Apollo Global Management chief economist Torsten Slok wrote on Oct. 24 that outside the technology sector, AI has delivered zero financial payoff. Revenue per worker and profit margins are rising at the Magnificent 7 and flat across the other 493 companies in the S&P 500. Outside tech, he writes, token spending is a cost center weighing on margins. The AI story “will be decided in the S&P 493, not the Mag 7.”

Slok has the beneficiaries right. Thus far, the value has sat with the companies selling AI infrastructure, not the enterprises consuming the technology. The Oct. 28 earnings reports made the point: for the quarter ended Sept. 30, Azure revenue grew 48% and Google Cloud 103%, driven in large part by AI demand. The buyers of that capacity reported job cuts, not margin gains. The more interesting discussion is what is happening inside those companies that keeps their spend from converting to realized gains, what has to change, and when it will.

The payoff is landing in people, not in companies

Surveys this year from the St. Louis Fed and the Census Bureau put work use of generative AI at about 40% of workers and adoption at about a quarter of firms. Individuals are running ahead of their employers. Employees are saving real hours with AI. Some teams, and sometimes whole divisions, are seeing real benefit too. But those gains are not adding up across the enterprise, because the core operating model has not changed. The workflows, the processes and the decision approvals were built around people doing the work, and the reviews, handoffs and sign-offs are still there even after AI has taken over the task. In an AI-enabled operation, a process built for human pacing becomes the constraint, which is why the gains stay in pockets instead of reaching the whole company.

What changing the operating model actually means

Executives nod at “operating model.” Boards ask what it means in practice. Three things rise to the surface, and each is hard to enact in the companies Slok is describing.

First, AI must be a CEO-owned initiative. That does not mean sponsorship. It means one accountable party for building a new operating system and making the difficult calls that come with it: declaring the outcome, setting the pace, redesigning the work, deciding how much trust to extend to a model, and choosing what gets measured. Nobody below the CEO carries the same authority to enforce those decisions in every function. The most common pattern I see in enterprises not realizing value is the opposite of ownership: the CEO has delegated AI to IT or a transformation office, approved a budget, and is waiting for a result that the organization has no mandate to produce.

Second, picking the right problems and rebuilding the work around them. The filter is whether a workflow is material, not whether it is interesting: does it move revenue, margin, risk or cycle time, and will one business leader own the outcome. Pass that filter and the hard work begins: deconstructing workflows to eliminate dated processes, redefining roles so that work once performed, reviewed and approved by people alone is shared with AI without losing accountability, resetting approval authorities to the point where judgment is now applied, and establishing a governance model that operates at runtime, because AI produces output at a volume and speed that no manual review can keep pace with. That is where most of the time goes.

Third, the AI itself. Slok is right that most companies do not need the latest model, and enterprises increasingly understand that. Less understood are the layers beneath that conversation. Start with whether an AI model is needed at all: significant improvements are available through deterministic automation that most organizations have never pursued, while many are using frontier models for routine tasks. Then comes infrastructure. Increasingly I am discussing the optimal mix of open and closed models, which revives the considerations of the cloud versus on-premises era with one difference: cloud took a decade to become multi-cloud, and model strategy is multi-vendor from the start. Finally, selecting a model is not a one-time event. That capability is very rare in most organizations today.

None of this is a technology project. It touches every function and lives inside core processes.

Boards then ask how long. Every executive team reaches the same realization: this is more expensive than we thought. Not because of token prices but because the build costs more than it returns until it is done. BCG’s Applied AI Index puts AI spending at 3.3% of revenue now, up from 1.7% in late 2025. That is Slok’s cost center, seen from the inside.

Why headcount will not shortcut it

The temptation is to skip the build and cut staff. Barclays announced roughly 5,000 job cuts on Oct. 22 and cited AI, and markets read it as the payoff arriving. Barclays’ own chief executive called AI “the icing on the cake,” and the bank has not said how much of the number is attributable to AI at all.

Measure the cost of the work

The right number has two sides. On one, the quality of the outcome: did the work meet the KPI or service level the business set for it. On the other, the full cost of producing it: the model, the data, the orchestration, the validation, the human handling of exceptions, the rework and the governance, of which token spend is one line and rarely the largest. Enterprise AI will be measured this way, paying for outcomes accepted rather than intelligence consumed, and almost no company can produce that number today. Their dashboards show what is being used, by model, and nothing about what that spend returned. So the bill arrives before the value does, and the market reads layoffs instead.

When

Given the ownership, redesign and measurement work still ahead at most companies, I would not expect widespread acknowledgement of specific, measurable AI benefits before the end of 2027. Widespread gains across the 493 are at least a couple of years away.

Until then, four questions tell you where your company stands:

Cost per outcome. Do we know what AI work costs and returns, or only how many people it touched?

Adoption. Who is using the tools, and in which workflows? Seat counts are not usage.

Sequencing. Are we building the operating model before we cut, or after?

Model discipline. Are we paying frontier prices for routine work?

Slok’s flat line is the average of 493 answers to those questions. It will move when the answers do.

Adam Dumey leads enterprise frontier lab strategy at World Wide Technology (WWT), a privately held technology firm. WWT sells the enterprise AI integration work this piece recommends and partners with NVIDIA, Dell, Cisco and Microsoft. He holds no positions in the companies mentioned.

The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.