Working With Machines
By Dan Martin
Walking on the beach with my 6-year-old granddaughter, I decided it was time for a teaching moment.
“Do you remember the last time we were here? The ocean was way out there. Today it’s right beside us. Why do you think that happens?”
She thought for a moment. “We have to ask ChatGPT.”
In a few years our relationship with AI has changed dramatically. When it began, AI seemed something like a search engine on steroids: a machine that could understand and respond in complete sentences. It became a useful assistant at work and then a reasonably capable teammate. Now it’s beginning to act autonomously — hopefully on our behalf. Every advance in AI changes the relationship between people and machines and therefore changes what leadership must do. These stages also represent increasing levels of delegation: first asking AI for information, then assigning it a specific task, then entrusting it with substantive work and finally authorizing it to make decisions and act. As the delegated authority increases, so does a leader’s responsibility to define boundaries, set verification standards, and establish escalation rules while remaining responsible for the outcome.
AI as Information
When Large Language Models (LLMs) became available, they were used to get information. They were fast, clear, confident, and apparently able to understand and almost instantly deliver information — in complete sentences — from a wide variety of relevant sources. Students discovered their ability to answer questions, solve complex problems and expound at length on almost any topic. They didn’t notice that machine-generated solutions to math problems did not guarantee student understanding.
“He got every problem on the take-home test right, Dad,” my son, Darius Martin, a university economics professor, commented. “But when he did his oral presentation, he didn’t know whether the supply-and-demand curve went up or down.”
However, AI was, and still is, imperfect. An AI may know everything about plumbing. That doesn’t mean it knows your plumbing. As Rob Clyde told me:
“If a robot plumber had been controlled by AI, it might have stopped the rattling. It would have cost $60,000. The AI recommended that I get a plumber to determine where to rip open the wall and fix the clamp. I realized that the problem occurred only when the sprinklers came on – the large water flow caused the rattling. I installed a hose timer to run the hose for a few minutes before the sprinklers engaged. The AI had the same facts and knew about hose timers, but didn’t figure out the $15 solution.”
AI knew plumbing. But Rob knew his house.
Access to almost unlimited information won’t solve the problem if it can’t be connected to local context.
AI as a Tool
In short order, people began using AI as a specialized assistant. As screenwriter David Brownfield, who creates TV murder mysteries, said:
“My character needs a creative way to kill someone. AI provided 25.”
David used AI to answer a specific question. Before AI, he’d have brainstormed with the writing team, emailed a colleague or Googled “murder.” He may even have strolled into a library.
Medicine offers a more consequential example. Physician and former Chief Medical Officer Campion Quinn pointed out that “AI is not a thing. It is many things.” In medicine, one AI may read an EKG or analyze an image; another may summarize a medical record; another may flag a patient at risk of sepsis. Different tools do different jobs.
The question isn’t whether “AI” is good in medicine. AI can be a useful tool when used properly in appropriate areas. The human being has to know what AI can do and what it can’t do. The human, in the end, is responsible for the result.
AI is a tool — but so is a hammer. Not every problem is a nail, and not every issue can be solved by AI.
AI as Coworker
Understand the Machine and its Limitations
The use of AI continues to evolve. Now, AI is not just answering questions; it is doing work. Rob Clyde says AI now writes large portions of software code. His experience also captures the paradox: AI can be extraordinarily capable in narrow ways while missing things that are obvious to an experienced person looking at the larger problem. People should understand the strengths, limitations, and biases of AI. It may appear to know “everything,” but what it produces reflects its training, its sources, and choices made by the people who built and deployed it. What it delivers also depends on the user’s own specific instructions. As Roy Austin, attorney and Director of the Howard University School of Law’s Artificial Intelligence Initiative, emphasized in our conversations about responsible AI, human beings must determine what is best, what is right, and what is truthful.
