Autonomous workers do not only create finished work. They also create pending decisions. If nobody burns that inventory down, the org quietly drowns in half-owned judgment.
Decision burndown treats every unresolved agent choice as operational inventory, not conversational noise or harmless blocked work.
Agents can finish steps faster than organizations can resolve exceptions. The result is not speed; it is a growing warehouse of tiny unanswered decisions.
Operators see which decisions deserve human attention, which need better policy, and which should expire without dragging the workflow.
Classic ops teams counted tickets, tasks, meetings, and cycle time. Agent ops needs a harsher metric: how many decisions did machine workers create that nobody has resolved?
A sales enrichment agent finds two conflicting company records. A renewal agent sees usage down but executive engagement up. A support triage agent detects refund language but no eligible policy. A forecast agent sees a rep note that contradicts CRM stage. None of these are big enough to become projects. All of them are big enough to make the agent pause or guess.
If the system has no decision burndown, those moments scatter across Slack threads, task comments, private agent memory, and “blocked” labels. Work appears calm while the control plane fills with invisible debt.
Items where the agent knows the best next action but lacks authority: discounts, customer-facing changes, account ownership shifts, public claims, or high-value routing.
Items where the agent has competing interpretations: duplicate records, stale instructions, ambiguous segment rules, conflicting success notes, or missing source evidence.
Items where delay changes the answer. If nobody responds by the timeout, the worker should suppress, rollback, route to safe default, or close without action.
Items that repeat often enough to become policy. If ten agents ask the same question, that is not review work; that is missing operating system.
When humans say a task is blocked, a manager can ask follow-up questions and read body language. When agents say blocked, the queue becomes a black box unless the system records the decision type.
“Blocked on approval” is not the same as “blocked on missing data.” “Blocked because the customer is risky” is not the same as “blocked because the worker exceeded scope.” Treating all of those as one status hides the real management problem.
A decision burndown splits the pile. Operators can clear five approvals in two minutes, assign one policy owner to repeated clarification, and let low-value expired items die instead of haunting dashboards for a week.
Approve, clarify, expire, codify, escalate, rollback, suppress, or route. No generic “blocked” without a type.
The person, role, or control group allowed to resolve the decision, plus the backup owner when the queue ages out.
How long the decision has waited, measured against the SLA of the workflow it is holding.
The pipeline, customer, reputation, compliance, or operational value affected by waiting or guessing.
What the agent does if nobody answers: safe send, no-send, route, rollback, suppress, or close.
An AI-native org can look productive while the real system is rotting. Agents complete easy paths, generate polished artifacts, and keep dashboards green. Meanwhile, every non-routine branch accumulates in decision purgatory.
The fix is not more meetings. It is a visible queue of unresolved judgment with ruthless aging rules. If a decision is high value, assign an owner. If it repeats, codify it. If it is low value, expire it. If it keeps blocking revenue, redesign the workflow.
This is management, not prompt engineering.
Pick one live agent workflow and require every pause to write a decision type, owner, age, value at risk, and default action.
Do not start with the loudest decisions. Start with the oldest unresolved ones. Age reveals where the org is avoiding ownership.
If the same decision appears three times, turn it into a rule, approval threshold, or escalation grammar. Do not keep paying humans to answer identical questions.
Give low-value decisions expiration rules. A queue that can only grow is not governance; it is landfill with nicer labels.
Before adding another agent, measure how many unresolved decisions your current agents are already creating.
The blunt rule: if a machine worker creates a decision, the control plane must either resolve it, expire it, or turn it into policy.