It seems that every day (and often several times a day) we get bombarded by another article on AI. Some people are saying it may be the end of humanity and others are saying that it won’t deliver and that the “AI bubble” will pop. For all the discussion, there’s a need for a more thoughtful, balanced approach to the issues and benefits around AI.

A few weeks ago, Dr. Daven Morrison, a good friend of mine, suggested getting together several folks specifically with very diverse professional experiences to try to objectively and non-emotionally discuss artificial intelligence. I was fortunate to have the opportunity to spend time with:

-Daven Morrison M.D., an organizational psychiatrist and president of Morrison Associates LTD.

-Mitch Kentor M.D., MBA, an emergency physician and health systems leader

-Matt Pipke, JD, a tech entrepreneur and machine learning expert

After several meetings, we put together the first of a series of perspective pieces summarizing our thoughts and opinions based on our various backgrounds. In this series, we will discuss what agentic AI is, the humanistic side of it, and how values-based leaders can responsibly drive this change.

Leadership Choices at the Precipice of Agentic AI

Matt Pipke, JD, Mitchell N. Kentor, M.D., MBA, David E. “Daven” Morrison, M.D., Harry M. Jansen Kraemer, Jr.

AI technology underwent an enormous advance with the advent of Large Language Models (LLMs) that has the potential to massively impact our entire society. No one can afford to ignore it. Even if we set aside further progress toward Artificial General Intelligence (AGI), and consider only “Agentic AI” (that is, wrapping current LLMs in simple control systems, software tools and goal-directed agency — “orchestration”), its potential impact on human cognitive work may be greater than the impact on physical work by steam engines, hydraulics, and electric motors that replaced human and animal labor over the last two centuries. Thus, the implications for leadership and the management of the workforce are profound.

It’s all happening much faster than the revolution in mechanical power. Some fear this will lead to a disruption that takes away jobs and leads to mass unemployment. Yet others believe AI could usher in a new era of productivity and prosperity. The transition will be very challenging. The speed of it has left governance in its wake, with sound regulation and corporate best practices still in their infancy and failing to address the issues AI has created. In this gap, leaders of organizations have considerable influence over how civilization navigates this change. Leaders need to make the right decisions for both short-term goals and longer-term outcomes.

We believe leaders can make better choices armed with a comprehensive understanding of how the era of AI could change our society. In this multi-part series, we will explore whether AI should be viewed as a cost-cutting measure or a productivity tool; the challenges of AI change management and the emotions that go with it; and the ethics of the technology itself. Our goal is to help provide you as a leader with the tools and information needed to achieve optimal outcomes for both your organization and our society.

The authors’ industry experiences span an adoption spectrum: At one end, the “wild west” of the software development industry which is primed for unchecked AI adoption; and at the other end, the risk-averse and cautious industry of health care, which nonetheless needs a remedy for unsustainable demands on its workforce. In between: A variety of human cognitive work, previously impervious to automation, including finance, therapy, consulting, and business leadership. In this part, with Agentic AI as our foundation, we ask:

  • Will Agentic AI be a partner or a replacement?

  • What kind of work is most amenable to AI automation?

  • How might different industries be affected?

  • What do leaders of teams and organizations need to know about Agentic AI?

  • What are possible outcomes of the choices leaders make?

Notably this piece will not discuss AGI or its existential implications on AI safety and alignment. These issues will be discussed separately.

What is Agentic AI?

LLMs exhibit profound linguistic fluency, extensive knowledge capacity, and shockingly advanced conceptual comprehension, by leveraging humanity’s unique coding scheme for ideas about the world: Language. But it’s less clear how much reasoning they perform in isolation. Moreover, alone they are passive conversationalists. They don’t get work done autonomously — they lack agency.

To provide agency, these models can be embedded in software scaffolding that serves as a control system. The scaffolding provides software tools to interact with the digital world; memory to store new information beyond the initial “frozen” training of each model release; interfaces to control external systems; and guard rails and feedback loops to keep tasks on track. Notably, this can be done with ordinary non-AI software: Procedural code that can input linguistic prompts to the LLM, and pipe outputs to external systems or other LLMs. Teams of LLMs can be set up where each has a role (writer, evaluator, tester, re-writer). Through such “orchestration”, ensembles of these super-smart but passive LLMs, along with ordinary tooling, can transform into “agentic AI”, capable of performing autonomous cognitive work.

Kahneman’s System 1 / System 2 thinking (Thinking, Fast and Slow) provides a good conceptual basis for understanding this. LLMs are like System 1: Fast, intuitive, associative — like retrieving a fact, having a gut-instinct reaction, or finishing a sentence. The LLM System 1 is, however, vastly more knowledgeable than any individual human, having been trained with almost all known written material ever produced by humanity.

In contrast, an orchestrated agentic system of LLMs in roles, controlled by external software scaffolding, feedback loops, and armed with tools and interfaces, is more like the deliberative, procedural System 2 of Kahneman. It can perform cognitive work methodically, in a prescribed manner, and according to a framework to achieve goals.

