The Hidden Architecture of Better Decisions: Why Cost Savings Follow Systems, Not Willpower

The real problem is not persuasion, it is decision architecture

What if the biggest cost savings in health care do not come from getting people to care more about price, but from changing the shape of the decision before anyone feels like they are making a tradeoff?

That is the deeper lesson hiding inside two seemingly different worlds: physician referral behavior and AI agents. In one, clinicians were nudged toward lower-cost imaging and procedure sites through goals, incentives, coaching, and feedback. In the other, intelligent agents are built from modules that perceive, plan, remember, reason, act, communicate, and learn. Put them together and a striking idea emerges: high-performing behavior is rarely the product of a single smart choice. It is the result of a well-designed system that makes the smart choice easier, more visible, and more repeatable than the dumb one.

That matters because most organizations still think about performance as a moral issue. We assume good decisions happen when the right person is sufficiently informed, motivated, and conscientious. But the evidence points somewhere else. People and organizations usually do not fail because they lack values. They fail because the environment, feedback loops, memory, and incentives are misaligned. In other words, the problem is less about character than architecture.

The health care example is especially revealing. Moderate changes in referral behavior translated into large cost savings, but only in some service types. That unevenness is not a disappointing footnote. It is the whole story. It shows that decision systems are not generic. They have friction, path dependence, and different degrees of “stickiness.” The same intervention can work well where decisions are frequent, standardized, and geographically flexible, but barely move behavior where relationships are personal or habits are deeply embedded.

That is exactly how AI agents are designed to think about the world. They do not simply react to input. They perceive, decompose, remember, reason, act, and adapt. The parallel is not superficial. It suggests that human institutions, especially in health care, could be redesigned with the same discipline we already use when building intelligent software.

Why some decisions change and others stay stuck

The most useful insight from the referral intervention is not that incentives work. Everyone already knows that. The useful insight is that incentives work only when they are attached to a decision path that is already somewhat movable.

Radiology is a good example. Imaging referrals often have several viable sites, visible price variation, and relatively low emotional complexity. A physician or practice can shift a patient toward a preferred location without rewriting the entire relationship around that choice. The decision is real, but it is not sacred. That means a system can intervene at multiple points: set a goal, offer a reward, show the alternatives, and provide feedback. The result is not just compliance, but a new default.

Specialist referrals, by contrast, are sticky. They are tied to long standing personal relationships and often occur less frequently. That makes them less like choosing a route on a map and more like changing a trusted collaborator in a long running project. No amount of monthly reporting fully overcomes the social and professional inertia embedded in that decision. Laboratory and gastroenterology decisions sit somewhere in between. Some behavior changes occur, but not enough to move the economics decisively.

This is the first mental model worth keeping: not all decisions are equally programmable. Some are high frequency and low relational cost. Others are low frequency but high trust, high identity, or high coordination cost. The best intervention strategy depends on the kind of choice you are trying to influence.

You can think of it as a spectrum of decision friction:

1. Low friction decisions: easy to compare, easy to switch, high standardization. These are the best candidates for feedback, default changes, and incentives.
2. Medium friction decisions: partially standardized, but still shaped by habits or workflow constraints. These need coaching, simplification, and visible alternatives.
3. High friction decisions: relational, identity laden, or deeply contextual. These are resistant to simple incentives and require trust, redesign of relationships, or changes in the underlying network itself.

The important point is that behavior change often fails when leaders treat all choices as if they were low friction. They are not. Good systems map the friction first, then choose the intervention.

The easiest way to waste an incentive is to attach it to a decision that the system has already made hard.

That is a lesson software engineers know instinctively. It is also why AI agent design is such a useful metaphor here. No one expects an intelligent system to improve simply because you tell it to “try harder.” It needs a perception layer, a planning layer, memory, and tools. Human organizations are no different.

---

The organization as an agent: perceive, plan, remember, act

AI agents are compelling because they separate intelligence into components. A model that merely produces outputs is not enough. It must first perceive the environment, plan across steps, remember what matters, reason under uncertainty, act through tools, communicate with others, and learn from feedback.

That framework is surprisingly useful for understanding why some health care interventions succeed while others stall.

1. Perception: what the system can see

The referral intervention worked partly because it made the hidden visible. Physicians and practices received individualized reports showing where they were already sending patients and where high-value alternatives existed. That is a perception intervention. It does not tell people what to want. It changes what they can see.

Organizations often assume their members already know the landscape. They do not. A physician may know a hospital by reputation but not its comparative cost, or a scheduler may know a familiar site but not the alternative with equivalent quality and lower price. In AI terms, the system is blind to relevant features of the environment unless someone builds a perception module.

That is why dashboards alone are often underpowered. A dashboard is not perception unless it is timely, relevant, and actionable. Otherwise it is just decorative cognition.

2. Planning: what the system can do next

The intervention did not merely say “use lower-cost sites.” It built a route from current behavior to better behavior: set a goal, identify acceptable sites, coach the practice, then review progress monthly. That is planning. It breaks a vague aspiration into manageable steps.

The same is true for AI agents. Planning matters because intelligent action is rarely one move. It is sequence management. The agent must decide what comes first, what depends on what, and where uncertainty can be tolerated.

In organizations, many improvement efforts fail because they announce the destination but never design the path. “Be more cost conscious” is not a plan. “For imaging orders, here are the preferred sites, here is the target for this month, here is how your team will get the information at the point of scheduling, and here is the feedback loop” is a plan.

3. Memory: what the system retains

A monthly report only matters if the system can remember last month. This is where memory enters. The intervention used repeated performance feedback and customized lists, effectively creating organizational memory around a behavior that would otherwise be forgotten after the moment of referral.