For David, the screenwriter, AI is part of the writing team, coming up, among other things, with a seasonal arc for a TV series. Elsewhere, it drafts briefs and contracts, summarizes meetings, analyzes CVs, writes software and contributes ideas. It is no longer simply retrieving information. It is doing work that, in the past, would have been done by a person. And, just as leaders learn the foibles and strengths of people, they have to learn the capabilities and limitations of machines.
Verify What the Machine Produces
People need to monitor and review, and that is challenging because of what Campion Quinn calls “automation bias.” If the AI is right 99 times in a row, it is very tempting to stop checking the hundredth time. The better the machine becomes, paradoxically, the more difficult it will be to provide human oversight.
An AI can only evaluate the information available to it. Campion described an AI listening to a medical visit. A patient may say the pain is “not too bad,” and the AI will accurately record what was said. But unless the machine also receives and correctly interprets the relevant visual and behavioral signals, it may miss the grimace, guarded moment, and hesitation visible to the physician sitting across from the patient. The AI may capture the words, but the physician can detect what is meant. Experienced professionals often recognize what is missing, what does not fit, or what has not been said. The goal of the person working with the machine is calibrated trust: neither automatically accepting nor needlessly rechecking everything. The leader’s job has shifted from doing every task to deciding what needs independent verification and what can reasonably be trusted.
Judge Outcomes, Not Activity
When AI not only provides information to help a person work, but also does work, the role of the leader also expands. A team should deliver an outcome that meets or exceeds requirements and promises. The responsible human’s reputation depends on how well that happens. In 2023, lawyers in a federal case submitted a brief containing judicial decisions supplied by ChatGPT. The cases looked authentic, complete with names, citations, and quotations, but several had never existed. When the lawyers asked ChatGPT whether the cases were real, the system assured them that they were. The court sanctioned the lawyers, not the machine. AI had made the error, but the professionals remained responsible for verifying its work. People are responsible and must know the subject and the context.
Used correctly, AI can be incredibly productive and remarkably fast. In software, for example, the role of the developer has changed. The metric shouldn’t be “lines of code developed per day.” AI can obliterate that metric. And the sheer volume of code that AI produces overwhelms historical code review practices. Now, verification no longer means that a person has inspected every line of code. It means designing systems that can detect whether the result meets the required standards. A leader must be very clear about the requirements, the guardrails, and the standards while also being innovative about ways to assess and confirm quality.
A leader needs to define and measure results, not the tasks leading up to the result.
AI as an Autonomous Agent
An AI becomes an agent when it moves from recommending an action to taking one. Increasingly, these agents are being assigned tasks that require them to act quickly and with some degree of autonomy. One of the most visible examples is the self-driving car. When introduced, it was merely a brainier cruise control that worked best on the highway and required the human driver to keep their hands on the steering wheel. Later systems could maneuver through parking lots or navigate toward a destination, but they still required supervision. Now fully autonomous taxis are readily available in cities like San Francisco. Available, but not flawless:
“Why aren’t we moving?” My wife asked from inside the locked Waymo that was stuck at the intersection of Bay Street and the Embarcadero.
The car was flummoxed by pedestrians who were certain that they wouldn’t be nudged out of the way by an irritated robot. The car had not moved for several minutes. Six additional Waymo cars, waiting patiently, were trapped behind our vehicle. There was also a human-driven car directly behind our Waymo. It was not waiting patiently.
According to my wife, it took me “far too long” to convince Waymo to let us leave.
Now AI is working autonomously on things that happen so quickly they can’t rely on timely human oversight. In cybersecurity, AI agents constantly scan for attacks that occur in milliseconds and must sometimes intercept and disable them before a human can intervene. Cybersecurity expert Chris Hetner said it is almost like having to shoot first and aim later. Cyber defenses often must act before a human has time to verify the details.
