Every AI deployment turns vague values into executable instructions. If leaders do not define purpose, knowledge, and reality, the system will borrow its philosophy from metrics, training data, and vendors.
Imagine a leadership team announcing its AI strategy in one sentence: We will achieve the same growth with 15 percent fewer people.
That may be a target. It may even be a responsible response to difficult economics. But it is not a compelling account of what the organization hopes to become. It tells employees which cost should fall, not which capability should rise. It describes subtraction as though subtraction were a future.
Michael Schrage raised a version of this problem in our recent conversation. The trouble with treating AI chiefly as a source of efficiency is not only that the ambition is small. It is that the ambition supplies the system with a purpose. The tools are then asked to optimize labor cost, cycle time, or output without anyone having to say what the work is ultimately for.
AI does not make philosophy irrelevant. It makes philosophy operational.
Once a model can recommend, rank, route, generate, negotiate, or act, words such as value, quality, loyalty, productivity, and risk can no longer remain inspiring abstractions. They have to become instructions. Someone, somewhere, must decide what counts as evidence, which categories matter, which tradeoffs are acceptable, and whose outcomes deserve priority.
If leadership does not make those choices deliberately, the choices do not disappear. They are inherited from the data, the metric, the workflow, or the vendor.
The danger is not only that AI will fail. It is that AI will faithfully optimize a purpose no one examined.
Philosophy is already in production
In his MIT Sloan Management Review essay, “Philosophy Eats AI,” Schrage argues that AI systems inevitably embody positions on three old philosophical questions.
Teleology asks what the system is for. Is a customer-service agent meant to close tickets, reduce expense, restore confidence, preserve a relationship, or help a customer make a better decision? These purposes overlap until they do not. At the moment of conflict, the system needs a priority.
Epistemology asks what counts as knowledge. Which evidence may the system trust? Does a recorded click outweigh a customer’s explanation? Is a prediction sufficient grounds for action? How should the system treat uncertainty, disagreement, or information that does not fit its prior patterns?
Ontology asks what kinds of things exist in the system’s world. Is a person a customer, a user, an account, a risk score, a member of a household, or a participant in a community? Which relationships are visible? Which are erased because the database has no field for them?
These questions can sound academic until a system denies a loan, selects a job candidate, prioritizes a patient, assigns a sales lead, or decides which employee appears unproductive. Then ontology becomes a data model, epistemology becomes an evidence rule, and teleology becomes an objective function.
Philosophy has entered production.
The practical mistake is to imagine that a technically neutral system waits for human values to be added later. There is no neutral ranking, neutral threshold, or neutral definition of success. Every deployment encodes a view of what matters, even when that view arrives disguised as a default setting.
When metrics become metaphysics
Organizations need metrics. The problem begins when a proxy stops being treated as a partial representation and becomes the reality the system is permitted to see.
Suppose a company defines loyalty as repeat purchases. That definition is measurable and useful. It is also incomplete. A customer might repurchase because switching is difficult, because a contract traps them, or because every alternative is worse. The metric registers all three as loyalty. An AI trained to increase the number may improve retention while degrading the relationship that retention is meant to represent.
The same problem appears when knowledge work is reduced to visible activity. Messages sent, documents produced, tickets closed, or hours online are traces of work. They are not the work itself. Judgment often creates value by preventing unnecessary activity, reframing the question, or recognizing that the requested output should not exist.
AI makes this confusion more consequential because it can optimize the proxy continuously and at scale. A weak metric in a monthly report can mislead a meeting. A weak metric embedded in an autonomous workflow can reorganize behavior across the company.
When a metric becomes the target, it can become the organization’s working definition of reality.
This is why the familiar warning that “the map is not the territory” becomes a governance issue. The model will travel the map it is given. Leaders remain responsible for asking what the map excludes, who drew its borders, and whether arriving at the marked destination would actually count as progress.
The accumulation of philosophical debt
Technical debt accumulates when short-term engineering choices make future systems harder to change. Organizations now face a related problem: philosophical debt.
Philosophical debt accumulates when fundamental questions are left unresolved but systems must still be built. The team does not agree on what a qualified lead is, so it automates the current scoring rule. Leaders have not settled what good service means, so the agent optimizes handle time. No one can define productive collaboration, so the dashboard counts individual output.
Each choice is defensible as a temporary convenience. Together, they harden into an operating philosophy no one remembers choosing.
The debt is easy to miss because the system can perform well by its own measures. It can become faster, cheaper, and more consistent. But consistency is not the same as correctness. A system can scale a category error with extraordinary reliability.
This is one reason vendor selection alone cannot be an AI strategy. Vendors can provide capable models, safeguards, and implementation patterns. They cannot decide what your customers should become, what obligations you have to employees, or which forms of value your business should protect. When leaders leave those questions blank, product defaults fill the space.
Agency raises the philosophical stakes
The stakes rise as AI moves from answering questions to taking action.
A conversational model can produce a flawed recommendation that a person ignores. An agentic system can convert the same flaw into a sequence of actions before anyone notices the premise. It can segment customers, change offers, move money, contact suppliers, or alter a workflow. Greater autonomy shortens the distance between an assumption and its consequences.
This does not mean every decision needs executive review. That would defeat much of the value of automation. It means the boundary between machine discretion and human judgment must reflect the moral and strategic weight of the decision, not merely the model’s confidence.
A model may be highly confident that a customer belongs in a low-value segment. Confidence does not answer whether the category is fair, whether the treatment will become self-fulfilling, or whether the company wants to build a relationship in which past value sets the ceiling on future attention.