AI agents distinguish between short-term and long-term memory because context matters. Humans need the same thing. If a practice keeps relearning the same lesson every month, the organization is stateless. Stateless organizations are busy, but not smart.

Memory is not just record keeping. It is how systems avoid starting from zero every time. It is also how they build identity around repeated choices. When a practice sees that it consistently sends imaging to a high-value site, the behavior begins to feel normal. What begins as an incentive can become a habit.

4. Reasoning: how the system chooses

The intervention suggests that choice architecture matters more than abstract preference. Physicians do not choose purely by price or purely by habit. They weigh quality, convenience, familiarity, workflow burden, patient needs, and interpersonal trust. AI agents do this explicitly through reasoning modules, evaluating different paths against goals and utility functions.

Organizations, by contrast, often pretend to reason while really just inheriting defaults. A referral is made to the known site because the known site is known. An office uses the familiar process because the familiar process is already embedded in the workflow. That is not reasoning. That is inertia dressed up as judgment.

The intervention changed the utility landscape. It did not eliminate judgment, but it made lower-cost alternatives easier to justify and easier to remember.

5. Action and communication: how the system executes and coordinates

The most underrated part of the referral intervention may be the front office. Schedulers and staff often mediate the actual site choice. That makes the intervention partly an action and communication system, not just a physician behavior system.

This is where many organizations miss the real bottleneck. Leaders address the person with authority, but the actual action is executed by the person with the workflow. AI agents understand this intuitively: decision and execution are separate modules. A system can reason perfectly and still fail if the action layer cannot access the right tools.

The same principle applies to health care referrals. If the preferred imaging center is not in the scheduler’s mental map, the decision never becomes real. If the preferred site is hard to reach, the incentive stays abstract. If the system does not communicate the available options in a usable format, nothing changes.

6. Learning: how the system improves over time

Finally, the intervention used repeated feedback to create learning. That matters because one-time education is not adaptation. Learning requires a loop: action, outcome, reflection, adjustment.

AI systems improve by evaluating outcomes and refining behavior. Organizations should be built the same way. Yet too many interventions are static. They are launched like posters, not like learning systems. They do not update by service type, friction level, or user response.

The mixed results across radiology, orthopedics, gastroenterology, laboratories, and specialist referrals are not evidence that learning failed. They are evidence that the system learned something important: different behaviors need different architectures.

---

The deepest design principle: reduce friction before you add force

There is a temptation, in both management and policy, to think that if a desired behavior does not happen, we need more force. More incentives. More rules. More reminders. More penalties.

But the better principle is often the opposite: reduce friction before you add force.

Friction is anything that makes the better choice harder to see, harder to execute, or harder to repeat. In the referral study, friction might be a lack of awareness of lower-cost sites, uncertainty about quality, habit, scheduling inconvenience, or the relational pull of existing specialist networks. The intervention reduced some of that friction, but not all of it. That is why it worked unevenly.

This principle generalizes far beyond health care. A company trying to get employees to use an internal AI tool should not begin with training alone. It should ask: Is the tool visible at the point of need? Does it fit existing workflow? Does it remember context? Can it act quickly? Does it communicate clearly? If any of those answers are no, more enthusiasm will not fix the problem.

The same logic applies to patient care. If the goal is lower-cost, high-value settings, then the system must make that choice the path of least resistance. That can mean customized site lists, scheduling support, feedback on performance, and incentives sized to actual behavioral barriers. It can also mean understanding which choices are fundamentally resistant to optimization because they depend on trust relationships that should not be flattened into cost alone.

This is a subtle but crucial distinction. Not every high-cost decision should be cheapened. Some choices are expensive because they are complex, unique, or relationship based. A good system does not blindly optimize everything. It distinguishes between waste, craftsmanship, and trust.

The goal is not to mechanize judgment. The goal is to remove the unnecessary effort around good judgment so that the right choice becomes easier to sustain.

That is why the phrase “high-value” matters so much. It is not simply low cost. It is lower cost with equivalent or better quality. The objective is not austerity. It is alignment. Cost savings matter because they indicate that the system is no longer paying a premium for avoidable friction.

---

Key Takeaways

• Map decision friction before designing an intervention. Ask whether the behavior is low, medium, or high friction. Different choices require different tools.
• Turn aspiration into a sequence. Goals alone are weak. Pair them with coaching, defaults, feedback, and clear next steps.
• Build memory into the workflow. If a team must relearn the same decision every time, the system is not learning. Make good behavior visible and repeatable.
• Treat communication as infrastructure, not decoration. The best decision is useless if the person who executes it cannot easily access the relevant information.
• Reduce friction before adding force. Incentives work best when they amplify an already feasible path. They do not rescue a broken one.

---

What this changes about how we think about improvement

The usual story about better decisions is too human and too simple: give people more information, and they will choose better. The real story is more interesting. Better decisions happen when a system is designed so that perception, planning, memory, reasoning, action, communication, and learning all point in the same direction.

That is why the connection between AI agents and physician referral behavior is more than a clever analogy. It is a blueprint for institutions. Organizations are not just collections of incentives or rules. They are decision engines. When they are badly designed, even intelligent people make expensive, inconsistent choices. When they are well designed, modest nudges produce outsized gains.

The most important question is no longer, “How do we convince people to do the right thing?” It is, “What architecture would make the right thing the easiest thing to do, the easiest thing to remember, and the easiest thing to repeat?”

Once you start asking that question, cost savings look less like austerity and more like intelligence.