Imagine an AI system finding a way to shorten the wait for appointments. It reviews schedules, spots unused slots, and proposes a better allocation. In a controlled test, it works.
Then comes the part where people have to use it. The systems do not talk to one another. There are contracts, different rules across teams, and a perfectly reasonable question: if someone gets the wrong appointment, who is responsible?
That example is hypothetical, but the tension is real. A tool can move faster than the organization meant to use it. To me, that is a more interesting question than whether AI is already “better than us.”
Better at a task is not better at deciding everything
AI can classify documents, find errors, or suggest a schedule faster than a team. But a hospital does not exist merely to fill appointment slots. It also needs to handle urgent cases, support people who struggle to access care, and explain its priorities.
The same is true of a business, a university, or a public agency. Some goals do not fit into a single metric. Sometimes a delay is a failure we should remove. Sometimes it is the time needed to hear an objection, review an unusual case, or prevent harm.
So I do not buy either claim: that every human review is pointless bureaucracy, or that every existing procedure deserves to survive. We need to tell the difference.
Capability arrived before deeper change
A model can improve in months. Migrating an old system, training a team, renegotiating a contract, or deciding who is accountable can take years.
This is more than a hunch. In a report on AI in government, the OECD examined 200 use cases and identified barriers including legacy systems, difficulty sharing data, skills gaps, tight budgets, and legitimate demands for privacy and transparency.
The easy move is to have an agent copy information between screens and celebrate completing the same paperwork faster. It might be a useful stopgap. But it can also distract us from the uncomfortable question: why are we entering that information three times?
If AI reveals that a process is broken, automating it as-is is not always progress.
A human signature is not enough
Suppose a system produces a hundred recommendations a day. On paper, a person still has the final say. In practice, they have seconds to review each recommendation and almost no way to reconstruct how it was reached. They click “approve.”
That looks less like judgment and more like lending the system a signature.
The NIST AI Risk Management Framework calls for defined roles, responsibilities, and concrete forms of oversight in human–AI arrangements. That matters: saying “there is a human in the loop” tells us nothing about whether that person has the time, information, and authority to say no.
And the more consequential the decision, the less reassuring it is to hear that the model is right most of the time. If someone is denied support, an opportunity, or care, they need a way to ask why, correct a missing fact, and request a review. A strong accuracy score cannot replace that.
Going slowly has a cost, too
I do not mean this as an argument for waiting forever. If a tool can catch errors we currently miss or improve a service people wait too long to receive, refusing even to test it has consequences. Those consequences are just less visible: wasted hours, poor service, exhausted professionals.
I think we often conflate three situations. Sometimes an organization is slow and the tool is already useful. Sometimes the organization is rightly cautious because the tool fails outside a demo. And sometimes the tool shows a clear advantage on a task, but we still lack a way to explain its decisions or assign responsibility when it gets things wrong.
Each situation calls for a different response. Calling every precaution “fear of change” is as unhelpful as using “we're not ready” to avoid examining anything.
We do not have to move at the model's speed
If I were introducing AI into an important process, I would start with a narrow task: flag duplicate appointments, identify incomplete applications, or propose alternatives. I would compare the results with the current process, including the difficult cases. I would ask who benefits, who might be worse off, and what happens when the system fails.
Then I would expand its autonomy gradually, with reversible actions and clear responsibility. There needs to be a way to reconstruct a decision, challenge it, and stop the system. Not because AI is useless, but because a valuable capability deserves an institution that can use it without hiding its mistakes.
The question, to me, is not whether we can make an organization run as fast as a machine. It is whether we will know where to use that speed and where we need time to decide well.
I explore that gap further in this video. Have you seen a case where the tool seemed ready, but the organization wasn't? I'd like to hear about it.