Managing AI Coworkers Means Managing Systems, Not Personalities

The phrase "AI coworker" is spreading as software moves from answering isolated prompts to completing longer tasks through connected tools. It is an accessible label, but a poor management model. A system has no employment relationship, professional judgment, personal accountability, or understanding of office consequences. It produces outputs and actions from models, instructions, data, permissions, and runtime controls. Managers should therefore treat delegated AI work as an operational system. The central questions are not whether the system appears diligent or collaborative. They are who owns the workflow, what authority has been granted, how output quality is tested, which workers are affected, and what happens when the system fails. The trend is from individual prompting to shared workflows Early workplace adoption often happened one employee and one conversation at a time. The next phase is more persistent: reusable instructions, scheduled runs, connected applications, shared agents, and tasks that continue without a person directing every step. OpenAI's May 2026 workspace agent material illustrates that direction by describing repeatable team workflows with shared tools, permissions, approvals, and administrative controls. It is a vendor product source, not evidence that every organization is ready for this model. Anthropic's June 2026 Economic Index adds a different kind of evidence from usage and survey data. It reports growing user expectations about the share of tasks AI may handle, while also noting that judgment and management were frequently named as capabilities AI lacks. The survey is drawn from Claude users and is not representative of the general workforce. Its value is in showing the gap managers must address: more work may be delegated even though contextual judgment remains difficult to encode. That combination changes the unit of management. A prompt can be reviewed as an individual artifact. A persistent workflow needs a lifecycle: design, authorization, testing, release, monitoring, incident handling, revision, and retirement. The coworker metaphor hides the accountability chain Anthropomorphic language encourages people to infer intent, common sense, loyalty, or learning that the system has not demonstrated. It can also obscure the humans and organizations that remain responsible. When an automated report is wrong, "the agent did it" is not an adequate root cause or accountability statement. Map the real chain instead. A business owner defines the outcome and accepts operational risk. A technical owner maintains integrations and controls. Data owners approve access to sources. Security and legal teams set requirements where needed. Reviewers assess outputs. A vendor supplies models or infrastructure under stated terms. Workers and customers may experience the effects and need a route to question or correct them. NIST's AI Risk Management Framework clearly calls for documented roles, clear lines of communication, defined human oversight, ongoing monitoring, and executive responsibility for deployment risk. Those are system management duties. Giving the software a friendly role name does not replace any of them. Four management rules replace personality based supervision Assign an outcome, not a broad identity. Define a bounded job such as checking invoices for missing fields or preparing a draft status report. Specify inputs, source of truth, completion criteria, forbidden actions, timing, and the human owner. Avoid labels such as "finance teammate" that imply an open ended remit. Grant capability by task. Access should be limited to the records, tools, actions, and duration required. Separate read, draft, and execute permissions. Require a new authorization when the destination or consequence changes. Personal accounts, undocumented integrations, and shared credentials make both security review and continuity harder. Evaluate behavior as well as the final artifact. A correct report can still come from an unsafe process, such as using an unapproved source or exposing data in a log. Test source selection, tool choice, arguments, policy compliance, stopping behavior, and escalation. Include ordinary cases, missing inputs, conflicting records, prompt injection, permission denial, timeouts, and duplicate requests. Treat change as a release. Models, connected systems, policies, and source data change. Version instructions and tool contracts, record approvals, run regression cases, monitor production, and maintain rollback and retirement procedures. A system that improved during one demonstration has not acquired a stable professional skill. Management must include workers affected by the system AI management is not only a technical quality program. Workflows can change pace, discretion, monitoring, task allocation, and how performance is judged. OECD research on algorithmic management found that surveyed employers reported benefits as well as concerns about accountability, explainability, and worker health. The study covers a broader class of management software than generative agents, but its lesson applies: managerial automation affects employment conditions, not just output efficiency. Involve the people who perform and receive the work before deployment. Ask where tacit knowledge changes the decision, which exceptions are common, what evidence reviewers need, and which mistakes create hidden rework. Provide a clear way to contest or correct automated results. Train workers on both use and limits, without making them personally liable for defects they cannot inspect or control. Managers should also distinguish assistance from surveillance. Logs needed for security and troubleshooting should have defined purposes, access, and retention. Do not quietly convert interaction telemetry into individual performance scoring. Where law, contracts, or collective agreements apply, engage the appropriate specialists and representatives. A portfolio view prevents shadow systems and duplicate risk As experimentation spreads, organizations need an inventory that is useful enough to maintain. For each workflow, record the purpose, owner, users, model or service, connected data, permissions, review level, deployment status, last evaluation, incidents, cost boundary, and retirement condition. The inventory should cover purchased products as well as employee built automations. Then classify workflows by consequence and reversibility. A private draft from public sources needs lighter controls than a customer message, a record update, a financial recommendation, or a decision affecting a worker. Raise review and authorization requirements with impact. Some proposed workflows should remain assistive; others should not be deployed. Central standards should make safe reuse easier. Provide approved connectors, evaluation templates, logging, incident channels, and shared workflow components. At the same time, keep named business owners close to each deployment. A central AI team cannot know every operational exception, while a local team should not define security and data rules alone. Measure dependable outcomes and organizational effects Do not judge an AI workflow by demo fluency, raw activity, or a vendor productivity claim. Establish a baseline for the existing process and define the outcome that matters: accuracy against a rubric, completion without unauthorized action, correction rate, cycle time, service quality, incident frequency, accessibility, or another mission specific measure. Include the cost of integration, review, failures, and maintenance. Sample successful runs as well as visible failures; silent drift often appears in work that looks complete. Review escalation patterns and worker feedback. Check whether people are doing more judgment rich work or merely absorbing more verification and exception handling. Segment results where aggregate averages could hide unequal effects on roles or groups. The management shift is ultimately straightforward. Delegated AI increases the amount of software mediated work, so organizations need clearer workflow ownership, narrower authority, stronger evaluation, and more deliberate worker participation. Calling the system a coworker may make the interface approachable. Running it responsibly requires the discipline used for consequential processes and production software. Sources Anthropic Economic Index report: Cadences https://www.anthropic.com/research/economic index june 2026 report Build, Govern, and Scale Workspace Agents in ChatGPT Enterprise https://academy.openai.com/en/public/events/build govern and scale workspace agents in chatgpt enterprise zs4dny9yey AI RMF Core https://airc.nist.gov/airmf resources/airmf/5 sec core/ Algorithmic management in the workplace https://www.oecd.org/en/publications/algorithmic management in the workplace 287c13c4 en.html?wcmmode=disabled%27.html Cover image credit Cover image by Robert Scoble , made available under CC BY 2.0 . WIKIVISE cropped and converted the source image.