The Dive
I keep hearing leaders talk about AI as a tool employees use.
I’m not sure that description holds up anymore.
Anthropic recently reported that Claude now leads 26% of its AI research and development work. In Anthropic’s definition, that means a human gives Claude a high-level objective and the AI completes most of the task from beginning to end, with the human supervising.
More than 90% of Anthropic’s AI research now involves Claude at least as a collaborator.
What Anthropic is not saying is that Claude is operating on its own. Humans are still involved. Humans still review the work. Humans are still accountable.
But that’s exactly what I find interesting.
Because if I give you an objective, you determine how to accomplish it, you work through the problems, and you bring me something that’s close to completed.
Which one of us managed the work?
That’s not a philosophical AI question; it’s a management question.
And I don't think most organizations have begun to answer it.
“Human in the loop” isn't enough.
We’ve gotten very comfortable with the phrase human in the loop.
It sounds responsible.
But I want to know where the human is in the loop.
Did the employee define the problem?
Did they decide what a good answer would look like?
Did they challenge assumptions along the way?
Could they spot the thing the AI missed?
Could they explain why the answer is right?
Or did they receive a polished result, read it, think, this looks good to me, and approve it?
Those are radically different kinds of human involvement.
And putting somebody's name on the final product does not magically make them accountable in any meaningful sense, does it?
If I am responsible for an answer I don't understand well enough to challenge, I don't have accountability. I have liability.
This is where the people question gets really interesting.
I think the bigger issue might be what happens to judgment.
We develop judgment by doing things before we supervise other people doing them.
An HR leader learns by working through difficult employee situations.
A manager learns by making operating decisions, getting some of them wrong and understanding why.
An engineer develops judgment by solving problems that don't have tidy answers.
That's the work we eventually learn to review, challenge, and delegate.
Now imagine removing a significant amount of that developmental work.
The productivity numbers might look fantastic: People move faster; output increases; teams accomplish more with fewer resources.
But what happens three years from now when the organization has a lot of people who are very good at approving AI-generated work and not enough people who know what producing quality work actually requires?
That's not an argument against AI. It's an argument for being much more intentional about what we automate and what we still need people to learn by thinking through and doing the work themselves.
You can't remove the work that creates expertise and simply assume the expertise will still show up later.
This is the part I think deserves much more attention than it’s getting.
When AI takes over pieces of knowledge work, we aren't just changing productivity. We're changing who (or what) has the initiative.
Who frames the problem?
Who decides what to try?
Who makes the intermediate decisions?
Who determines when the work is done?
Who knows enough to say, No. That's not right. Start over!
Those used to be pretty good descriptions of managing work.
Now, increasingly, an AI system may be making some of those decisions.
That doesn't mean the human disappears.
It means the human role has to be better defined.
And I don't think “review the output” is good enough.
The employee who reviews the work needs enough context, judgment, and expertise to challenge the output. Otherwise we're creating an accountability structure in which the human carries the responsibility while the machine controls the process.
That's a dangerous mismatch. Not because the AI is “in charge.” Because nobody has decided what being in charge means anymore.
The Save
If your organization is moving quickly on AI, don't just ask where you can automate work. Pick a few important workflows and ask:
Who defines the problem?
Is the employee deciding what needs to be solved, or simply passing an objective to AI?Who chooses the approach?
Does the employee understand why the work is being done this way?What does the employee have to know before approving the answer?
“That looks right” is not good enough.What would make the employee reject the AI's recommendation?
If nobody can answer that question, you may not have sufficient human oversight.What capability is this person developing?
This is especially true for early-career employees: If AI is now doing the work people used to learn from, where will that learning happen?
And one more:
Who could do this work if the AI couldn't?
That question might tell you more about the resilience of your organization than any AI adoption metric on your dashboard.
One thing to try this week
Take one workflow where AI is already doing significant work.
Don't ask, “Is there a human in the loop?”
Draw the loop.
Write down who defines the problem, who chooses the approach, who executes, who evaluates, and who has the authority to reject the result.
Then look at what the human is actually doing.
You may discover that you've automated a task.
You may discover that you've redesigned a job.
Or you may discover that you've quietly transferred much more authority than anyone realized.
Every company using AI is making these choices right now.
The question is whether we're making them deliberately.
Where has AI moved beyond assisting people in your organization and started determining how the work gets done?
Reply and tell me. I read every one.
Talk soon,
Anita
This is Issue 11 of Live Wire, a weekly jolt of candid thinking for CEOs, founders, and people leaders who know culture isn't a side project, it's how a business performs. I'm building this alongside my upcoming book: Everyone Sells. If this issue landed, forward it to a leader deciding where AI should stop and human judgment should begin.