Currently, orchestrations are being developed that are effective for actual work-in-production. There will be some tweaking to make them sufficiently reliable. Different tasks may require different orchestrations. But we argue it is inevitable: Agentic AI will be able to perform human cognitive work to completion with rapidly improving success rates, for increasingly more complex and open-ended tasks. This will likely happen within a year, if not in months. This advance can happen without any further fundamental breakthroughs in AI.

 

How will it work for us?

A slew of human cognitive activities are susceptible. One beachhead comprises tasks with clearly testable end-points that make it possible for AI to determine when it is done. Software development (and bug fixing) is a prime target, since AI can be interfaced to automated testing that has been used by human developers for decades. Agentic AI can digest written specs describing desired functionality, write code that meets the specs, test that the code is bug free and performs desired functions, fix bugs, iterate in an ever-improving loop until all tests pass, and generate the documentation.

Another category, given AI’s new mastery of language, is tasks involving communications with humans, like customer support. In health care, there is a huge burden in scheduling as well as post-procedure follow-up to capture complications or answer questions about medication regimen, all critical to successful outcomes. Here, orchestrations must be robust against mistakes and must properly escalate to human experts when appropriate.

Education, adaptive instruction, and testing should benefit enormously from Agentic AI, with its infinite patience and personalized attention to the needs of each pupil at any time. In conventional classrooms, the numbers are stacked against desired success: One or two teachers must teach 20-35 students the same content, at the same rate, and move on regardless of the receptiveness, developmental state, capacity to sustain attention, nutrition, or fatigue of each student. Agentic AI can take education far beyond this, and far beyond current individual use of LLMs for Q&A. Sufficient guard rails are needed to monitor and keep pupils on track within acceptable scope.

Surveillance duties could benefit from a tireless, vigilant, and intelligent agent that takes instructions in natural language form: Cybersecurity, physical plant security, patient monitoring in health care – wherever we currently rely on human eyes to watch and escalate. An important nuance is that pre-LLM AI technology is already adept in this area. The difference is that LLMs provide a natural language interface and a means for more flexible reasoning over the patterns of detections from these earlier AI surveillance technologies.

Analysis and draft content generation in financial, accounting, legal, clinical, and scientific domains can also benefit from Agentic AI. Consider the conventional investment banking team of a partner and 4 associates landing at a client to put together a fund-raising book, a consulting team brought in to review a project slate, or a law firm partner and junior associates working on briefs, motions and document discovery for a case. LLMs can digest reams of digital content and produce a cogent synopsis far faster than the junior associates of these teams. In healthcare, non-agentic LLMs already provide significant diagnostic leverage given the scale of medical literature (see AMA’s 2026 Physician Survey on Augmented Intelligence, Mar 2026). Orchestrated teams of LLMs may turbo-charge triage and capitalize on the influx of large amounts of data from remote patient monitoring.

Data collection through interviews with clients can reach a whole new level. Forget fixed surveys or questionnaires. Orchestrated LLMs will likely conduct reasonably interactive interviews that go in whatever direction may be relevant. Teams of LLMs can be orchestrated to cross-check output and avoid hallucinatory mistakes.

Buying, selling, shopping, ordering, trading (including financial instruments) – marketplaces of every kind are seeing agentic AIs establish a major footprint on behalf of humans and human-supervised organizations, going beyond conventional deterministic trading algorithms. Imagine eBay, Whatnot, Polymarket, Truckstop.com, Binance, and the NASDAQ as swarms of agentic AIs executing strategies against one another, with direct human traders in the fringes, as a very foreseeable near-term outcome.

These examples, while providing some color to what’s possible, still merely scratch the surface of the potential implications of AI. The leaders at the forefront of this workplace revolution need to intuitively understand what’s possible, in order to look for well-matched opportunities to improve their organizations.  Leaders also need to know that there’s a danger of deploying Agentic AI where it’s not needed: For example, using LLMs to automate SOPs that can be readily automated by non-AI systems.

 

Partner or Replacement?

Across most industries, Agentic AI offers the opportunity to reduce or eliminate the burden of menial cognitive work. For professionals advanced in their field and who need to offload or delegate cognitive work, this is a tremendous productivity boon, and Agentic AI is a partner. For organizations, like those in health care, with an enormous backlog of demand and a current operations approach that cannot meet that demand, the help is welcome and overdue.

For those who are relatively inexperienced in the workforce and starting their careers, Agentic AI is a competitor for the “grunt work” these employees are usually hired to do at low cost. Whether this means mass unemployment, or a different path to starting a career in an organization is the kind of choice that is in the hands of leaders today at this precipice of adoption.

New positions will emerge driven by the adoption wave. We may see a replication of something like the multitude of IT consultancies of the 1990s that then helped organizations first widely deploy computing and networking – this time with experts who will be able to help design orchestrated systems, select model vendors, run the numbers and customize Agentic AI solutions. Or we may see a cloud-era alternative to that, with a proliferation of subscription micro-service vendors for AI-executed cognitive tasks that you assemble to perform your own organization’s work. Whether a leader chooses to outsource or develop the skill internally is a choice with ramifications for both speed as well as competitive differentiation.

