The Decision Flow Problem

In 1987, George Stalk Jr. of the Boston Consulting Group published a deceptively simple observation about corporate operations. He called them the “Rules of Response,” and they described how time moves through value-delivery systems. The insight was this:

"Products and services receive value for only 0.05 to 5 percent of the time they spend in a company's system. The rest is waiting."

Stalk was talking about physical goods. A heavy vehicle manufacturer takes forty-five days to prepare an order for assembly, but only sixteen hours to actually assemble each vehicle. The vehicle is being worked on for less than 1 percent of the time it spends in the system. The remainder is waiting: waiting for a batch to complete, waiting for the batch ahead of it, and, critically, waiting for management to make and execute the decision to move it to the next step.

That third category of waiting is where things get interesting for us.

The Invisible Batch: Decisions

Stalk’s three sources of waiting time divide almost equally. A third of the waste comes from waiting for batches to fill. A third from waiting for prior batches to clear. And a third from waiting for someone to decide what happens next.

Now consider what a supply chain planning organization does. It doesn’t move physical goods. It moves decisions through an organization. A demand planner detects a shift in customer ordering patterns. That signal must flow to the supply planner, who adjusts replenishment. The change propagates to the MPS manager, who resequences production. The allocation manager redistributes available-to-promise. The procurement analyst places revised purchase orders.

Each of these is a decision. And each decision, like Stalk’s physical goods, spends almost all of its time waiting, not being made.

The .05 to 5 Rule applies to decisions too. A procurement analyst might spend five minutes deciding to expedite a purchase order. But the information that triggered that decision may have been sitting in a report for three days, waiting to be noticed. Before that, it waited two days in an exception queue. Before that, the underlying demand signal waited a week to be aggregated into the next planning cycle.

The decision itself takes minutes. The time it spends in the system? Weeks.

Why Decisions Wait

In 2018, Ajay Agrawal, Joshua Gans, and Avi Goldfarb, economists at the University of Toronto’s Rotman School of Management, published Prediction Machines. Their thesis was elegant:

"AI is best understood as a drop in the cost of prediction. And when the cost of prediction drops, the economics of every decision that depends on prediction changes."

But they made a subtler point that’s often overlooked. Every decision, they argued, has an anatomy: data goes in, prediction happens, judgment is applied, and action comes out. When prediction was expensive, organizations batched decisions. They gathered data weekly, ran forecasts monthly, and held planning meetings quarterly. This wasn’t because monthly was the right cadence for decisions, it was because prediction was expensive enough that you had to batch it to justify the cost.

Planning organizations batch decisions for the same reason factories batch production runs: to amortize the setup cost across enough units to make it economical.

The “setup cost” for a supply chain decision isn’t a die change on a press. It’s the time a planner spends gathering context. Pulling up inventory positions across six warehouses. Cross-referencing the forecast with the latest sales orders. Checking supplier lead times. Reviewing the capacity plan. Reading three emails from the regional sales manager about a customer promotion.

By the time the planner has assembled enough context to make a good decision, forty-five minutes have elapsed. The decision itself, increase the safety stock target by 200 units at the Dallas DC, takes seconds.

"Two-thirds of supply chain organizations spend 30% or more of their time firefighting rather than planning."

This is Stalk’s .05 to 5 Rule, applied to information work. The value-adding act (the decision) occupies a tiny fraction of the total cycle time. The rest is information gathering, context assembly, and organizational coordination.

What Happens When Decisions Become Cheap

Agrawal, Gans, and Goldfarb observed that when technology makes something cheap, you use more of it. When electricity made lighting cheap, factories didn’t just replace gas lamps, they redesigned buildings, extended operating hours, and invented the assembly line.

So what happens when you make supply chain decisions cheap? Not cheaper in quality — cheaper in cycle time. What if the elapsed time from “demand signal changes” to “corrective action taken” dropped from days to seconds?

This is the promise of autonomous supply chain agents. Not that any single decision is better than what a skilled planner would make. A well-trained agent making an inventory rebalancing decision and an experienced supply chain analyst making the same decision will arrive at similar conclusions. The atomic decision quality is comparable.

The difference is everything that surrounds the decision.

The agent doesn’t spend forty-five minutes gathering context. It already has context — it’s continuously monitoring inventory positions, demand signals, supplier status, and capacity constraints across every site in the network, simultaneously. It doesn’t wait for the weekly planning meeting to surface an exception. It detects the exception the moment the underlying data changes. It doesn’t wait for a handoff between the demand planner and the supply planner and the allocation manager. It operates across the full decision chain in a single pass.

"With advanced system support, 80 to 90 percent of all planning tasks can be automated."

A human planner can process perhaps twenty to thirty meaningful decisions per day. Not because they’re slow thinkers, because each decision requires extensive context gathering, cross-functional coordination, and organizational navigation. The decision itself is the easy part. The hard part is everything before and after.

An autonomous agent processes the same decisions in seconds, twenty-four hours a day, seven days a week. Not because it’s smarter. Because it has eliminated the waiting time between decisions.

