The Economics of AI

Picture a single page in an executive committee pack. The property group behind it is invented, but the page is not unusual. A residential property group that lets apartments through its own website, an app and two listing portals has tried an assistant that drafts listing descriptions. For one month the assistant wrote 4,000 descriptions for apartments in one region, and the model bill came to 600, about 0.15 per description. The request on the page is simple: take the same assistant to every region and every channel, about 40,000 descriptions a month, for a model bill of 6,000.

Many executives would sign it. The arithmetic is correct, the amount is modest and the pilot worked. Yet the page answers a question nobody should be asking. It tells you what an answer costs. It does not tell you what the outcome costs: accurate listings, in every channel, kept right every time a rent changes, an apartment is refurbished or a regulator changes what a listing must disclose.

The model price is one line of the bill

That gap is the idea of this chapter, and the last big idea of Module 01: the model price is one line of the bill.

The price of an answer is the number everybody sees. It is small, it appears on an invoice, and it keeps falling. As Why AI, Why Now? showed, the cost of querying a model as capable as the first ChatGPT fell more than 280-fold between November 2022 and October 2024¹. A number that falls that fast invites the conclusion that AI is nearly free.

The cost of an outcome is different. An outcome is a business result, such as a listing that is correct, a tenant query that is resolved or a contract that is reviewed, delivered reliably at the scale of the whole business. It includes the model, but it also includes the people who connect the system to the property records, the data that keeps it accurate, the checks that catch its mistakes and the time thousands of employees spend learning to work with it.

AI economics, in other words, are system economics. This chapter gives you no price list, because prices change every quarter. It gives you the reason the subject deserves your attention now. Module 08, AI Economics, gives you the tools.

Why the bill is a leadership question now

For most of the last decade, AI spending was a research budget: a data science team, some cloud capacity, a few experiments. That has changed quickly. The FinOps Foundation, whose members manage technology spending in large organizations, asks every year whether its practitioners are responsible for AI costs. In 2024, 31 percent said yes. In 2026, 98 percent did².

The reason is not only that more organizations use AI. It is that an AI bill behaves like a meter: it moves with every request, every user and every design choice. A license costs the same whether people use it well or badly. A model bill does not. Module 08’s opening chapter, also called The Economics of AI, explains what that metered behavior means for budgets.

Economics also decides which pilots survive. In July 2024 Gartner predicted that at least 30 percent of generative AI projects would be abandoned after proof of concept by the end of 2025. Escalating costs were one of the four reasons it named, alongside poor data quality, inadequate risk controls and unclear business value³. A prediction is not a measurement, but the pattern it describes will be familiar to anyone who has watched a promising demonstration stall when the full cost of running it became visible. A pilot that cost almost nothing can still die of economics, because the pilot never showed what the outcome would cost.

Much of the bill never reaches the invoice

What, then, is missing from the page in the committee pack? Engineers answered that question for machine learning more than a decade ago. In a widely cited paper, a team at Google drew a real-world machine-learning system as a large diagram in which the model code was one small box. Around it sat data collection, data verification, configuration, monitoring, serving infrastructure and more. Only a small fraction of a real system, they wrote, is the model itself, and the surrounding infrastructure is vast and complex⁴.

Generative AI did not change that shape. For the property group, the rest of the bill looks like this.

None of these lines is exotic. The engineers and lettings experts who built the pilot were not free because they were already on the payroll; their time came out of other work. The pilot read from a clean extract of one region; the rollout needs live links to every property, rent and availability record. In the pilot, a lettings manager read every description. At 40,000 a month, someone has to decide which descriptions need a human check and what happens when a wrong claim about a home, such as its size, its energy rating or its rent, reaches a prospective tenant. And the letting agents, property managers and portal managers who will use the assistant need time to learn a new way of working.

A pilot can skip most of this. Production cannot. Module 08 sorts these lines into cost families and teaches how to count them over the life of a system, starting with Understanding AI Total Cost of Ownership.

The rest of the bill arrives in other budgets

There is a second reason the model price misleads, and it is organizational rather than technical: the rest of the bill arrives in other budgets.

In the property group, IT pays for the integration. The data team pays for cleaning property records. Security pays for the access reviews. The lettings team gives up hours so its people can check and correct descriptions, and the portal team adapts every listing to each portal’s rules. Training comes out of the people budget. Only the model invoice, paid by finance, has AI in its name. Each budget holder sees a modest, defensible number. Nobody sees the total, so nobody can say whether the outcome is worth what it costs.

That is why AI economics is a leadership question and not a procurement question. Procurement can negotiate the price of the model. Only the executive who owns the outcome can decide whether the whole bill is worth paying, and only if someone puts the whole bill in one place. How organizations tag, allocate and show those costs back to the teams that cause them is the subject of Cost Optimization and AI FinOps in Module 08.

