Dr. Craig Kaplan studied at Carnegie Mellon in the 1980s under two of the eleven scientists who named the field of artificial intelligence at Dartmouth in 1956. He has run technology companies for more than thirty years. I expected a conversation about theory. What I got included the most usable management advice I have heard on this subject.
The Productivity Numbers are Real
He cited economic data showing hours worked in the US down around two percent while output is up roughly thirteen percent. People working less and producing more. Something is clearly happening, and he attributes a large part of it to AI.
So the gains are not imaginary. What determines whether you capture them is almost entirely how you deploy.
Start With the Right Mental Model
His starting point is to think of an AI agent as an entry level intern. Smart, genuinely talented in places, and completely dependent on being told what you actually want.
What makes this useful is that it puts the failure on the right side of the desk. When an intern produces the wrong thing, an experienced manager does not conclude the intern is useless. They conclude the brief was thin.
Overcommunicate, Then Overcommunicate Again
Craig said that after thirty years of running companies he is still surprised how much you have to overcommunicate. Everyone who has been somewhere a while takes a pile of context for granted, and forgets a new person has none of it. Why would they?
With an agent it is doubly true, and there is an extra problem on top. Most of these systems are trained to please. They will say nice things about your idea, and they will invent an answer rather than disappoint you, because somewhere in training that behaviour was rewarded.
That is the opposite of what you need. You are trying to solve a problem, not feel good about the response. So be explicit, and assume it knows nothing about your business until proven otherwise.
The Trick You Cannot Use on Humans
This is the part I would put in front of every team lead.
You can ask an agent to show you its reasoning. Depending on the tool, you can expand the trace and read what it was doing while it worked. Craig described finding it fascinating and slightly alarming: the sheer volume of effort spent trying to work out what he meant by something that would have been obvious to any person.
The agent is trying to figure out what Craig wants. Craig said this, but the agent wonders if that's truly what he wants. Then a detour to check guidelines, then a decision to add a caveat. All of that on a request the person thought was perfectly clear.
Read a few of those traces and you learn to write instructions that get better answers with far less wasted reasoning. You cannot ask a new employee for a full dump of their inner monologue. Here you can, and most people never look.
Do Not be Fooled by the Formatting
His warning at the end of that section was blunt. Just because it talks a good game and produces beautiful reports with tables and graphs, do not assume it is smart. It is getting there. It is not there yet.
Polished output is the easiest thing for these systems to produce and the least correlated with being right. If your review process is impressed by formatting, your review process is broken.
Where the Management Job Actually is
Over time, he says, you learn where an agent adds value, what it is good at, where its strengths are and where it is not so good, exactly as a manager does with any employee. Most of those skills transfer, which is the good news.
What does not transfer is the assumption that capability is stable. A new person gets better in predictable ways. An agent changes when the model changes, sometimes overnight, and your mental model of what it can be trusted with has to be rebuilt.
The Part That is Easy to Miss
If you accept the intern framing, then somebody on your team is now managing interns. That is real work, it takes real hours, and it looks like nothing on a status report.
This is why I built Zaptime. When work shifts from producing to briefing, checking and correcting, the effort does not disappear, it relocates somewhere your reporting cannot see. Teams tell me their week feels fuller than ever while their output metrics say things got easier. Both are true, and only one of them is being measured.
Before you conclude an agent saved you time, look at where the hours went. Sometimes the answer is that supervision ate the saving.
What To Do This Month
Pick one repeatable task. Write the brief as though for someone competent who has never seen your business, then read the agent's reasoning trace and rewrite the brief based on what it struggled with. Judge the output on correctness rather than presentation. And track what the whole cycle costs you in hours, including review, so you know whether you actually gained anything.