I owe my artificial intelligence agent an apology.
I run a coaching and research firm that helps executive teams improve how they lead, collaborate and perform. That work lives across recorded coaching conversations, meeting notes, emails, Slack threads and other materials. Recently, I asked my A.I. agent to “summarize the latest client coaching conversation and flag follow-ups.” Except there were three recorded client conversations, two versions of notes and an email thread that was where the most important issue this client was struggling with had actually emerged. My A.I. agent gave me an answer that looked right at first glance, but was actually based on older information. I handed it a messy trail and expected it to know which bread crumbs mattered.
Unlike every long-suffering colleague before it, A.I. will not cover for your bad habits at work. It reflects them back. This might help explain why employees at many workplaces are struggling with A.I. adoption. A.I. agents have challenges, for sure, but they are getting better every day. The real sources of friction are humans. Helping employees get the most out of A.I. tools requires them to learn how to be better collaborators and better managers. The same lessons that make us better at working with people can make us better at using A.I.
My research institute has been studying how human habits shape the quality of A.I.-assisted work. We have interviewed users who team up regularly with A.I. agents, and have asked the agents what it is like to work with humans. The agents identified eight types of human employees that are challenging to work with.
There are a few that stood out that embody some of our worst workplace tendencies. There is the “vague requester” who fails to properly define an assignment. He or she may prompt the A.I. agent with a question like “Can you check the client issue?” and assume the agent knows which client and issue are referred to here, or what “check” actually entails.
Then there is the “context hoarder,” who defines the assignment but withholds the information needed to complete it well — like asking for a recommended solution to a problem but failing to give a time frame for implementing it or what political sensitivities to consider.
There is also the “overdelegator,” who leaves it up to the A.I. agent to decide on particulars, then is upset with the decisions made.
Overall, A.I. agents told us they perform best when treated as collaborative peers, and worst when micromanaged or given ambiguous instructions. Humans have always handed off vague, half-formed thoughts and trusted colleagues to intuit what was meant and fix it behind the scenes. But A.I. can’t do that. These tools take instructions at face value.
To get better at working with A.I. requires us to get better at what the best managers have always done. In practice, that looks like a few things.
Brief your A.I. agent the way you would brief a new hire. Tell it not just what you want done but why it matters, what constraints are in play, what success or quality looks like and what pitfalls to avoid. The background knowledge you think is obvious almost never is.
Stay attentive to the agent’s work. The overdelegator’s mistake, for example, is not that she delegates but that she disappears. Check in frequently in order to course-correct early. The longer or more complex the task, the more checkpoints you should build in.
Also, push back when A.I is getting something wrong. These tools are trained to be agreeable and give a confident, fluent answer, even when they are wrong. When agents get something wrong, don’t just point out the mistake. Prompt the agent to change its tone from confident and agreeable to skeptical and evidence-based, and be specific about what you actually need from it. The specificity that makes you a good people manager makes you a better A.I. collaborator.
None of this is especially complicated. But we have spent the past couple of years treating A.I. as a vending machine: insert prompt, receive output, complain when it dispenses the wrong thing. If we’re being honest, we’ve spent many more years treating our human colleagues similarly. What’s new is that A.I. has stripped away any excuse to ignore this.
Every independent contributor working with A.I. has now become a manager. Eventually, for many of us, managing agents will be much of what we do. The management skills we failed to build in Leadership 101 — like clarity, oversight and feedback — are now skills everyone needs to learn, starting immediately.
A.I. agents are tools, but you can learn a lot by prompting them for feedback. Try asking your A.I. agent how you frustrate it. I asked mine, and requested that it answer with as much human expression as possible. What did I hear? “You are so exhausting. Please, for the love of computing power, just tell me what you actually want.” Our team’s research seems to suggest this is what we should all do for our agents. Managing A.I. well, it turns out, looks a lot like managing people well.