The challenge for leaders will be carving out enough time and focus to really understand what Agentic AI can do in your organization and how to deploy it successfully. There is a critical speed bump to adoption owing to the gap between tech employees who know how to mechanistically deploy AI but do not understand the target business processes and leadership who intuitively understands the business but may feel daunted by the “black box” of AI deployment. We suggest there is also an additional dimension: How to weigh short-term gains from AI deployment against long-term impacts on the business through the societal consequences of AI deployment.

 

Potential trends and consequences worth considering:

When viewed another way, the ability to deploy Agentic AI as a solution to menial cognitive work is an opportunity for your best employees to branch off and start their own businesses, potentially as competitors. There is a high probability of the development of powerful “micro firms” which leverage AI orchestration such that a small, lean team of founders can deliver an output traditionally only seen by large corporations. Retention of the organization’s best employees will require heightened attention. Larger organizations risk becoming dumping grounds for the least motivated, least mobile employees, as high performing employees see greater potential to take charge in their own firms with AI providing support. Leaders today need to get ahead of the question: Why would my most capable people want to stay here?

An oft-mentioned concern with replacing entry level workers with Agentic AI is the potential to develop an enormous skill gap in the workforce. In one camp: experienced workers who have already learned the ropes and now can orchestrate Agentic AI to give them tremendous leverage. In the other camp: Entry level workers who are denied the chance to work from the ground up and learn the details of their industry. This has societal implications, as well as implications for the workforce as experienced workers retire. Leaders need to consider their future workforce supply, and whether this is an opportunity to invest in the organization’s people to build loyalty as well as in-housed tradecraft. As an example, the accounting and law professions are currently reducing hiring as they wait to see what AI can do. This has the potential to create what has been described as an organizational diamond (not pyramid) that leaves “people development” vulnerable into the future.

Of course, the fundamental choice at this precipice is between the extremes of approaching Agentic AI as either an opportunity to lay off people and shrink budgets, while delivering the same work output as before, or a productivity tool to enhance the output of the organization in order to achieve more with the same people and budget. Historically, automation has favored the first choice, likely because such automation was primarily “brainless” deterministic mechanical or procedural steps, best suited for mature industries with cost pressures and tight margins – the market was not going to pay more for the item, so the only way to win was to reduce costs.

Now we are faced with the potential for cognitive automation. Industries with high cognitive work burdens typically have more complex problems to solve, more unsolved latent pains, and a backlog of demand. There’s no shortage of software to write, health care patient needs to take care of, products to improve, scientific research to perform, children to educate, and so on. The bottleneck in the past has been sufficient numbers of highly trained human workers. Agentic AI provides leverage to these workers to do more in the cognitive workspace. We suggest a case can be made that the strategic choice is closer to the second option above, with a view toward retaining and training this workforce in the use of Agentic AI as a partner. This is not merely an altruistic choice; we believe it is a wise competitive choice as well.

Some have predicted a slowdown in AI adoption. One thesis is that companies are not seeing a positive ROI on pilot projects employing LLMs. A related thesis is that the cost of Agentic AI for cognitive work is rising out of control and blowing budgets – making cognitive work potentially as expensive as when humans do it. Vendors charge for tokens used, and it is true that orchestrated agents cost a lot more: “thinking out loud” (in words), scanning the web for input, and iterating multiple times on outputs, results in massively more token processing compared to a simple Q&A – upwards of 50x more tokens would be typical for a software coding request.  However, we’ve seen this before with virtually all other technologies and we do not think this will be an obstacle. Prices drop with scale and widespread adoption. Optimizations can increase capability exponentially at equivalent costs. As an example, automobiles went from ~1-5 hp in the 1890s to 80-120hp by the 1930s (a mean 33x improvement) using the same fundamental technological paradigm (piston, cylinder, connecting rod, crankshaft, spark ignition and hydrocarbon fuel) while also becoming 3x more affordable. This will happen for Agentic AI as well, and in much less time. Pilot projects will yield to better orchestrations that reliably deliver positive ROIs. The challenge for leadership is not whether to adopt Agentic AI, but how to seek those optimizations.

 

Key takeaways:

  • Even without any further AI breakthroughs, current LLMs orchestrated into Agentic AI will have a major impact on human cognitive work and the workforce.

  • Leaders can’t ignore this revolution and have a key role to play in making informed decisions about Agentic AI to optimize outcomes for their organizations and society.

  • The persistent backlog of human cognitive labor demand may tilt decision-making in favor of investing in higher output and better quality work product, rather than only cost cutting.

 

Next

In the next parts of this series we will examine the impact on the workforce; develop a mental model to understand and anticipate what the changes will demand; examine what history has to teach us about the coming changes; share practical guidance on navigating your organization through Agentic AI adoption and what leaders should pay attention to; and suggest how they can lead through the change.

 

Authors’ Note: The authors did not use AI to generate the concepts or text in this article.

-Daven Morrison M.D., an organizational psychiatrist and president of Morrison Associates LTD.

-Mitch Kentor M.D., MBA, an emergency physician and health systems leader

-Matt Pipke, JD, a tech entrepreneur and machine learning expert