Businesses are graphs, not lines

Every operating model deck draws the business the same way: plan, source, make, move, sell, serve. Six boxes, five arrows, left to right. It is a tidy picture and it decides a great deal, because it is also how budgets, teams, systems and now AI use cases get carved up.

The picture is wrong. Those decisions do not happen in sequence. They happen at once, and each one pulls on the others.

The line breaks even when everyone does their job well

Procter & Gamble noticed something odd about Pampers. Babies consume nappies at a steady rate and retail sales were fairly stable, yet distributors’ orders swung widely and P&G’s own orders to its suppliers swung more. Hewlett-Packard found the same in printers: reseller orders moved far more than sales, and orders to its chip division more still.¹

Lee, Padmanabhan and Whang named this the bullwhip effect and traced it to four causes: demand forecast updating, order batching, price fluctuation, and rationing and shortage gaming. Their central finding is the uncomfortable part. The distortion is “a consequence of the players’ rational behavior within the supply chain’s infrastructure.” Nobody is being careless. Each node is optimising correctly on what it can see.¹

John Sterman had shown the same thing in the laboratory in 1989. Give capable people one link of a four-stage chain and a single step up in demand, and they produce large oscillations, because most largely ignore the orders already in the pipeline.³ Switch between the two views below.

Industry studies of grocery found over 100 days of inventory in the chain, and the industry’s ECR reports put the opportunity from removing the inefficiency at $30 billion.¹

Price moves demand, demand moves inventory

One of those four causes is a pricing decision. Campbell Soup ran its trade promotion on chicken noodle soup every January. Retailers responded by forward buying, in some cases a year’s supply. To meet the spike, the chicken-boning plants had to start running overtime in October.²

A marketing calendar was setting the factory’s shift pattern three months earlier. On the org chart those two functions sit at opposite ends of the line. In the business they are adjacent nodes.

When Campbell connected them, through continuous replenishment and pricing terms that removed the incentive to forward buy, in-stock rates rose from 98.5% to 99.2%, retailer inventory fell from about four weeks to two, and sales through participating retailers grew twice as fast as through the rest.² Higher availability, less stock, more sales, at the same time.

Inventory moves cash

In 2000, facing component shortages, Cisco locked in long-term purchase commitments on the strength of bookings. Everyone padded orders to secure allocation. As one Cisco supplier put it: “if they ask for 100, they’ll get 80. So they ask for 120 to get 100.” When demand turned, quarterly sales fell 30%, Cisco wrote off about $2.2 billion of inventory and cut 8,500 jobs.⁷

Cisco had some of the most advanced real-time reporting of its day. The data was integrated. The decisions were not: sales forecasts, supplier commitments and the balance sheet were each managed along their own line.

Run it as one system and both lines move

The clean experimental evidence comes from retail, where the price–demand–inventory loop is tight.

• Zara. Clearance markdowns were set manually. Caro and Gallien replaced this with a demand forecast coupled to a price optimisation over remaining stock. A controlled field test in Belgium and Ireland in 2008 raised clearance revenue by about 6%; Zara rolled it out worldwide.⁴
• Rue La La. Ferreira, Lee and Simchi-Levi built demand prediction and price optimisation as one tool for first-time products. Revenue rose by almost 10%.⁵
• Across industries. McKinsey assessed more than 170 companies over five years. Those with mature integrated business planning showed one to two points more EBIT, service levels 5 to 20 points higher, capital intensity and freight costs 10–15% lower, and 40–50% fewer customer penalties and missed sales.⁶

The usual trade-off between growth and margin exists inside a function. Across functions it often disappears, because the waste sits in the gaps between them.

Automate the line and you make the old mistakes faster

This is where most AI roadmaps go wrong. The use-case list mirrors the boxes: a forecasting model for planning, a pricing model for sales, a replenishment model for supply. Each is trained on its own metric and each can improve while the business does not.

The bullwhip research says why. The first cause on the list is each node re-forecasting from the orders it receives.¹ A faster, more responsive node amplifies whatever signal it receives. Elmachtoub and Grigas make the same point formally: a prediction model trained to minimise its own error, independent of the decision it feeds, is solving the wrong problem. Train it on the quality of the downstream decision and simple models can beat more accurate ones.⁸

The field data agrees. In McKinsey’s survey, the single attribute most associated with EBIT impact from generative AI, out of 25 tested, was redesigning workflows, and only 21% of adopters had done it.⁹ In software, Google’s DORA study found that each 25% rise in AI adoption came with a 1.5% fall in delivery throughput and a 7.2% fall in stability: one step sped up, the system slowed down.¹⁰

So adopting AI means changing the process first. Decide which decisions must be made together, then build for that.

“You’ve never run a business.”

True. I have spent 14 years inside the data of businesses that do: marketing mix models, retail assortment, pricing, supply chains. Data does not respect the org chart. A promotion shows up in the stock-out report. A stock-out shows up in next quarter’s churn. A payment term shows up in the forecast.

The same pattern is there in a P&G nappy line in the 1990s, a soup factory in October and a network-equipment maker in 2001.

Operators see their own node in depth. The data shows the edges. That is the view from which the line looks most obviously wrong.

References

1. H. L. Lee, V. Padmanabhan and S. Whang, “The Bullwhip Effect in Supply Chains,” Sloan Management Review, Spring 1997. PDF
2. M. L. Fisher, “What Is the Right Supply Chain for Your Product?” Harvard Business Review, March–April 1997. Link
3. J. D. Sterman, “Modeling Managerial Behavior: Misperceptions of Feedback in a Dynamic Decision Making Experiment,” Management Science 35(3), 1989. MIT working paper
4. F. Caro and J. Gallien, “Clearance Pricing Optimization for a Fast-Fashion Retailer,” Operations Research 60(6), 2012, pp. 1404–1422. Link
5. K. J. Ferreira, B. H. A. Lee and D. Simchi-Levi, “Analytics for an Online Retailer: Demand Forecasting and Price Optimization,” Manufacturing & Service Operations Management, 2016. Summary
6. E. Dumitrescu, M. Jochim, A. Sankur and K. Shah, “A better way to drive your business,” McKinsey, May 2022. Link
7. S. Berinato, “What Went Wrong at Cisco in 2001,” CIO, 1 August 2001. Link
8. A. N. Elmachtoub and P. Grigas, “Smart ‘Predict, then Optimize’,” Management Science 68(1), 2022. arXiv
9. McKinsey, The state of AI: How organizations are rewiring to capture value, March 2025. Link
10. Google Cloud DORA, Accelerate State of DevOps Report 2024; figures as summarised by RedMonk. Link
11. H. A. Simon, “The Architecture of Complexity,” Proceedings of the American Philosophical Society 106(6), 1962, pp. 467–482.