Dor Sasson uses recent AI pricing and subscription disruptions to argue that post-hoc invoices are too late to control unpredictable workloads. He proposes synchronous entitlement and balance checks on the request path, followed by asynchronous reconciliation, treating the comparison with banking as an architectural analogy rather than a claim that AI vendors are licensed banks.
The talk explores concurrent agents drawing from a shared credit pool, hold-and-settle patterns, idempotency and auditabilityIdempotency is the property that repeating an operation has the same intended effect as performing it once.. Dor Sasson also discusses different credit sources, organizational budget hierarchiesAI agent cost governance attributes, limits and optimizes resource spending across complete AI agent runs. and the value of keeping metering data close to the workload.
The concluding design principles are to reserve before inference, settle actual use afterwardUsage-based pricing charges customers according to measured consumption of a service rather than only a fixed recurring fee. and provide visibilityAI cost observability explains which AI runs, users, models, tools and behaviors generated resource spending. into which users, teams, models and agents consume an allocation. Dor Sasson presents these as infrastructure requirements for predictable AI spending, with examples and estimates attributed to his talk.
Watch on YouTube