Anything — person or machine — empowered and capable of making autonomous decisions must have clear, understood, and enforceable rules of engagement that define what it can do, when it must stop, escalate, or return control to a person. As Greg Godbout, CEO of Flamelit & AI for Natural Disasters, observed in one of our discussions, we would not give a new employee unlimited authority simply because we hired them. But in some cases, it is necessary to empower an AI — and that means the rules of engagement must be clear, comprehensive, and consistent. And we may have to live with the fact that these rules won’t be fail-safe — especially when dealing with machines. We must design the rules so that in the event of a miscue, the damage is limited. This challenge has been considered in the world of fiction. Unfortunately, it’s easier to implement Isaac Asimov’s Three Laws of Robotics in fiction.
Leading People and Machines
As the role of the machine expands, so does the responsibility of the leader.
| Machine Role | Leader’s Responsibility |
| Information Source | Supply context and question the answer |
| Specialized Tool | Choose the task and evaluate the result |
| Coworker | Define the work, calibrate trust and verify the outcome |
| Autonomous Agent |
Define authority, escalation rules, and limits on harm |
For centuries we learned how to lead people. During the past fifty years we learned how to lead organizations supported by increasingly sophisticated technology. Today’s challenge is different: we must learn to lead teams in which some capable members are not human.
AI strategist Bjorn Austraat described this division of responsibility succinctly: “AI may be good — even extremely good — at the how. People provide the what and, more importantly, the why.” Machines contribute speed, scale, and execution. People remain responsible for purpose, judgment, wisdom, and accountability. The challenge is no longer simply learning how to use machines. It is learning how to lead teams that include them.
About the Author
Dan Martin is the Program and Relationship Manager at Eleven Canterbury. He is the host of the video series, “Conversations With the Experts.” Each episode features senior leaders from Eleven Canterbury’s global network who share candid insights on timely issues, including the role of AI in business. Watch episodes featuring AI experts here.
Contributors
Roy Austin Jr. is an attorney and executive whose career has focused on civil rights, technology, litigation, corporate responsibility, and public policy. He directs the Howard University School of Law’s Artificial Intelligence Initiative and previously served as Meta’s Vice President of Civil Rights and Deputy General Counsel and as a Deputy Assistant to the President of the United States.
Bjorn Austraat advises organizations on enterprise AI strategy, implementation, and organizational change through Kinetic Cognition. His work focuses on helping companies move from broad AI ambitions to practical applications that produce measurable results.
David Brownfield is an award-winning writer, storyteller, and media executive with more than twenty years of experience developing and producing content across traditional and digital platforms. His work has also included branding, nonprofit communications, organizational change, and programs designed to broaden participation in the entertainment industry.
Rob Clyde is a technology executive, company founder, and experienced board director whose career has encompassed cybersecurity, artificial intelligence, cloud computing, high-performance computing, and software development. A former chief technology officer of Symantec, he has helped build and lead technology companies from early growth through major expansion and acquisition.
Greg Godbout is the CEO of Flamelit & AI for Natural Disasters and Chief Innovation Officer of Global Clean Energy, where he leads the company’s AI division. He previously served as Chief Technology Officer of the Environmental Protection Agency and co-founded and served as the first Executive Director of 18F, the federal government’s digital-services organization.
Christopher Hetner advises boards, executives and investors on cybersecurity, artificial intelligence and technology risk. Drawing on senior experience in government, financial services and critical infrastructure, he helps leadership teams translate technical threats into business, financial and governance decisions.
Darius Martin is an economist and university professor with teaching experience in the United States, Lebanon and Iceland. He holds a Ph.D. in economics from the University of California, Santa Barbara, and is an associate professor at Western Washington University.
Campion Quinn, M.D., is a physician executive, author, and strategist with more than two decades of experience in clinical medicine, medical communications, and the life-sciences industry. As CEO of Rockville Medical, he advises pharmaceutical, biotechnology, and medical-technology organizations on medical affairs, clinical strategy, evidence, regulatory alignment, and AI-assisted communications.
