In a July 28 review of enterprise AI adoption, Microsoft wrote that agentic workflows need to be built, observed, and tuned against the outcomes an organization seeks. It also emphasized governance, observability, security, and continuous improvement. The article includes Microsoft's product and customer narrative; it is not evidence that another company will obtain the same results.

The transferable management question is simpler. After an export team creates agents for inquiries, translation, quotation preparation, content, documentation, and follow-up, who decides which ones should continue, be consolidated, or be retired?

Creation needs a corresponding exit mechanism

Agents often begin as a demonstration and then acquire access to a spreadsheet, inbox, group, or business system. Six months later, several may summarize the same messages, translate the same files, or generate overlapping reminders using different prompts and knowledge sources.

The organization can claim that it has many AI tools while lacking a clear picture of daily users, error rates, maintenance owners, or process value. This is agent inventory without agent operations.

Every agent should begin with a task definition, audience, input boundary, allowed action, human handoff, quality measure, maintenance owner, review date, and exit condition. An agent with no active user, persistent correction burden, outdated knowledge, or substantial overlap should enter consolidation or retirement review.

Experimental and production status should remain separate. A limited trial can work with minimized examples and manual transfer, but access to live mailboxes, customer records, pricing data, or system writes requires a new review. Production status should confirm identity, permissions, logging, exception ownership, data retention, and rollback rather than inherit approval from a demonstration.

Observe the task outcome, not the call count

Invocation volume shows that a tool was triggered. It does not show that work was completed correctly. An inquiry-classification agent can be assessed through label consistency, manual corrections, and downstream routing. A quotation-preparation agent needs field completeness, source traceability, and human confirmation. A content agent needs factual-error, duplication, and pre-publication rejection records.

Each task requires its own evidence. A common dashboard may show status and ownership, but it should not collapse unlike activities into a single “AI productivity” score.

Cost also needs a complete definition. Model fees are only one component. Review time, exception handling, duplicated integrations, source maintenance, and the cost of correcting an external error can exceed the visible usage charge. Portfolio review should compare the full operating burden with the task value and the available non-agent alternative.

Versioning is part of observability. A model, prompt, source, permission, or workflow change can alter behavior. The team should retain the version, reason, test set, approval, and rollback point so an error can be traced to a specific change rather than attributed vaguely to AI.

What this means for Chinese exporters

Lean export organizations pay a real maintenance cost for unused agents: permissions must be reviewed, source data must be refreshed, prompts must be maintained, and exceptions must be handled. Portfolio governance concentrates those resources on a small number of workflows with accountable ownership and verified quality.

Retirement does not mean deleting the evidence. The task definition, historical outputs, issue log, and retirement reason should be archived. Credentials, scheduled jobs, write access, and integrations must be revoked. If the business condition later returns, the team can reassess a documented design rather than unknowingly extend an obsolete background process.

The portfolio review should also distinguish experimentation from production. A limited experiment may remain useful without permanent system access. Production status should require a current owner, monitored quality, approved data, and a functioning human escalation path.

Action checklist

1. Maintain an agent register covering the task, users, owner, source data, permissions, version, and review date. 2. Assign task-specific quality measures instead of treating calls, messages, or generated words as business outcomes. 3. Review unused, overlapping, correction-heavy, and persistently failing agents every month. 4. Standardize the knowledge source and output contract before consolidating two agents. 5. Revoke credentials, schedules, integrations, and write permissions when retiring an agent. 6. Require regression testing and a rollback point for material model, prompt, source, or permission changes. 7. Keep production status only for agents with active use, evidence, ownership, and a working exception path.

Sources