By: Bob Morse
In a post from a few months ago, I assessed AI against the framework of invention vs. Innovation. Invention is about creating new capabilities, whereas innovation turns those capabilities into tangible, commercial value.
Tremendous capital is flowing into the “invention” around AI, some $5 trillion at latest count in announced infrastructure to support the compute largely for frontier models advancing the boundaries of the possible. This over-investment in invention compares with a relative under-investment in innovation. How even should we think about the innovation phase, the applications that will be introduced into the real economy with a sustainable long-run profitable model?
This piece proposes a framework for the innovation phase of AI: applying Peter Drucker’s lessons from the rise of the knowledge worker to the age of AI.
As AI systems become trusted to make some decisions previously taken by human employees, they are best thought of as acting like a knowledge worker. The term “knowledge worker” was famously coined by management theorist Peter Drucker in 1959. Previously, corporations had large workforces with primarily manual skills; in the mid-20th century, corporations began building large labor forces with primarily thinking skills. These knowledge workers could not be managed with the same techniques used to manage manual workers. Drucker’s contribution was to define the knowledge worker and educate a generation about how to manage, measure and compensate folks who think for a living. Those frameworks for managing are still the backbone of organizational design today.
For AI to deliver on its promise as an innovation, it must move beyond chatbots of whom we ask questions. It will become systems which can actually make decisions. In common usage today, the term “agentic AI” means an AI system to whom you delegate a certain decision-making authority. (If the system cannot make decisions, it is merely a tool.)
To understand AI systems that we trust to make decisions, let’s look back at how Peter Drucker defined what makes a knowledge worker an Executive.
“Every knowledge worker in modern organization is an ‘executive’ if, by virtue of his position or knowledge, he is responsible for a contribution that materially affects the capacity of the organization to perform and to obtain results.” – Peter Drucker, The Effective Executive (1967)
Today the language on the future of AI applications uses words like co-pilot, agent, or AI assistant, which implies a smart but junior and subservient employee. That is the wrong model. As we delegate decisions that matter to AI, which must happen for AI to earn its keep over the long haul, then those systems are not interns or co-pilots. They are, by definition, AI Executives.
The form which innovation will take, making good the massive capital put into the invention phase, will be the rise of the AI Executive.
Thinking about your AI system as an executive (or perhaps many executives) is a bit of a scary concept at first. Let’s look a bit closer at this analogy and see how it holds up.
One defining characteristic of a knowledge worker is that typically they know much more about how to complete their task than their boss. The international transfer tax pricing expert knows much more about how to do that specific task than the CFO does. The CFO can’t tell the knowledge worker exactly how to do their job, they instead set the goals and outcomes, provide the support, and appropriately reward the results. Again, in the words of Drucker (Management: Tasks, Responsibilities, Practices 1973):
“The knowledge worker cannot be supervised closely or in detail. He can only be helped. But he must direct himself, and he must direct himself toward performance and contribution, that is, toward effectiveness.”
The analogy to LLMs is strong. No one knows exactly how LLMs reach the decisions they reach. We as the ‘boss’ of LLM can only direct them towards effectiveness. Classic SaaS workflow software functions more like the manual skill worker, where the designer of the software knows how each action is taken and defines each branch on a decision tree. So far, so good.
Now let’s look briefly at how Drucker measured the knowledge worker (from The Effective Executive, again) and evaluate the analogy to the AI Executive.
“For the manual worker, we need only efficiency, that is, the ability to do things right rather than the ability to get the right things done. The knowledge worker is certainly not defined by quantity. Neither is he defined by his costs. He is defined by his results.”
With knowledge work, we measure, value and pay for the output. If the knowledge work is the design of a new fashionable sneaker, we want the best design. Whether the design was produced by one genius in the basement over a long weekend, or a team of 200 in a global product management hierarchy, we care about the best design, not how much labor went into it. In shorthand:
Physical labor >> measure by time worked
Knowledge work >> measure by output
The read-across to the shift from classic SaaS to Agentic AI, then, will be:
SaaS >> pay per user per month
Agentic AI >> pay for outcomes
This change is revolutionary, not evolutionary, for the investment community which has been investing in SaaS software. Considering just the sphere I know best, the private equity industry is by my count now leveraged long more than $1 trillion in the seat-license model. For a SaaS company with little customer churn and the ability to raise prices, the so-called net retention for Annual Recurring Revenue typically exceeds 100%. That is for an annual cohort of signed customers, the price rises on renewing customers offset the departing customers, so that annual vintage of customers in total has revenue that is flat to slightly growing over time. And because it is recurring and has high margin, the lending community and over time equity investing community began to think of those as bond-like streams of payments, against which loans are advanced and valuations are set.
Shifting from ARR to outcomes-based pricing is terrifying to some in that it pulls away the struts supporting this idea they are financing a bond-like stream of payments. If it is variable, even if in direction it is growing, that is much less popular to lend against or place a valuation multiple on. Moving from ARR to outcomes based pricing will upend the foundation on which so much of the SaaS investment community has relied.
Let’s wrap this up with an example. We are indeed now seeing users for some of our AI software platforms beginning to feel comfortable delegating some decisions to the machine. For instance, Netstock provides inventory management software to mid-size companies. Over the past few years, we have gone from “just” really great inventory management analytics to a conversational AI interface that suggests the order to change or cancel. Users have grown to trust those recommendations. Indeed, in the Netstock annual survey, 24% of users now say they would be comfortable completely delegating inventory decisions to Netstock AI, and another 50% say they would be willing to partially delegate, with oversight or shared control. This is a recognition that the Netstock AI engine has gotten reliably very good at the game of optimizing inventory, and so here some version of “human + machine” seems a likely path of agentic AI adoption. Those employees can now turn their attention to other initiatives at their companies.
How do you pay for an AI system that is managing your inventory? It’s not per user per month, I think we can all agree on that, as there are no longer users in the traditional sense. The answer is that the value is defined by the improvements in fill rate and reduced inventory carrying costs, hard data economic measures, which the customer can see and pay for on the basis of value received, not the number of their employees using a software package. In considering pay, we must acknowledge some fundamental differences in the risk-bearing ability of human executives, who have mortgages and risk preferences, as compared to AI programs, which do not.
These are the great opportunities of our day: to not just create but to manage AI Executives so that they in fact produce useful outputs in the world, and to price these AI Executives in a new way, graduating beyond the user-seat-license model.
It is possible. At Netstock, we took a great ‘classic’ SaaS company and introduced an AI agent that users grew to trust and rely upon. Today three-quarters of the clients are willing to delegate their inventory management decisions to the Netstock system (and some to delegate completely). Build software which clients trust with their decisions. No longer a system of record, your software will become a system of agency. That is a huge opportunity and the one we are spending all our time and capital pursuing.
Written By: Bob Morse
Co-Founder & Managing Partner, Strattam Capital
December 2025