Name the recurring business outcome, not the software action. “Qualify expansion signals” is a job. “Summarize Slack threads” is only a task.
// AI-NATIVE ORG DESIGN · GOVERNED AGENT LABOR · June 5, 2026
AI-Native Orgs Need Work Contracts Before Agents Become Shadow Headcount
If an agent owns recurring work without a contract, the company has hired shadow headcount with no manager, no job description, and no clean exit.
AI-native companies do not fail because they use too many agents.
They fail because nobody knows which agents are actually employed by the operating system.
A machine worker starts as a workflow. Then it handles a queue. Then another team copies the pattern. Then it gets permission to update records, message buyers, route work, summarize calls, draft strategy, approve exceptions, or trigger follow-up. Six weeks later the org has thirty semi-autonomous workers doing real labor, but the org chart still pretends nothing changed.
That is shadow headcount. It does not show up in HR. It does not have a manager. It does not get reviewed. It does not have a termination path. But it influences revenue, customer trust, internal priorities, and the shape of human work.
AI-native org design needs work contracts for agents before autonomy scales into invisible labor sprawl.
A prompt is not a job description.
A prompt can tell an agent how to act. A work contract tells the company why the agent exists, who owns it, what authority it has, what quality means, when it must escalate, and how it can be removed without breaking the operating rhythm.
Agents need managers before they need more tools.
The highest-leverage control is not another integration. It is named ownership. Every recurring machine worker should have one accountable human who can narrow scope, change standards, pause execution, and retire the worker when its job stops being worth the risk.
Define what the agent can read, write, recommend, send, approve, spend, escalate, and learn from without additional review.
Attach the worker to a person, not a department. Departments do not notice drift. Named operators do.
Set concrete acceptance checks: accuracy, latency, tone, evidence, exception rate, customer visibility, and rollback cost.
Specify how to pause, replace, archive, or delete the worker without losing context or leaving half-owned work in production.
Grounded examples
Where contracts stop shadow headcount
The lead-routing worker
An agent starts assigning inbound accounts by fit. Without a contract, it quietly changes territories, rep load, and SLA expectations. With one, routing authority, exception rules, and sales ownership are explicit.
The renewal-risk worker
A churn agent flags accounts and drafts save plans. The contract defines when it can notify CSMs, when it can request executive help, and which risk claims require evidence before entering a customer record.
The discount-advice worker
A pricing agent recommends concessions. The contract separates advice from approval, ties suggestions to margin policy, and records who accepted exceptions when the machine sees urgency.
Build the agent work contract before expanding permissions
The contract should be small enough to maintain and strict enough to matter. It is not legal theater. It is the operating document that lets humans manage machine labor like labor instead of treating it as background automation.
Purpose
The business result the agent is responsible for producing or improving.
Kill any worker whose purpose is only “save time” without a measurable operating outcome.
Scope
The queues, systems, customers, records, regions, segments, and decision types the agent can touch.
Scope creep should require an amended contract, not a casual prompt edit.
Authority
What the agent can do alone, what it can recommend, and what needs human approval.
Authority should shrink automatically after errors, anomalies, or unresolved review debt.
Evidence
The proof required before the agent changes state, escalates, messages, forecasts, or spends scarce resources.
No evidence means no durable action. Drafts are cheap. Operating changes are not.
Exit
The process for pausing, replacing, retraining, or retiring the worker.
If nobody knows how to fire the agent, it is already too embedded.
Operator tests
Four checks for this week
List every agent that handles work more than once a week. If it touches a business process repeatedly, it counts.
Assign one human owner to each worker. If ownership is shared, the contract is already weak.
Write down which actions the worker can execute and which ones it can only recommend.
For one high-risk agent, document the pause switch, data handoff, replacement path, and cleanup owner.
The AI-native org chart is not complete until machine workers have contracts, managers, and exits.
Concrete takeaway for this week
Create one work contract for the most embedded agent in your company.
- Name the job. Describe the business outcome the worker owns, not the tool it uses.
- Bound the authority. List what it can read, write, send, approve, escalate, and learn from.
- Attach an owner. Give one human the power to change scope, reject output, reduce access, or retire the worker.
- Set the exit. Define the pause path and what happens to queued work, memory, logs, and downstream dependencies.
The point is not bureaucracy. The point is leverage you can govern. Agents should make the company faster, but every durable worker needs a contract before it becomes part of the operating system.