The Agency Gap®: When Having the Answer Stops Being the Same as Making the Decision
There is a question I think we need to start asking about AI.
Not:
Is there a human in the loop?
But:
Is the human in the loop still thinking?
Because putting a person somewhere in a process does not necessarily mean they are exercising judgment, making a decision or taking responsibility for what happens next.
They may simply be approving what the machine has already decided.
And increasingly, the evidence suggests that distinction matters.
I call the space between having the capacity, information and authority to act, and actually exercising that capacity, the Agency Gap®.
It isn’t a problem created by AI.
I’ve seen it throughout my career.
People know something needs to change, but don’t change it.
They have enough information to make a decision, but ask for another report.
They have the authority to act, but wait for someone else.
They know a conversation needs to happen, but avoid it.
They can see an opportunity, but don’t move towards it.
Knowing is not agency.
Agency is what happens when we exercise judgment and turn what we know into purposeful action.
And that is becoming considerably more important in an AI world.
AI changes the conditions
AI gives us access to extraordinary amounts of intelligence.
It can analyse, compare, summarise, recommend, model, challenge and increasingly act.
That can make humans considerably more capable.
But there is another possibility.
The better AI becomes, the easier it becomes for us to stop doing some of the thinking ourselves.
New IBM research makes the problem unusually visible.
Its 2026 global CHRO study surveyed 1,500 CHROs and senior executives responsible for workforce strategy and 8,800 employees.
71% of CHROs said the ability to supervise, validate and override AI outputs is the workforce’s most essential capability.
Yet only 29% of employees ranked judgment as important.
Meanwhile, 60% of employees worry that AI is eroding their skills, with critical thinking the capability most commonly identified as declining. (IBM)
That’s quite a gap.
The people designing organisations around AI increasingly recognise that humans need to be able to challenge it.
The humans using it don’t necessarily recognise judgment as something they need to practise.
The danger isn’t necessarily a bad AI answer
Sometimes AI will be wrong.
But I’m increasingly interested in what happens when it is usually right.
Highly reliable systems can change human behaviour.
Research into meaningful human oversight warns that people supervising reliable automated systems can become complacent, over-trust plausible outputs and gradually move from meaningful oversight towards something closer to rubber-stamping.
Researchers therefore distinguish between AI’s operative agency, its ability to produce a solution, and human evaluative agency: our capacity to understand, assess, contest and, when necessary, override it. (Springer)
I think that distinction matters enormously for leadership.
Because the future relationship shouldn’t be:
AI thinks. Human approves.
Nor should it be:
AI recommends. Human automatically disagrees to prove they’re still in charge.
Both are poor judgment.
The capability we need is much more sophisticated.
- When should I trust this?
- When should I test it?
- What doesn’t it know?
- What context might be missing?
- Whose perspective isn’t represented?
- When should I seek another intelligence?
- And when should I override it?
That is human agency.
This is not an argument against AI
Quite the opposite.
I’ve spent the last few years exploring what becomes possible when human and artificial intelligence work well together.
I don’t want humans doing work machines can do considerably better.
And the research doesn’t support a simplistic story in which humans always blindly follow machines.
A 2026 study involving 2,015 military personnel, for example, found significant algorithm aversion in high-stakes scenarios rather than the expected automation bias. (Sage Journals)
Other research shows human and AI judgments can complement one another precisely because their biases and strengths differ. (ScienceDirect)
So the goal isn’t maximum human intervention.
It’s better allocation of intelligence and better human judgment.
AI should exercise the capabilities at which it excels.
Humans should not waste their intelligence reproducing work the machine has already done brilliantly.
But nor should we confuse delegation with abdication.
The Agency Gap® in practice
Imagine a senior leader receiving an AI-generated recommendation on an investment.
It is beautifully argued.
The data looks convincing.
The assumptions appear reasonable.
The conclusion broadly aligns with what the leader already suspected.
