Enterprise AI Workflows: State, Approvals and Recovery

The demo works. The process still fails.

An agent reads a customer application, calls an identity service and opens a case. The demonstration ends successfully. Production does not. A screening provider times out, a document arrives late, an approver is unavailable and a retry creates the same case twice. The missing component is not another model. It is a business process that persists state and survives exceptions.

The difference between a tool call and an outcome

Tools perform actions; workflows make commitments. A workflow knows the case it is processing, the owner of each transition, the conditions for waiting, the evidence required to proceed and the definition of completion. It can suspend for an event, retry selectively or escalate to a human without losing its place.

Conceptual reference pattern; enterprise policies and integrations determine implementation.

Reference architecture: customer onboarding

StageActorInvariantIntakePortal + workflow engineOne case ID per applicationDocument reviewAgent + retrieval toolsSource references preservedRisk checksEnterprise APIsAuthoritative results retainedExceptionHuman approverExplicit accountable decisionFulfilmentCore systemsIdempotent update with reconciliationCompletionWorkflow engineBusiness outcome and evidence recorded

Five design principles

Durable state: persist business status independently from transient conversation context. Explicit contracts: distinguish suggestions from authorized commands. Event awareness: wait for authoritative events instead of repeatedly polling with an LLM. Compensation: specify what happens after partial completion. Observability: correlate traces with case IDs and business outcomes.

Where agents add value

Agents are useful at ambiguous, document-rich or judgment-supporting stages: classify incoming material, extract supporting evidence, suggest exception routing and summarize a case for human review. Deterministic rules and transactions should continue to enforce invariants where correctness matters.

Measure completed work, not agent activity

Track end-to-end cycle time, touchless completion rate, rework, exception resolution time, failed handoffs and cost per successfully completed case. A workflow that invokes fewer agents but closes more cases is the better system.

Archy Platform brings agent orchestration and enterprise integration into a broader operating model. See archy-ai.com.

The hardest transition is the one after an interruption

Imagine that a document assessment finishes while an analyst is correcting the customer’s legal entity. The assessment can be internally consistent and still belong to the wrong version of the case. Persist the case revision with the assessment, then compare it before accepting the result. If the revision changed, reopen the affected activity rather than allowing a plausible but obsolete recommendation to advance the process.

The same principle applies to approvals. Approval should refer to a particular proposed action and evidence set. A subsequent change to the payee, amount, document or policy can invalidate it. The workflow should make that relationship explicit. A generic approval flag attached to the conversation does not capture what the person actually agreed to.

Prove recovery before increasing autonomy

A useful acceptance exercise deliberately interrupts the process after an external update but before acknowledgement. On restart, the coordinator must determine whether the update happened, reconcile the result and avoid repeating it. Another exercise revokes a permission while the case waits for a human. Resumption should recheck current authority instead of treating yesterday’s access as a permanent grant.

Start with one business outcome, one system of record and a named operational owner. Define the waiting states and manual recovery path before expanding the agent’s role. A workflow that exposes a clear exception for an operator is more useful than one that retries indefinitely and conceals the unfinished case behind a successful model response.

Explore Archy Platform.

Example recovery contract for an onboarding workflow

StateCompletion evidenceRecovery ruleAwaiting evidenceRequired documents and versions recordedResume the same case when missing evidence arrivesScreening pendingScreening request ID and response statusReconcile a timeout before submitting againHuman review requiredDecision attached to the proposed action and evidenceExpire stale approvals when material evidence changesReady to commitPolicy checks pass and approval remains validUse an idempotency key for the external writeCompletedDestination confirms the accepted resultReconcile discrepancies without reporting false completion

This is a reference design. The business owner must define the states, deadlines and approval rules for the actual process.

For the related architecture decisions, see AI Workflow Orchestration: Beyond Agent Tool Calls; AI Agent Governance: Identity, Authority and Accountability.