Prices move, and not only down

A third reason to look past the price tag is that the price itself is not stable ground. The price per unit of model output has fallen fast, and may keep falling. But organizations do not buy units; they buy work done. As systems move from answering a question to carrying out a task in several steps, the amount of model output each task consumes grows.

In August 2026, Gartner forecast that the inference cost of each agentic workflow would rise more than fivefold through 2028, even as the price per unit kept falling, because each new generation of capability uses more, and often more expensive, output⁵. That is a forecast, and forecasts about AI have been wrong in both directions. The lesson does not depend on the exact number. A budget built on today’s price per answer can be wrong in either direction within a year; a budget built on what each outcome needs is far more durable. Module 08 calls the pattern in which cheaper units lead to a bigger total bill a rebound, and explains it in its first chapter.

Two parts of the ledger sit elsewhere in the course. The value an initiative creates came in the previous chapter, Measuring AI Business Value, and Module 03 turns it into a business case. The cost of things going wrong belongs to Module 06, starting with The AI Risk Landscape. Engineering levers such as routing simple work to smaller models, caching repeated answers and batching requests belong to Cost Optimization and AI FinOps, in Module 08, which takes up the whole cost side in depth. For now, one habit is enough: never judge an AI initiative by its model bill alone.

Story: from six weeks to four days

The economics of AI are not only a story about hidden costs. Seen whole, they also explain why some organizations change what they can afford to do. Zalando, a European online fashion retailer, gives a documented example.

Before generative AI, the imagery for an editorial campaign was produced the conventional way: photo shoots with models, photographers and stylists, planned and delivered over weeks. In fashion, where trends spread on social media can be short-lived, that lead time limited which moments the company could respond to. In May 2025, Matthias Haase, Zalando’s vice president of content solutions, told Reuters what had changed. By the company’s account, generative AI had cut the time to produce imagery from six to eight weeks to around three to four days, and reduced its costs by 90 percent. Zalando also said that around 70 percent of its editorial campaign images in the fourth quarter of 2024 were AI-generated, and that it was developing AI “digital twins” of real models, so that a model featured in a campaign can appear as an exact replica on the app’s product pages⁶. These are the company’s own figures, not an independent measurement.

Three things in this account are worth a leader’s attention. First, the numbers are about an outcome, not an answer. Nobody reported the price of a generated image; Zalando reported what a campaign’s imagery cost and how long it took. Second, the value claim was not about quality. Haase said the point was not that AI content is better, but “how new, how relevant it is to our customers”⁶. Speed to a trend was the outcome being bought.

Third, and most useful for the next budget meeting, the account leaves the whole bill unstated. The article does not say what the 90 percent includes: the cost of building digital twins, the fees paid to the real models whose likeness they use, the creative staff who direct and review the images, or the rights and approvals each image needs. That is not a criticism of Zalando, which owes the press no cost breakdown. It is the habit this chapter asks for. When someone shows you a striking AI saving, yours or a competitor’s, ask what the outcome costs and which lines the number contains.

What this means for leaders

The model price will dominate conversations about AI cost, because it is the number that arrives on an invoice. Your job is to widen the frame. Treat a cheap pilot as good news about learning, not as a forecast for production. Ask for the cost of an outcome, not of an answer. Insist that one owner sees the whole bill, wherever its lines are paid. And build budgets on what each outcome needs rather than on today’s unit price, because the unit price will not stay where it is.

None of this requires you to become a cost accountant. It requires you to refuse a page like the one at the start of this chapter until it shows the rest of the bill.

Check yourself

1. If the model costs almost nothing per answer, the AI solution is cheap.MythFact
2. The price of querying a model of fixed capability fell more than 280-fold between late 2022 and late 2024.MythFact
3. A pilot’s model bill multiplied by the new volume is a reliable rollout budget.MythFact
4. Most FinOps practitioners now manage AI spending.MythFact
5. Because unit prices keep falling, the total cost of AI work will fall too.MythFact
6. Zalando’s reported 90% saving tells you the full cost of its AI campaign imagery.MythFact

Reflection: find the missing lines

What comes next

This chapter closes Module 01, which has made one argument from sixteen angles: AI is a technological, economic and organizational change, and leaders have to lead it. In four lines, the module said this.

Transformation: like electricity before it, AI pays off only when organizations redesign the work around it. Competition: the models are open to almost everyone, so advantage comes from what you build around them and how fast your organization learns. Work: language made AI usable by almost anyone, agents can now act, and jobs change task by task. Leadership: hold the components together, start where feedback is fast, measure value rather than activity, and count the whole bill.

The decisions ahead are about what to trust, what to ask a vendor, what to fund and what to stop. Making them well needs a clear picture of the technology itself: not an engineering course, but an executive mental model. Module 02 builds it, starting with the most basic question of all, in What Exactly Is Artificial Intelligence?