The easiest response is:
“That makes sense. Proceed.”
But agency requires something more.
Not endless scepticism.
Not another committee.
Not paralysis disguised as governance.
It requires the leader to remain cognitively present.
What would have to be true for this recommendation to be wrong?
What has the model privileged?
What hasn’t it seen?
What do I know about our people, customers, organisation or environment that isn’t represented here?
What does someone who disagrees with this see?
And perhaps the hardest question:
What do I think?
That final question may become one of the most important leadership questions of the AI era.
We need to practise agency
Judgment isn’t something we possess permanently because we once became experienced.
It needs exercise.
If AI increasingly drafts the first answer, performs the analysis, proposes the strategy and recommends the decision, leaders will need to become much more deliberate about maintaining their own capacity to think.
That doesn’t mean refusing AI.
It means changing how we work with it.
One useful practice is what I call Dynamic Deliberation®.
Before accepting an important AI-supported recommendation, deliberately move between different positions.
1. Think before you ask
On important questions, establish your own initial view before asking AI.
It doesn’t need to be complete.
But don’t always let the machine create the first cognitive anchor.
2. Ask AI to challenge, not merely confirm
Don’t just ask:
“Is this a good idea?”
Ask:
“Make the strongest case against this.”
“What assumptions am I making?”
“What evidence would change this conclusion?”
“What am I likely to be missing?”
AI can be an extraordinary thinking partner when we stop using it merely as an answer machine.
3. Bring in another human intelligence
Some decisions need lived experience, moral judgment, organisational memory, dissent, empathy or expertise that isn’t contained in the model or in you.
The fact that AI enables us to make a decision alone doesn’t mean we should.
Cognitive diversity is not inefficiency.
Sometimes another mind is exactly the friction better thinking requires.
4. Notice your own deference
Pay attention to moments when you think:
“Well, AI says…”
That sentence is worth interrogating.
Are you using AI as evidence?
Or as authority?
There is an important difference.
5. Make the judgment explicit
Before acting, articulate:
What do I believe we should do, and why?
If you cannot explain the decision without hiding behind the AI recommendation, you may not yet have exercised meaningful judgment.
6. Retain the Human Override®
Human Override® doesn’t mean humans should always overrule machines.
It means preserving the capability, confidence and authority to do so when human judgment says the situation requires it.
Accountability without meaningful ability to challenge is not genuine human control.
7. Learn from the outcome
After important decisions, don’t only ask whether AI was right.
Ask:
Where was AI stronger than us?
Where did human context improve the decision?
What did we miss?
When did we defer too easily?
When were we too sceptical?
What should we do differently next time?
That is how judgment develops rather than atrophies.
This may become one of the defining leadership capabilities
AI will become more capable.
That is precisely why human agency matters more, not less.
The scarce capability may no longer be possessing the information.
It may be knowing what to do with extraordinary amounts of intelligence when it becomes available.
For leaders, that means learning to integrate different forms of intelligence, specialist expertise, lived experience, data, diverse human perspectives, intuition and artificial intelligence, without surrendering judgment to any one of them.
I’ve increasingly been thinking of this as Integrative Leadership.
The leader becomes less the person expected to possess every answer and more the Conductor: bringing the right intelligence to the right problem, knowing which instrument should lead, listening for what is missing and ultimately taking responsibility for the whole.
Because there is a profound difference between having intelligence available and exercising intelligent leadership.
And perhaps that is the paradox of this moment.
We are rapidly solving the problem of access to intelligence.
We may simultaneously be creating a new problem around our willingness to exercise our own.
The more intelligence we have access to, the more important human agency becomes.
So perhaps every leader should occasionally ask themselves:
Am I using AI to expand my capacity to think and act, or am I slowly handing that capacity away?
The answer may matter considerably more than how good we become at prompting it.
Grow Bold. Think Big. Stay Human.
Carol Glover, The Maverick Maker®