“Human in the loop” is therefore too vague to be a governance principle. Which human? At what point? With what information? Empowered to question which premise? Accountable to whom?
Expert judgment should not be a ceremonial inspection step attached to the end of an automated process. It must help define the categories, evidence, exceptions, and purposes that shape the process from the beginning.
Critical thinking is not a soft skill
Schrage asks a bracing question in the episode: How do you contribute meaningful value when you are no longer the smartest creature in the room?
One answer is that intelligence and judgment are not interchangeable.
A model may generate more options than a team, retrieve more information than an expert, and expose patterns no individual could detect. It still cannot relieve leaders of the obligation to decide what deserves to be optimized. In fact, greater machine capability makes that obligation harder to avoid.
Critical thinking in an AI-enabled organization is not simply the ability to spot hallucinations. It is the ability to examine the frame:
- What problem have we decided this is?
- What must be true for this recommendation to make sense?
- Which values have been converted into variables?
- What does the system make invisible?
- Which outcome would falsify our theory of value?
- Who benefits if the model is right, and who absorbs the cost if it is wrong?
These questions are sometimes dismissed as slowing innovation. Used well, they do the opposite. They prevent a company from scaling a misunderstanding.
Promptathons as organizational inquiry
One of Schrage’s most exciting practices is the promptathon, a collaborative setting in which people use generative AI to explore problems, alternatives, and opportunities together. The important word is collaborative.
Most workplace AI adoption begins as private experimentation. Individuals learn to write faster, summarize more, or automate pieces of their jobs. That can produce real gains, but it also keeps the organization’s assumptions hidden inside personal workflows.
A well-designed promptathon can surface those assumptions. Different functions can ask the same model to describe a valuable customer, a successful project, or an acceptable tradeoff. The disagreement among the answers is not noise. It is evidence that the organization has been using shared words without a shared meaning.
That makes the promptathon more than a prompt-engineering exercise. It can become a form of institutional philosophy conducted through prototypes. Teams can test how a definition behaves before embedding it in a system. They can compare what sales, service, finance, operations, and customers each mean by “value.” They can discover which disagreements require leadership decisions and which can remain productively plural.
The goal is not consensus on every abstraction. It is enough clarity to know which purpose governs a particular system, which evidence it may use, and when competing values require escalation.
A philosophy brief for every consequential AI system
AI governance often begins with inventories, risk levels, privacy reviews, and technical controls. Those are necessary. They do not answer the prior question of what the system should be trying to accomplish.
Before a consequential AI system enters production, its owners should be able to produce a short philosophy brief. It need not quote philosophers. It should answer seven practical questions:
- What is this system for? State the human or organizational outcome, not merely the task it performs.
- What evidence counts? Identify the data the system may trust, the uncertainty it must expose, and the claims it may not infer.
- What world does it recognize? Name the people, relationships, categories, and time horizons represented in the model, along with the important realities that remain outside it.
- Which tradeoffs require a person? Specify the conflicts the system may resolve and those that need accountable human judgment.
- Who should become more capable? Describe how customers, employees, or partners gain agency rather than merely becoming easier to manage.
- What counter-metric could reveal a false victory? Pair the primary objective with evidence that would show the system is winning the measure while losing the purpose.
- Who bears the downside? Make distribution visible. An average benefit can conceal concentrated harm.
The brief will not eliminate disagreement. Its value is that it makes disagreement inspectable before software turns it into routine.
The strongest AI strategy is not “What can this model do?” It is “What should this system make possible, for whom, and at what cost?”
From cost effective to cost affective
Near the end of our conversation, Schrage plays with the difference between effect and affect. Organizations want cost-effective AI, systems that achieve an outcome with fewer resources. He argues that they should also want “cost affective” AI, systems designed with human feeling, motivation, and identity in view.
The phrase is more than wordplay. Technology changes not only what people produce, but how they experience themselves while producing it.
An AI system can make an employee more capable, curious, and willing to exercise judgment. It can also make the same employee passive, monitored, and reluctant to think beyond the recommendation. A customer agent can help someone understand a difficult choice, or it can perfect the art of moving them through a funnel. Both systems may reduce cost. Only one enlarges the person.
This suggests a more ambitious way to measure return on AI. Alongside time saved, errors reduced, and revenue created, ask whether the system improves the human capacity around it. Do people make better judgments after using it? Do teams ask better questions? Can customers act with greater understanding? Does expertise become more transferable, or merely more concentrated in the machine?
Schrage has long argued that companies should ask not only how to create more valuable innovation, but how innovation can create more valuable people. That shift changes the unit of analysis. The person is no longer a cost adjacent to the technology. Human capability becomes one of the technology’s intended outputs.
The advantage is clarity
AI capability will continue to spread. Competitors will gain access to similar models, interfaces, and agents. The durable advantage will not be possession of the tool alone.
It will be the clarity to give powerful tools a purpose worth pursuing, the discipline to distinguish evidence from convenience, and the imagination to define value as more than whatever happens to be measurable.
Leaders sometimes hope AI will help them avoid hard choices by revealing the objectively best answer. More often, it reveals how much judgment was concealed inside the question. It forces a company to say what a good customer relationship is, what productive work looks like, which risks are acceptable, and what kind of organization it intends to become.
Those are philosophical questions, but they are not luxuries. They are design requirements.
AI cannot decide what your business is for. It can only make your answer more consequential.
This essay grows out of Eric Pratum’s conversation with Michael Schrage on The Unfolding Thought Podcast. Watch or listen to “Why Philosophy Will Matter More Than AI.”