Stalk’s Rules, Reframed

Let’s revisit BCG’s Rules of Response through the lens of decision flow:

The .05 to 5 Rule for Decisions: In most planning organizations, actual decision-making represents 0.05 to 5 percent of the total elapsed time. The rest is gathering information, waiting for the planning cycle, coordinating across functions, and shepherding the decision through approval workflows.

The 3/3 Rule for Decisions: The waiting time divides roughly into thirds: waiting for enough information to accumulate (the “batch”), waiting for other decisions ahead in the queue (the “sequence”), and waiting for organizational bandwidth to process the decision (the “capacity”).

The 1/4-2-20 Rule for Decisions: For every quartering of the decision cycle time, labor productivity doubles and costs fall by 20 percent. A company that moves from weekly planning cycles to continuous agent-driven planning doesn’t just make faster decisions, it fundamentally changes the economics of its planning organization.

The 3 x 2 Rule for Decisions: Companies that compress their decision cycle times grow at three times the industry average with twice the profit margins. This is the competitive moat: not better algorithms, but faster organizational metabolism.

"Time-based competitors enjoy growth rates of three times the average, and twice the profit margin within their respective industries."

The Economic Inversion

What Stalk described for physical goods, and Agrawal described for prediction, is part of a broader structural shift that Jordi Visser calls the agentic inversion: the moment when AI severs capital’s dependence on labor for cognitive work.

"Companies growing top-line revenue over 20% annually, while operating expenses and headcount grow at just 2%. It's never happened before."

The economics are stark. Once an AI agent has been trained to make a class of supply chain decisions, the marginal cost of replicating that agent to handle increased volume is near-zero. An agent that can evaluate one inventory rebalancing decision can evaluate a million. This breaks the linear link between operational scale and headcount cost that has defined planning organizations for decades.

Ronald Coase famously argued that firms exist because the transaction costs of using the market for every step of production are too high. It’s cheaper to hire a planner and coordinate internally than to contract out each decision. But when agents radically reduce these coordination costs, discovering context, evaluating trade-offs, negotiating across functions, and executing actions autonomously, the economic rationale for large planning organizations inverts.

"Agentic AI is about replacing coordination, not just tasks. When coordination becomes programmable, the function that defined firm boundaries becomes automatable."

This is the deeper implication of cheap decisions. It’s not just that individual choices happen faster. It’s that the entire cost structure of the planning function changes. The setup cost per decision drops toward zero, which means batching is no longer necessary, which means cycle times collapse, which means fewer people are needed for routine coordination, which means the planning organization can be smaller, faster, and more focused on the judgment calls that actually require human expertise.

The planners who remain become more valuable, not less. Their role shifts from processing decisions (the 95% that was waiting time) to providing judgment on the exceptions that agents escalate, the novel, ambiguous, politically sensitive decisions where thirty years of institutional knowledge actually matters. The volume of human decisions shrinks. The value of each one increases.

"This is not a recession. It's a reordering of the economic contract between labor and capital, accelerated by AI's exponential reach."

The Real Argument for Agents

The case for autonomous supply chain agents is not that they make better individual decisions than humans. That argument is difficult to prove and easy to contest. A seasoned planner with thirty years of experience and deep institutional knowledge will sometimes outperform any algorithm on a complex, ambiguous, novel situation.

The case is that agents eliminate the waiting time between decisions.

A supply chain doesn’t fail because one person made one bad call. It fails because a thousand small signals went unnoticed for too long. A slight uptick in demand at three retail sites. A two-day delay at a port that affects four inbound shipments. A quality hold on one production batch. Each of these, detected and addressed in hours, is a minor adjustment. Left unattended for days or weeks, because the planning organization only reviews exceptions on Tuesdays, or because the analyst responsible was on vacation, or because the signal was buried on page four of a forty-page report — they compound into stockouts, expediting costs, and missed service levels.

Agents don’t sleep. They don’t take vacation. They don’t have forty-five-minute context-gathering sessions. They don’t batch decisions into weekly cycles because the setup cost of each decision is effectively zero.

They make the system better not by making any single decision better, but by making all decisions sooner.

The Economics of Continuous Correction

Agrawal and his colleagues would frame it this way: agents make prediction cheap enough that you can afford to predict, and therefore decide, continuously rather than periodically. When prediction was expensive (gathering all that context manually), you predicted weekly. When prediction is cheap (agents monitoring everything in real-time), you predict continuously.

And Stalk would frame it this way: the company that compresses its decision cycle time from weekly to continuous has applied the 1/4-2-20 Rule not once, but repeatedly. Each quartering, weekly to daily, daily to hourly, hourly to real-time, doubles productivity and cuts costs by 20 percent.

The multiplication is what matters. Not 20 percent once. Twenty percent compounded across four quarterings of cycle time. The fast-response planning organization doesn’t just plan better. It plans in a fundamentally different mode, continuous correction rather than periodic replanning.

This is what the building materials manufacturer in Stalk’s paper achieved when it cut order-to-delivery from five weeks to one. Not by working harder. By eliminating the waiting time in its system. Its growth rate exceeded the industry average by more than three to one. Its return on assets was double the competition.

