When not to use AI: the work to keep human

A lawyer submitted a brief with six case citations. Every case was invented by AI. The citations looked completely real. The lawyer paid the price.

The question every CEO asks after they get comfortable with AI is the right one: where do I stop?

Not "is AI good or bad" (it is neither). Not "should I use it" (yes, on the right things). The question is: which classes of work stay human regardless of how capable the model gets?

The core principle is this: AI as input is fine. AI as the decider or unverified shipper is the line.

The model has no skin in the game. It cannot be sued, fired, or stood in front of your board. When a decision needs someone to own the outcome, that someone is human. Accountability cannot be delegated to a tool. Signing your name is the act AI cannot do.

The capability map for CEOs gives you a task-by-task breakdown of where AI is reliable and where it is not. This piece is the decision rule that sits above that map: not which tasks are technically difficult, but which whole categories of work require a human owner regardless of capability.

Here are the five.

Irreversible decisions that ship before you can verify them

Rule: if a confident-wrong answer reaches the outside world before you catch it, keep a human on it.

Regulatory filings. Contract terms. Financial disclosures. These are not cases where AI is necessarily bad at the task. They are cases where the cost of a single confident mistake is catastrophic and non-recoverable.

Courts have already sanctioned lawyers for briefs containing AI-fabricated case citations that looked completely real. The model was confident. The cases did not exist. The lawyer stood in front of the judge. The model did not.

This is a pattern, not an edge case. Because of how AI generates text, plausible-sounding output and accurate output feel identical in the moment. On reversible decisions where you can catch the error before it ships, that is manageable. On anything that files, signs, or discloses, you cannot afford to rely on catching it.

Use AI to draft, structure, and pressure-test. Keep a human to verify and sign.

Hiring, firing, and anything touching protected-class status

Rule: you are legally accountable for discriminatory outcomes whether or not an algorithm made the call. "The AI did it" is not a defense.

AI resume screening with no human review layer is the canonical trap. A Stanford study found that AI screening tools produced racially disparate outcomes. Newer laws, including Colorado's AI Act and the EU AI Act, require human oversight and reasonable care against algorithmic discrimination. These requirements exist because the accountability sits with you, the employer, not with the software vendor.

This does not mean AI is useless in hiring. It means the decision layer is human. Use AI to structure job descriptions, synthesize your own interview notes, or draft offer letters. The moment AI is making the call on who advances without a human check, you have handed off accountability you legally cannot hand off.

The same logic extends to performance reviews, compensation decisions, and terminations. If a decision touches protected-class status and goes wrong, you will explain it, not the model.

For a deeper grounding on the bias mechanics, see what is adverse impact.

High-context people calls

Rule: if the decision turns on history, trust, and things only you know about a specific person, keep it human.

Letting go of a long-tenured employee is not a task the model can do for you. Not because it lacks the words. Because it lacks the context. It does not know what that person has been through. It does not know the promise you made at the offsite. It does not know the conversation you had with their manager six weeks ago that changed your read on the situation.

The same applies to a delicate conversation with a co-founder, a fractured investor relationship, or a team member you are about to promote over peers who also expected it.

AI is genuinely useful for preparation: draft the talking points, pressure-test the logic, anticipate the objections. But the call itself, and the judgment call embedded in it, belongs to the human who has the full context and will be in the room.

Anything where you sign your name or stand in front of the result

Rule: signing is the act AI cannot do. The human who signs owns it.

Board letters. Investor updates. Public statements. These carry your name and your authority. The model can write a first draft you are proud of. The moment you sign it, you have certified its contents. You, not the model.

This matters most when the output will be acted on by someone who cannot verify it independently: your board, your investors, your employees reading a company-wide email. They trust the letter because you wrote it. If the model invented a number, a commitment, or a fact, and you signed it without checking, that is your number, your commitment, your fact.

Use AI to draft. Read everything before it leaves. The signatory is always human.

Novel strategic calls, and the risk of eroding your own judgment

Rule: use AI to pressure-test your thinking, not to replace it on questions that have no clean answer yet.

There is a category of decision that sits at the genuine frontier: a market call no one has made before, a pivot under conditions that have not existed, a bet on a person in a role that has never existed at your company. These are exactly the calls where AI sounds most confident and is least reliable. The model has seen every pattern that existed before its training cutoff. It has seen none of yours.

More importantly: research warns that over-reliance on AI can dull critical thinking over time. Agency decay is a real risk. The operator who outsources every hard call to the model stops building the judgment muscle that makes the call worth making.

Pressure-testing a decision with AI is one of the highest-value uses of the tool: feed it the context, ask for counter-arguments, ask where the logic breaks. That is the model doing what it is actually good at. Then you decide. That is you doing what only you can do.

What to do next

Map your current AI use against these five categories.

For each place where AI is already involved, ask: is this AI as input, or AI as decider? If a confident-wrong answer shipped without a human check, what would it cost you?

Most operators find one or two live exposures when they look honestly. Fix those first. Then keep going with everything else.

The goal is not to use AI less. It is to use it in the right shape. The five categories above are not limits on your leverage, they are the fence that keeps the leverage clean.