If agents can assign work across teams, the org needs a routing constitution before machine labor starts moving priorities.
A routing constitution is the written rule set for how autonomous workers create, prioritize, escalate, and close work across teams.
Machine workers do not just execute tasks. They redirect human attention, consume specialist capacity, and change what the organization treats as urgent.
AI-native teams get faster handoffs without letting every agent become an unofficial manager of everyone else's queue.
Traditional org design assumes people negotiate priorities through managers, meetings, rituals, and escalation paths. Agentic work breaks that assumption. A machine worker can create tickets, tag owners, rewrite priorities, schedule follow-ups, and mark dependencies before a manager sees the pattern.
That is not a small workflow detail. It is labor allocation. The agent is deciding whose time gets spent, which customer gets urgency, which internal promise counts as real, and which queue absorbs the operational debt.
A routing constitution makes that authority explicit before it becomes invisible management.
Define which agents may create cross-team work, which teams they may touch, and which requests must remain drafts until a human owner accepts them.
Require proof before work becomes assignable: customer impact, revenue exposure, SLA breach, compliance risk, verified data, or manager-approved exception.
Ban vague urgency. Agents should route work as severity, deadline, dependency, dollar exposure, confidence, and rollback cost.
Decide who can mark routed work done, rejected, duplicated, blocked, or escalated so agents do not close loops by satisfying their own local goal.
Bad routing rarely looks like chaos at first. It looks productive. Tickets appear with clean summaries. Owners are tagged. Follow-ups are scheduled. Dashboards show motion. Everyone feels the machine is helping.
Then the hidden cost lands. Product is answering sales exceptions. RevOps is cleaning enrichment fallout. Customer success is carrying campaign promises. Finance is explaining discount logic created by a growth worker. Every agent has a local reason. No one owns the total load.
That is how AI-native organizations get slower while every individual workflow appears faster.
Which worker created the assignment, which policy or objective triggered it, and which data was used as evidence.
The team, role, or named accountable owner that may accept, reject, delegate, or demand more proof.
The customer impact, revenue amount, deadline, compliance exposure, or operating dependency that justifies the work.
The point where a machine-created request becomes committed work instead of a suggested task in a triage queue.
The route for disputing bad assignments, overloaded teams, conflicting priorities, or agents that repeatedly route weak work.
The first wave of agent ops focused on tool permissions: can the worker read the CRM, write to support, update tasks, or send messages? That matters, but it misses the deeper issue. The most expensive permission is not write access. It is the ability to redirect people.
For a Head of Growth, the value is immediate. You can let agents discover work, draft tasks, and propose owners without giving them the right to conscript RevOps, sales, product, success, or finance into every local optimization.
That separation keeps machine labor useful. Discovery can be fast. Assignment stays governed.
Teams are receiving agent-created work that feels mandatory even though no owner accepted the priority.
Machine workers use vague labels like important, blocker, or high priority instead of evidence-backed severity and deadlines.
Agents keep routing the same weak request because rejected work is not fed back into policy, examples, or permission limits.
Pick one agent workflow that regularly creates work for another team: renewals, enrichment cleanup, campaign QA, pricing, support, or sales assist.
State whether the worker can draft, suggest, assign, escalate, or close work for each receiving team.
Make every routed task carry source data, confidence, customer impact, deadline, owner, and why it cannot remain inside the originating queue.
Track acceptance rate, rejection reasons, duplicate work, stale assignments, and downstream load created per accepted agent request.
Before letting another autonomous worker assign cross-team work, write a one-page routing constitution.
The blunt rule: agents may discover work at machine speed, but they should earn the right to assign work through governed routing quality.