The same economics apply to the flow of decisions through a planning organization. The company that detects and corrects course in seconds will outperform the company that detects and corrects course in weeks, not because its planners are smarter, but because its decision cycle time is shorter.

That is what agents make cheaper, faster, and better. Not the decision. The flow.

Agrawal, Gans, and Goldfarb, economists at the University of Toronto’s Rotman School and co-founders of the Creative Destruction Lab, observed a pattern in technology adoption: the resistance to new technology breaks when it makes something unambiguously cheaper, faster, or better. Autonomous agents do all three simultaneously. They make decisions cheaper by eliminating the 45-minute context-gathering cost. They make them faster by compressing weeks into seconds. And here’s the insight that Stalk’s work makes clear: increasing velocity doesn’t just make things faster, it makes them better. The building materials company didn’t just deliver faster. Its defect rates fell, its inventory turns doubled, and its market share grew. Speed created quality.

The same is true for decision flow. A company that detects a demand shift in seconds and corrects course in minutes doesn’t just respond faster, it responds before the problem compounds. The small correction at hour one prevents the 847-exception Monday morning. Detection velocity, decision velocity, correction velocity, each compression makes the entire system more resilient, more accurate, and more valuable.

Velocity creates value.

Humans Ain’t Going Nowhere

"AI automates tasks, not purpose. Tasks get automated, but humans still own outcomes. If your job is tasks, AI threatens it. If it's purpose, AI makes you more powerful."

Huang’s distinction between tasks and purpose is the key to understanding what happens to supply chain planners in an agent-driven world. The tasks, gathering data, checking exception reports, cross-referencing forecasts, sending emails to coordinate across functions, are the 95% that was waiting time. Those are what agents eliminate.

What remains is purpose. And in supply chain planning, purpose has two dimensions:

Context curation. Agents are only as good as the context they receive. The demand planner who knows that a particular retailer always over-orders in Q4 and then cancels in January, that’s institutional knowledge that no algorithm can discover from transactional data alone. The supply planner who has a personal relationship with the plant manager in Mexico and knows that the published capacity numbers are conservative by 15%, that context changes every downstream decision. Humans curate, enrich, and correct the context that agents consume. They feed executive directives into the system: “prioritize service level over cost this quarter.” They interpret ambiguous signals that agents can’t parse: a rumor about a competitor’s new product, a regulatory change in a key market, a shift in consumer sentiment that hasn’t shown up in the numbers yet.

Judgment on exceptions. When an agent encounters a situation that falls outside its training distribution, a novel combination of demand spike, supplier failure, and capacity constraint that it has never seen before, it escalates. Not because it’s broken, but because it knows the limits of its own competence. The human planner provides the judgment: Is this a temporary disruption or a structural shift? Should we absorb the cost or pass it to the customer? Is this the moment to switch suppliers or ride it out?

The override itself becomes training data. When a planner overrides an agent’s recommendation and the outcome is better, the system learns. When the override leads to a worse outcome, the system learns that too. Over time, the agents internalize the organization’s decision culture, not by replacing human judgment, but by absorbing it.

This is the flip side of the economic inversion. The volume of human decisions shrinks dramatically. But each remaining decision is higher-stakes, higher-context, and higher-value. The planner who used to make thirty routine decisions a day now makes three that matter. Their job title might stay the same. Their actual role, the purpose behind the tasks, becomes more important, not less.

Huang is right. If your job is tasks, AI threatens it. If your job is purpose, curating context, providing judgment, feeding the learning flywheel, AI makes you more powerful than you’ve ever been.

The building materials company in Stalk’s paper didn’t win by making better products. It won by moving faster. The supply chain company of the future won’t win by making better individual decisions. It will win by detecting sooner, deciding faster, and correcting continuously. The atomic decision quality between a skilled planner and a well-trained agent may be comparable. But the velocity gap is unbridgeable. And velocity creates value.

---

This is the first in a series on the economics of autonomous supply chain planning.

References

• Stalk, G. Jr. (1987). Rules of Response. The Boston Consulting Group, Perspectives No. 317.
• Stalk, G. Jr. & Hout, T. M. (1990). Competing Against Time: How Time-Based Competition is Reshaping Global Markets. Free Press.
• Agrawal, A., Gans, J., & Goldfarb, A. (2018). Prediction Machines: The Simple Economics of Artificial Intelligence. Harvard Business Review Press.
• Agrawal, A., Gans, J., & Goldfarb, A. (2022). Power and Prediction: The Disruptive Economics of Artificial Intelligence. Harvard Business Review Press.
• Gartner (2023). Manage Supply Chain Disruption Without Firefighting.
• McKinsey & Company (2024). Supply Chain 4.0, The Next-Generation Digital Supply Chain.
• Visser, J. (2025). The Domino Effect: AI, Labor, and the Unraveling of the Old Economic Order. Visser Labs.
• Coase, R. (1937). The Nature of the Firm. Economica.
• Hadfield, G. & Koh, A. (2025). An Economy of AI Agents. Johns Hopkins / MIT. arXiv:2509.01063.

Remember: Velocity creates value.