tl;dr: Real autonomy means explicit permission to decide, not a vague invitation to use judgment. Leaders need to distinguish capability from decision rights, define when permission changes, and give a named human owner authority over the boundary. The test is whether a defensible decision within that boundary remains genuinely delegated when its result disappoints you.
Imagine telling a team to use its judgment, then discovering that it has.
A customer received a remedy you would not have chosen. The decision was reasonable. The team thought it had permission. You thought it would check first.
Nobody can find the point where those two beliefs were supposed to meet.
The scene is hypothetical. The design problem is specific: "Use your judgment" expresses confidence in a person without saying what that person may decide.
Add an AI assistant and the ambiguity becomes more consequential. Preparing a response, choosing a remedy, and issuing it are different permissions, even when the software makes them look like one convenient workflow.
Autonomy needs an explicit boundary between what an actor can do and what the organization has authorized it to do.
Otherwise, freedom is something leaders announce and everyone else has to interpret.
Permission should survive an unwelcome result
If you give someone authority only until they choose differently from you, you have given them a guessing exercise. They’re supposed to act independently while reproducing the decision you haven’t shared.
A mind-reading requirement, now listed under empowerment.
Meaningful autonomy asks more of the leader. You have to distinguish a decision you dislike from one made outside the agreed authority. Those aren’t automatically the same thing.
An unwelcome outcome may expose poor judgment. It may also expose an inadequate boundary, an overlooked condition, or ordinary uncertainty. Your response should depend on which happened. Surprise alone is a poor basis for deciding that someone exceeded their mandate.
Negligence still matters. So do consequences. But assess the conduct against the authority that existed before the result. Don’t invent the authority afterward.
The earlier TBAI edition The Delegation Trap: AI Can Do Your Work – But It Can’t Own The Outcome examined why responsibility remains human when execution moves elsewhere.
Permission needs the same clarity, before execution begins.
Give the freedom a usable shape
One practical approach is an autonomy contract: a short, explicit agreement about the decisions a person, team, or system may make without returning for permission.
It’s a proposed operating practice. It isn’t a validated management formula or a new legal instrument; its value has to be tested in the work it governs.
The questions are concrete:
- What purpose is this authority meant to serve, and which trade-offs remain unacceptable?
- Which decisions may the actor make, and which may it only prepare or recommend?
- What is the scope: customers, systems, data, access, time, and resources?
- Which conditions require a pause or escalation, and who receives it?
- What evidence must be retained so a consequential decision can be understood?
- Who owns the boundary, can change it, and is responsible for organizing repair?
The agreement needs enough detail to distinguish a permitted action from an exception. It needn’t anticipate every possible action. If every choice still requires a fresh interpretation from headquarters, you’ve documented a dependency rather than granted authority.
For AI-enabled work, ownership deserves particular care. The NIST AI Risk Management Framework Core calls for documented risk responsibilities and communication lines, executive responsibility for AI risk decisions, and differentiated roles for human-AI oversight.
Those are outcomes in a voluntary framework, not proof that organizations have implemented them. They do help sharpen the question: what is the named owner actually responsible for? A name beside the system doesn’t answer it.
The person answerable for the boundary needs authority to change it. Otherwise, accountability is a title with the controls removed.
Capability is not a decision right
Imagine a support team authorized to resolve routine cases within an agreed remedy limit. Cases outside that limit go to a named decision-maker.
An AI assistant could gather the relevant information and prepare a proposed remedy. That would not, by itself, authorize the assistant to issue it. The team might hold the right to approve and execute the remedy while the assistant holds only the right to prepare it.
That distinction belongs in the workflow and its access permissions. Agreement around a meeting table doesn’t set the permissions.
The same hypothetical case exposes another gap. A limit tells you where to stop; a route tells you what happens next. "Escalate exceptions" is incomplete unless somebody can receive the case, decide it, and tell the team what happens while it waits. Silence must not expand the assistant’s permission.
Nor should technical success quietly expand the mandate. Preparing ten sound recommendations does not confer permission to send the eleventh.
Demonstrated judgment may justify broader decision rights for a human team. Testing may support a carefully scoped expansion for an AI system. Either way, the owner must explicitly decide to change the boundary.
Decide when permission changes
A boundary needn’t be a permanent restriction. It can specify when an actor may continue, must pause, or needs different authority.
Google’s published Example Error Budget Policy makes this concrete in service operations. Releases proceed under the release policy while the service meets its objective. Exhausting the error budget triggers a halt on changes and releases, with urgent and security exceptions. Disputes escalate to the CTO.
The example explicitly says the halt is not punishment. It gives the team permission to focus on reliability.
This is an illustrative service policy, not a universal business risk budget or a claim about how every Google team operates. The useful principle is narrower: an agreed condition can change permission. A leader’s reaction to bad news needn’t decide it.
For AI agents, Anthropic’s engineering guidance, Building effective agents, describes human checkpoints, stopping conditions such as iteration limits, and sandbox testing. It also warns about higher costs and compounding errors.
That’s vendor guidance, not independent proof of improved outcomes. The design lesson still holds: an open-ended task does not require unlimited permission.
Checkpoints and stopping conditions don’t guarantee safety. If nobody can enforce the boundary, its presence on paper offers little reason for confidence.
Some decisions should remain outside the boundary
There is a fair objection here. Boundaries can become bureaucracy with a more attractive name.
A contract that reserves every consequential choice for the leader leaves the team free to make only decisions nobody cares about. Calling that autonomy doesn’t improve it.
Match authority to consequence. Delegating more is not an end in itself.
A contained internal experiment may warrant more freedom than an external commitment. But reversible needs careful handling. You might undo a system change without undoing what a customer disclosed, believed, or acted upon. A correction is not always a restoration.
Where consequences affect safety, rights, or obligations, appropriate domain review may justify keeping the decision outside autonomous execution altogether. Nothing in this argument requires every activity to become autonomous.
It does require honesty about the boundary. If you’re unwilling to delegate a decision, say so. If you delegate it, let the recipient exercise that authority within its terms, including when the choice would not have been yours.
Start with one decision you keep taking back
Choose a recurring decision that is supposedly delegated but still returns to you.
Write down the authority its recipient actually holds. Separate preparation from commitment. Name the limits, the conditions that change permission, and the person who can decide an exception. Check that the workflow permits what the agreement promises and prevents what it forbids.
Then examine your side of the arrangement. Could someone make a defensible choice inside that boundary, achieve an imperfect result, and still retain your support for the authority you granted?
If the answer is no, the missing boundary may be yours.
Grant enough authority for judgment to matter, and make its limits clear enough to defend. The demanding human task remains: recognizing who can use the freedom well, rather than simply produce a convincing account of having done so.
Rob
P.S. Which supposedly delegated decision would you take back simply because your team made a defensible choice you wouldn’t have made?
If you are a leader or founder trying to move from answer-based authority to responsible orientation, let’s talk. I help organizations design the human judgment, context, and decision structures that keep AI-enabled speed connected to consequence.
Find out more at www.robkonrad.com and www.thinkingbeyond.ai.
References
- NIST, AI Risk Management Framework Core, GOVERN 2.1, 2.3, and 3.2. Voluntary framework guidance.
- Steven Thurgood, Google SRE Workbook, Example Error Budget Policy. Illustrative service policy, 2018.
- Anthropic, Building effective agents, 2024. Engineering guidance; the page notes subsequent tooling changes.
- TBAI, The Delegation Trap: AI Can Do Your Work – But It Can’t Own The Outcome. Previous LinkedIn newsletter edition.






