Module 08 Synthesis — Making AI Economically Sustainable

In August 2026 Gartner published a forecast, reported in the trade press, that reads at first like a contradiction. The price of the tokens that AI models consume keeps falling. Yet Gartner expects the inference cost of a single agentic workflow, an AI system that plans and carries out a multistep task, to rise more than fivefold through 2028. Its explanation is simple: each new generation of capability uses more tokens, and often more expensive ones, to do more ambitious work¹. Gartner’s release blocks automated retrieval, so the figures here were checked against press coverage of it.

Put the question to a leadership team before showing the answer. If the unit price falls, what happens to the cost of the work? The intuitive answer is that it falls too. The forecast says the opposite, and it is not an isolated warning. A year earlier Gartner had predicted that more than 40 percent of agentic AI projects would be canceled by the end of 2027, and escalating costs were the first of the three reasons it gave, ahead of unclear business value and inadequate risk controls².

Neither forecast says that the business cases behind these projects were wrong on the day they were approved. They say something more uncomfortable: the numbers moved afterward, and an organization that is not watching finds out late.

The core idea

Module 03 asked whether an AI capability creates value and how to prove it. This module has asked a different question: what does it cost to keep creating that value, year after year, at the volume the capability will actually reach?

The answer can be put as one definition. An AI capability is economically sustainable when each outcome it delivers earns more than it costs, its whole life is paid for at the volume it really reaches, and someone keeps both true as prices, usage and the work change. That is three tests, and a capability has to pass all three.

The first test is about one unit of work. The second is about the whole system over its life. The third is the one business cases most easily leave out, because it cannot be calculated once. It is a habit of looking again.

Eight chapters, one chain of questions

Each chapter of this module produced one number or one discipline that a sustainable case needs. Sorted by the test they serve, they form a chain of questions, and each link depends on the one before it.

The order matters. A break-even volume built on a cost per request rather than a cost per resolved outcome will be too optimistic, because it ignores retries and human review. A build-versus-buy comparison built on an incomplete TCO will favor whichever option hides more of its costs. The chain is only as strong as its earliest weak link, which is why the module started with the bill and ended with the decision.

The three tests, worked once

Suppose, as an illustration, a container shipping line uses AI to check shipping instructions before a vessel’s cutoff, so that documents with errors are fixed before they become amended bills of lading. The outcome that matters is a document cleared without manual rework. The figures below are invented for the example and used consistently in the rest of this chapter.

Each cleared document is worth 9 to the line, in avoided handling and amendment costs. Its variable cost is 3: about 2 for model calls and compute, and about 1 for the people who review the 8 percent of documents the system flags. So each outcome leaves a contribution of 6. The first test passes.

The fixed cost per year is 1.5 million: 0.4 million of the 1.2 million build, spread over a three-year life, 0.8 million to operate the service and 0.3 million for changes. Break-even is 1.5 million divided by 6, or 250,000 documents a year³. The line expects 400,000. Over three years that makes a total cost of ownership of 8.1 million, using the module’s single definition: build 1.2 million, run 2.4 million, operate 3.6 million (the 2.4 million fixed team plus 1.2 million of exception review) and change 0.9 million. Value over the same life is 10.8 million, so the system nets 2.7 million, or 0.9 million a year. The second test passes.

On the day of approval, this is a good case. The third test asks what happens next.

Economics drift after approval

The numbers in a business case are a snapshot. Four forces move them once the system is live, and every one of them was the subject of an earlier chapter.

Price and volume move together, and not in the direction budgets assume. The Economics of AI showed why cheaper answers can produce bigger bills, and the forecast that opened this chapter is the same effect in agentic form. Mix and exceptions shift as a system meets harder cases than the pilot did; Data, Integration and Operational Costs showed how a small rise in the exception rate can swamp the model bill. Steps and contracts arrive on dates rather than gradually: a capacity tier, a commitment that no longer matches use, a renewal at a new price. The value of the outcome can change too, when the process, the market or a regulation changes what a resolved case is worth.

None of these is exotic, and leaders are right to expect them. When Gartner surveyed more than 300 CIOs in mid-2024, over 90 percent said managing cost limited their ability to get value from AI, and Gartner estimated that a CIO who did not understand how generative AI costs scale could make a 500 to 1,000 percent error in cost calculations⁴. In June 2026 it predicted that at least half of generative AI projects would overrun their budgets through 2028, blaming poor architectural choices and a lack of operational know-how, according to press coverage of a report that is not public⁵. Both are forecasts and estimates, not measurements, and the second is reported rather than read at source. The direction is what matters.

What drift does to the numbers

Return to the shipping line in its second year. Two ordinary things happen. The team adds an agentic second check that reasons through each document again before release. It works, but it raises the cost per cleared document from 3 to 4.5, and because the old process caught the same errors downstream, the value per document stays at 9. Then a trade slowdown cuts volume from 400,000 documents to 320,000.

The arithmetic is short. The contribution per document falls from 6 to 4.5, so 400,000 documents now earn 1.8 million against 1.5 million of fixed cost, a net of 0.3 million. At 320,000 documents they earn 1.44 million, a loss of 0.06 million. Break-even has moved from 250,000 documents to about 333,000 without anyone deciding to move it.

