Accountable Enterprise Agents A Critical Review of Authorization Execution and Recovery

Enterprise agents combine language models, persistent state, and executable tools to pursue goals across business systems. Their practical significance lies in the transition from generating information to changing organizational state. This critical review examines authorization, execution, and recovery at that transition and develops disconfirmable tests of the proposed controls. A purposively selected literature corpus links reasoning and tool-use methods with agent benchmarks, retrieval and memory research, security studies, and organizational risk guidance. The review is interpretive rather than systematic and does not report new experiments. Its central argument is that execution authority should be separated from model-generated plans. A reference architecture is developed around authenticated task intake, evidence retrieval, bounded planning, external policy enforcement, transactional execution, and independent outcome verification. The architecture is applied to a hypothetical procurement workflow to show why correct language output, valid API calls, and authorized business outcomes are different properties. The synthesis identifies coupled failures involving stale evidence, ambiguous delegation, repeated execution, and incomplete recovery. It proposes an evaluation framework covering task completion, policy adherence, repeatability, recovery, human workload, and cost per verified outcome. Future research should test controls under policy changes, asynchronous failures, and organizational handoffs rather than assume that benchmark success establishes production readiness. Enterprise agent design is best understood as the design of accountable execution systems in which language models propose actions within explicitly constrained authority.