaviors within those bounds.

Execution Control Plane

USER REQUEST(UNBOUNDED INTENT)FIRST-ORDER LEVERSBudgetingHow much can I spendper request?RoutingWhat model shouldI use?EscalationAuthorityCan the system domore work?TerminationWhen does executionstop?BOUNDED OUTPUTREQUEST KILLEDPRICINGMCP / PROMPTNORMALIZATIONEVALUATIONFALLBACKS /PROTOCOLSSECOND-ORDER INPUTS
  • Budgeting sets a ceiling on how much work any single request is allowed to trigger, so one prompt cannot result in runaway costs.
    Pricing sets the denominator for gross margin, while budgeting sets the numerator by backsolving a maximum allowable cost per request.
  • Routing determines which execution path a request follows within the budget, including model choice, selecting the cheapest path that can effectively respond to the request.
    Classification structures and refines inputs to standardize requests while improving quality, enabling routing to select the right path. Without it, routing is guessing.
  • Escalation determines whether the system is allowed to execute again after an initial attempt, including limits on retries, model upgrades, and incremental budget.
    Evaluation measures whether an execution meets confidence, coverage, compliance, or policy requirements, producing signals that inform escalation decisions.
  • Termination defines when execution must stop, typically when budgets, time limits, or retry limits are exhausted.
    Protocols define how the system behaves once termination is triggered, including fallbacks, partial responses, and cleanup.

Execution controls are the single most important determinant of long-term margins in AI systems.