This is part four of Human Systems for AI Work.
I want AI to take work off my plate. I do not want it to quietly take ownership of decisions I am still responsible for.
The distinction is easy to state and harder to maintain. Preparing a recommendation can involve choosing which evidence matters, defining success, and deciding whose concerns deserve attention. By the time a person approves the result, much of the consequential judgment may already be embedded in it.
My boundary is not that AI must never participate in important work. It is that people must retain ownership of the purpose, acceptable consequences, and commitments made through that work. Those responsibilities can move between authorized people. They cannot be discharged by accepting a model's answer.
That still leaves substantial room for useful assistance.
Delegate preparation, not responsibility
Consider a manager deciding whether to change a team's working hours. AI could summarize scheduling constraints, compare proposals, identify unanswered questions, and draft an announcement.
Those are useful tasks. But somebody must decide what counts as an acceptable burden on the people involved.
A proposal that maximizes shared availability might disadvantage a caregiver. A proposal that treats every stated preference equally might miss the difference between convenience and necessity. The model cannot resolve those questions simply by producing a more detailed schedule.
The manager needs to talk with the affected people, decide which obligations apply, and explain the tradeoff. If the schedule causes harm, “the recommendation looked reasonable” does not resolve the problem.
I separate the work into two questions: What can the system help me understand? What am I committing to after considering that help?
The second question should have a human owner before the first produces an answer.
Keep ownership of the goal
A request such as “improve our customer support” leaves almost everything important unspecified.
Improvement could mean shorter waits, more accurate answers, fewer repeat contacts, lower cost, or better treatment of difficult cases. These goals can conflict. A system asked to recommend an improvement may choose an interpretation that sounds sensible without making the choice visible.
Before asking for recommendations, I want a short statement of purpose: what we are trying to improve, for whom, and what we refuse to sacrifice to get there.
For a support team, that might be: reduce unnecessary waiting without making customers repeat their story or making exceptions harder to reach. That statement does not answer every operational question. It makes several otherwise attractive recommendations unacceptable.
AI can challenge the goal. It can point out contradictions or suggest a clearer formulation. That is assistance with judgment, not authority to settle it.
The failure mode is accepting the system's definition of success because it arrived inside a polished plan. Review the goal separately from the proposed method.
Keep ownership of acceptable harm
A recommendation often contains a sentence like “the benefits outweigh the risks.” I want to know who is entitled to make that judgment and who bears the downside.
Suppose a small organization is considering replacing individual onboarding sessions with AI-assisted self-service. The proposal may save staff time. It may also leave people with unusual circumstances unable to complete the process.
A list of benefits and risks is not enough. Someone needs to determine whether the people most likely to struggle have a usable alternative, whether they can ask for help without penalty, and whether the organization can actually provide that help.
AI can help identify potential effects. It cannot supply the missing agreement of the people affected. Nor should a generated description of their needs substitute for hearing from them.
This does not mean every decision requires unanimous consent. Some roles legitimately include authority to decide despite disagreement. The responsibility is to use that authority openly rather than presenting a contested value judgment as an objective result.
When a recommendation uses words like reasonable, fair, or acceptable, treat them as invitations to identify the standard and its owner.
Keep ownership of commitments
Drafting a promise and making a promise are different acts.
An AI-written customer message might offer a refund, a delivery date, or an exception to a policy. Each could be reasonable. Each could also exceed the authority of the person sending it.
I want the commitment reviewed independently of the writing. The question is not just whether the message is clear. It is whether we can and should do what it says.
For an ordinary message, this review can be brief. Identify the promise, confirm the person authorized to make it, and check that the organization can fulfill it. Higher-consequence commitments deserve more scrutiny.
The same boundary applies to apologies. AI can help someone find direct language and remove defensive phrasing. It cannot make the speaker accept responsibility. An apology approved for tone but unsupported by changed behavior is still inadequate.
There is nothing inherently more sincere about typing every word yourself. What matters is whether you understand, mean, and will stand behind the words you send.
Do not outsource the encounter with an affected person
Some work requires a person to be available for a response, not just to deliver information.
A difficult performance conversation is one example. AI can help a manager distinguish observations from interpretations, organize relevant evidence, and prepare questions. That preparation may improve the conversation.
But a generated assessment should not become a substitute for hearing the employee's account. Missing context may change the meaning of the evidence. The employee may need to challenge the standard itself, not merely correct a date in the summary.
Keeping a human in the process means providing someone who can listen, reconsider, and act. A person who only repeats the recommendation is not offering meaningful recourse.
This principle should not become a demand for unnecessary meetings. Some people prefer written communication, need accessibility accommodations, or benefit from time to respond. The essential feature is a real opportunity to influence the decision, not a particular communication format.
Ask whether the affected person can reach someone with authority to change the outcome. If not, adding a human signature has not solved the problem.
Make approval an actual decision
A final review is weak when the reviewer lacks time, evidence, or permission to disagree.
Imagine receiving a completed proposal minutes before a deadline. The recommendation is confident, the supporting material is long, and the announcement is already drafted. Formally, you can reject it. Practically, the workflow has made acceptance the easiest path.
I prefer to reserve the decision point before producing the final communication. Review the purpose and evidence while alternatives are still cheap to consider. Draft the announcement after choosing, rather than letting the announcement create pressure to proceed.
For consequential work, the owner should be able to explain why this option was chosen, what uncertainty remains, and what would justify reconsidering it. They do not need to repeat every detail. They do need an account that is more specific than “I reviewed it.”
If they cannot give that account, the next step is clarification, not approval.
Leave a record of the boundary
The agreement can be short. For a staffing proposal, I would want something like this in ordinary language:
The AI will organize the supplied workload information and compare scheduling options. It will not rank employees or communicate a decision. The team lead owns the recommendation, checks constraints directly with staff, and obtains the required approval before changing hours.
That is enough to distinguish assistance from authority. It also makes missing work visible. If nobody has checked the constraints with staff, the proposal is not ready just because the comparison is complete.
Record the boundary where the work lives, not only in someone's memory. Revisit it if the task expands. Permission to summarize feedback should not silently become permission to decide which complaints deserve a response.
A boundary also needs a stop condition. Missing evidence, unclear ownership, or a newly discovered consequence can be reasons to pause rather than improvise.
Humans are not automatically better
Human ownership is not a claim that people are unbiased, consistent, or especially good at every judgment. They are not.
A human-only process can be careless, arbitrary, or inaccessible. AI assistance may expose inconsistencies, make explanations easier to understand, or help someone participate who would otherwise struggle. Refusing all assistance in consequential work can preserve an already poor process.
The requirement is accountable judgment, not unaided judgment.
It is also reasonable to automate bounded actions under an established policy. An organization need not ask a person to approve every routine appointment reminder. It should have a person responsible for the policy, the exceptions, and the consequences when the process fails.
The important question is whether the automation applies an authorized decision or quietly creates a new one. A reminder following an agreed schedule is different from deciding that someone no longer deserves service because they missed it.
More reliable AI may change which tasks are sensible to automate. Reliability alone does not decide who has the right to set the policy.
A practical closing
Before your next consequential AI-assisted task, name the person who owns the decision. State what the AI may prepare and what remains outside its authority.
Then ask that owner to describe the commitment in their own words, including who could be worse off and how the decision can be challenged.
If those questions feel disproportionate, narrow them to the actual consequence. If nobody can answer them, do not solve the gap by asking the model for a more confident recommendation.
You can delegate a great deal of work without delegating the obligation to understand what you are doing.