Which Business Decisions Should Never Be Delegated to AI?

By Victor Hoang · Updated Oct 11, 2026

Most conversations about what to hand to AI start with capability. Can the model do it? Is it good enough yet? That's the wrong first question, because capability changes every quarter and the answer will keep moving. A better question is one that doesn't expire: if this decision is wrong, who answers for it, and can we take it back?

Our view is that the line between delegable and non-delegable decisions isn't task difficulty. It's accountability and reversibility. A hard analytical task with an easy undo is a fine thing to automate. A simple-looking decision that ends someone's job, commits the company to a contract, or can't be reversed once it's made should stay with a named human, even when the AI's recommendation is better than the average manager's. "Better on average" is a statement about the population of decisions. Accountability is always about one.

Key Facts

• Under GDPR Article 22, people have the right not to be subject to a decision based solely on automated processing that produces legal effects or similarly significantly affects them. Where narrow exceptions apply, the controller must still provide at least the right to obtain human intervention, to express a point of view, and to contest the decision.
• The EU AI Act's Annex III lists employment uses as high-risk, including recruitment or selection, decisions affecting terms of work relationships, promotion or termination, task allocation based on behavior or personal traits, and performance monitoring.
• Article 14 of the EU AI Act requires high-risk systems to be designed so the people overseeing them can decide not to use the system, and can ignore, override, or reverse its output. It also asks them to stay aware of the tendency to over-rely on what the system recommends (automation bias).

Why Difficulty Is the Wrong Test

Difficulty sounds like a sensible filter. Let AI handle the easy stuff, keep people on the hard stuff. In practice it fails in both directions.

Some hard tasks are perfect for automation. Reconciling ten thousand invoices against purchase orders is tedious and error-prone for humans, and a wrong match is caught and fixed in the next close. Forecasting demand across 4,000 SKUs is beyond any planner's patience. Hard, yes. But the cost of a wrong answer is small and recoverable, and someone can check it.

Some easy tasks should never be automated. Declining a refund is easy. So is flagging an employee as a flight risk, or approving a payment above a threshold. The mechanics are trivial. The consequences land on a person, a relationship, or the balance sheet, and they're hard to walk back.

So we'd drop the question of how hard the work is and use two others.

Accountability. When this goes wrong, which human is answerable to a customer, a regulator, a board, or a court? If you can't name one, you haven't delegated a decision, you've dissolved it. An AI system can't be fired, sued, or asked to explain itself to a worker in a meeting. The organization carries the responsibility either way. The only variable is whether someone with authority is actually exercising judgment on its behalf.

Reversibility. If the decision is wrong, how fast and how cheaply can you detect it and undo it? A mispriced quote that a manager reviews before it goes out is reversible. A termination, a denied claim, a public statement, a wire transfer, a signed contract: once they land, the reversal costs far more than the original decision, or isn't available at all.

The Four Categories That Stay Human

We'd put decisions in the non-delegable bucket when they meet any one of these tests.

1. Decisions about people's livelihoods

Hiring, firing, promotion, pay, discipline, scheduling that changes someone's income. Regulators have reached the same place from a different direction. The EU AI Act treats recruitment, promotion, termination, task allocation, and performance monitoring as high-risk uses. GDPR gives individuals a right not to be subject to solely automated decisions with significant effects on them.

Think about what the person on the other end needs: someone who can hear their side, change their mind, and put their name on the outcome. A model can screen, summarize, and surface patterns. A manager decides and owns it.

2. Irreversible commitments

Anything that can't be undone, or undone only at ruinous cost: signing a contract, wiring funds, closing a facility, publishing a statement, deleting records under a retention duty, shipping a product with a safety implication. The test is simple. If you wouldn't let a new hire do it unsupervised on day one, don't let a system do it unsupervised on day one, however fluent its reasoning sounds.

3. Decisions that set the rules

Policy, pricing architecture, risk appetite, what the company will and won't do, how much AI error the business can tolerate. These aren't decisions inside the system, they're decisions about the system. We cover that last one in how much AI error a business can afford. Letting AI choose its own boundaries is like asking a department to set its own budget and audit itself.

4. Decisions where the consequence has to be owned in public

Apologies to a customer after a failure. Layoff communication. A response to a regulator. A disclosure to investors. These need a human voice because the audience is judging accountability, not accuracy. A technically correct message from nobody reads as evasion.