This is the pattern AI Unit Economics and Economics at Scale warned about when it asked which assumption moves the answer most. Here two assumptions moved at once, one on each side of the contribution line. The point is not that the second check was a bad idea. It may be worth keeping at a lower cost. The point is that the case no longer describes the system, and only someone looking at the actuals would know.

Reopen the case on triggers, not anniversaries

An annual review would have caught the shipping line’s loss about a year late. The better practice is to write the conditions that reopen the case into the approval itself, as numbers someone watches every month. FinOps practice calls this making cost data timely and giving every cost an owner⁶⁷. Cost Optimization and AI FinOps described that operating loop; the triggers are what connect it to the business case.

The thresholds come from the case. A 20 percent rise in cost per outcome, from 3 to 3.6, cuts the annual net from 0.9 million to 0.66 million: a warning, not a crisis. A volume forecast below 300,000 cuts the annual net from 0.9 million to 0.3 million or less. In the second-year story, the cost trigger would have fired within two months of the new check going live, and the volume trigger as soon as the trade forecast changed.

When a trigger fires, the response is to rerun the three tests with actual figures, and the answer is one of four.

Repairing the unit cost follows the order set out in Cost Optimization and AI FinOps: question the work before the price. Changing route uses the flip point and exit price from Build vs Buy Economics and Investment Decisions. Retiring is a legitimate answer, not a failure; money already spent is gone whichever way the decision goes. How these single-system decisions add up across many initiatives, including when to stop funding one, is the work of Module 09’s Prioritizing the AI Portfolio.

Story: the checkout that changed route

A clear recent public example of a company running the third test on a working AI system comes from retail, and from a company that could afford to keep almost anything.

Before. Amazon’s Just Walk Out technology uses cameras, shelf sensors and deep learning to work out what shoppers take from the shelves and charge them as they leave, with no checkout line. Amazon put it into many of its Amazon Fresh grocery stores in the United States⁸. The system worked, in the sense that shoppers could walk out and receive a receipt. Whether it paid, in a full-size grocery store with large baskets and thousands of products, is a different question.

After. In April 2024 Amazon said it would remove Just Walk Out from its US Fresh stores and replace it with smart carts that shoppers scan as they go. Its stated reason was what customers wanted: to find nearby products and deals, see a running receipt and know how much they had saved⁸. Press coverage the same week, citing The Information, reported that in 2022 about 700 of every 1,000 Just Walk Out sales had gone through human review by a team in India. Amazon disputed that account, saying the associates’ main role was to label video so the model could improve⁹. Amazon did not withdraw the technology. It kept it in its smaller stores and in third-party venues where baskets are small and speed is worth most, about 170 locations such as airports, stadiums and hospitals by August 2024, and announced a new model meant to raise accuracy and lower the cost of installing the system. The trade press noted that grocery retailers had largely chosen cheaper self-checkout instead¹⁰.

Amazon has not published the economics of either setting, and the review figure is reported and disputed, so the story should be read as a pattern rather than a verdict. Read through the module’s lens, the public record fits the three tests closely. In a large grocery store, each basket is long and varied, so whatever human checking a hard case needs, Data, Integration and Operational Costs showed how quickly it can swamp the model; that is the first test. The cameras and sensors are fixed cost per store, and a cheaper alternative existed; that is the second. And the response was not a single keep-or-kill decision. It was a change of route by setting: smart carts where the system did not fit, the original technology where it did, and investment to lower its unit cost. That is the third test working as it should, on actual figures, while the capability was still in use.

What this means for leaders

The practical consequence of this module is a change in what an AI approval is. It is not a verdict on a business case. It is the start of a contract between the sponsor, finance and the team: here is what one outcome costs and is worth, here is the volume that pays for the whole life, here is who watches, and here are the numbers that will bring us back to this table. A capability that cannot state those terms is not ready for funding at scale, however good the pilot looked.

Two habits make the contract work. The first is to keep cost and value on the same page, so that no one can cut cost without seeing what it does to outcomes, or claim value without seeing what it costs. The second is to treat a fired trigger as routine. The shipping line’s second year is not a scandal; it is what the economics of live AI systems look like.

Check yourself

1. If token prices keep falling, the cost of an AI workflow will fall too.MythFact
2. A capability whose value per outcome exceeds its cost per outcome is economically sustainable.MythFact
3. Break-even volume should be calculated on cost per resolved outcome, not cost per request.MythFact
4. A system that works better than before can still have worse economics.MythFact
5. An annual review is enough to keep an AI business case current.MythFact
6. In 2024 Amazon stopped using its cashierless checkout technology altogether.MythFact

Reflection: write the terms

What comes next

This module has given you a way to judge whether an AI capability can pay for itself and keep doing so. That judgment assumes the organization can deliver the capability in the first place: that the data, the skills, the governance and the operating rhythm exist to build it, run it and watch it. Module 09 turns from economics to the roadmap, and it starts with that assumption. The next chapter, Assessing Enterprise AI Readiness, asks what an organization must have in place before its AI ambitions, however sound their economics, can be delivered.