From Tool Success to Business Truth: Contract-Based Verification of Agentic Outcomes

AI agents increasingly transform enterprise systems through tool calls, workflow transitions, approvals, and database updates. Yet a successful tool response does not establish that an intended business result occurred, persisted, remained authorized, or was legitimately declared complete. This paper develops a contract-based account of agentic outcome verification that separates invocation success, observed effects, evidence sufficiency, authorization, workflow composition, and final business claims. It defines an Outcome Verification Contract (OVC) that binds an explicitly scoped goal predicate to a state oracle, time horizon, provenance requirements, authority conditions, and adjudication rule. A three-valued completion judgment prevents missing observations from silently becoming positive evidence. Conditional propositions show why successful component operations cannot entail global completion without additional assumptions; why certain outcomes are observationally indistinguishable to an agent with restricted access; and why truthful completion is time-, scope-, and authority-relative. Six constructed enterprise scenarios expose contract violations and define discriminating witnesses. A preregisterable evaluation protocol compares tool-status heuristics, conventional workflow verification, and contract-explicit adjudication under matched evidence, independently maintained ground truth, and operational costs. The paper reports no new agent deployment, business-system experiment, causal effect, or measured performance improvement. Its contribution is a vendor-neutral, falsifiable specification for deciding when operational traces warrant a claim about business reality, and for preserving an explicit unknown state when the available evidence does not suffice.