A Decision Framework

Score any decision on four questions, then act on the combination.

Now add the reverse test. If a decision answers "no" to all four, it's a strong candidate for delegation, with sampling and monitoring. That covers a lot of ground: routing tickets, drafting internal summaries, reconciling data, scoring leads for a human rep to review, scheduling meetings. The goal isn't to slow AI down. It's to spend scarce human attention where it carries weight.

Two answers in the middle deserve a note. A decision can be reversible but high-volume, such as thousands of small credit adjustments. There, no individual approval is practical, so the human owns the policy, the thresholds, and the sampling. And a decision can be irreversible but low-stakes, like sending an internal email. Those can go automatic. It's the combination of significant, hard to reverse, and personal that draws the line.

Illustrative Example: One Workflow, Three Decisions

Consider a hypothetical 300-person company using AI in its hiring process. The numbers below are invented to show the logic, not to benchmark anything.

Notice that AI does most of the labor, and humans keep the decisions that matter. That's the shape we'd expect in most functions: the machine takes the volume, and the human keeps the verdicts.

When AI Recommends, the Human Has to Be Real

Keeping a human "in the loop" isn't enough if the human is a formality. The EU AI Act's oversight article spells out what real oversight looks like: the overseer must be able to decide not to use the system, disregard or override its output, and stop it, and must stay aware of the tendency to over-rely on the output. Treat that as a design standard even if you operate nowhere near the EU.

Practical tests for whether your oversight is real:

• Can the reviewer say no without penalty? If overriding the system takes a form, a justification, and a manager's sign-off, while accepting it takes one click, you've built a rubber stamp. We look at that failure in why human-in-the-loop turns into a rubber stamp.
• Does the reviewer have the time and the context? A person approving 200 recommendations an hour isn't deciding.
• Does the reviewer know the domain? Oversight is only as good as the person's ability to tell a good recommendation from a plausible one. The cost of that checking is the subject of the verification tax.
• Is the decision recorded with a named owner? When something goes wrong, "the system suggested it" can't be the answer. A name, a date, and a reason should be on file.

What About Agents?

Autonomous agents sharpen all of this, because they act rather than recommend. An agent that drafts an email is a tool. An agent that sends it, issues the refund, updates the contract, and closes the ticket has taken the decision. That's why the line has to be drawn at the action, not at the model. Give an agent permission to do the reversible, low-stakes steps, and require a human approval at any step that crosses one of the four tests above. We go deeper on this boundary in copilots versus agents and on the controls leaders should demand in AI agent security questions.

What to Do Next

For a CEO, COO, or general counsel who wants a clear position by the end of the quarter:

1. List your top 20 decisions that AI already touches. Include the informal ones: the spreadsheet a manager built with a chatbot, the screening step in recruiting, the refund bot. You can't draw a line around decisions you haven't found.
2. Run each through the four questions in the table. Mark each as delegable, delegable with sampling, or human-owned.
3. Name an owner for every human-owned decision. A person, not a committee, and not "the AI team." Put it in the process document.
4. Make overriding easy. Remove friction from saying no to the system. Track the override rate. A rate near zero isn't proof the AI is excellent, it may be proof nobody is looking.
5. Gate agent actions by reversibility. Allow autonomous action on reversible steps and require approval on the rest. This ties directly into the broader work of closing the governance gap.
6. Check your legal exposure. Where you hire, lend, insure, or monitor staff, the rules on automated decisions are already in force or arriving. Our piece on AI compliance that survives regulation changes is a good starting point for counsel.
7. Revisit the list every six months. Capabilities move. The accountability test doesn't, but some decisions that were irreversible may become cheap to undo, and some low-stakes ones may gain consequences as you scale.

The point isn't to distrust AI. It's to be clear about who is responsible when it's wrong, because it will be, sometimes. The companies that handle that well will move faster, since their people will know exactly which decisions they can hand over and which they can't.

Learn More

• The Governance Gap: What Leaders Get Wrong About AI at Work
• The Verification Tax: Why AI Time Savings Shrink Once Someone Checks the Work
• AI Copilots vs. AI Agents: Understanding the Difference Matters
• AI Compliance That Survives Regulation Changes
• AI Agent Security Questions for Leaders