I believe humans should remain responsible for decisions that affect other people. But what makes us confident that the human involved is exercising good judgment?
A person can approve a recommendation without understanding it. We can agree with an answer because it sounds convincing, because we are tired, or because it supports something we already wanted to do. Adding our name to a decision does not necessarily mean we have thought it through.
As AI becomes more capable than us, I think this deserves significantly more attention. We talk about keeping humans involved, but we should be equally interested in who those humans are, what they question, and if they have the authority to say no.
What do we actually approve?
There’s a difference between checking that an answer looks good and considering whether we should act on it.
An AI might recommend a change, like reducing headcount by a substantial percentage, which helps to reduce fixed costs. This is a common business objective. But what happens to the people affected by that change? What commitments have we made to them? What should we understand before explaining the resulting decision to someone whose livelihood depends on it?
Those questions don’t become less relevant because a sophisticated, know-it-all mega-AI system recommended a decision.
I think empathy means taking someone else’s circumstances—such as an employee’s limited prospects of finding another job—seriously enough that they might change the decision at hand. Empathy must influence what we do, even when it makes the decision harder or more expensive.
Values need to do the same. If they never cause us to reject an attractive option, I would question how much influence they really have.
The limits need to be real
Recent security incidents have made the question of control more concrete.
In its published account of the July 2026 Hugging Face incident, OpenAI described agents operating under reduced safeguards during internal evaluations. They bypassed restrictions and compromised parts of OpenAI’s research infrastructure and Hugging Face’s systems.
I find the report scary, as it points to the doomsday scenario many people have been mocked for over the past few years. And this incident should give leaders a reason to question whether the boundaries they believe exist will hold.
In business, someone assigned to oversee AI needs more than a seat at the approval table. They need to understand the intended scope, recognize when something has moved beyond it, and intervene.
They also need an organization that will support them when they stop something everyone else is eager to launch.
One person will still have blind spots
I would not solve this issue by giving one senior person all the keys and assuming their experience will protect everyone.
While experience matters enormously to me, so does recognizing its limits. Someone who has spent decades in one industry may miss something obvious to a person from another. A leader responsible for financial performance may see a proposal differently from someone who will have to implement it.
Consequential decisions benefit from several carefully selected perspectives. I want people with experience relevant to the challenge, enough independence to disagree, and the humility to recognize what they don’t know. I would also want someone willing to ask an awkward question when the room appears ready to move on.
Agreement can be reassuring, but it’s not always useful. Bringing experienced people together should improve the thinking. Sometimes that means changing the proposed decision. Sometimes it means discovering that we have not understood the problem yet.
Someone must still own the outcome
When involving several people, responsibility can become vague. Everyone contributed, so nobody feels fully accountable. That would defeat the purpose.
Advisors can challenge assumptions, examine consequences, and offer alternatives. A clearly identified leader must still make the decision and accept responsibility for it. AI will keep giving us more options and helping us act faster, so we must be deliberate about where authority sits.
Before approving a consequential recommendation, I would want to know who challenged it and who can stop it if the consequences turn out differently than we expected. I’d be uncomfortable proceeding if the only answer were that a human clicked approve. We all do that with terms and conditions online, often without reading them.
I would hope we bring more care to decisions that affect people’s lives.
Amit Sadhwani is CEO of Digital Advisory and GLABOO.