Banks have spent years investing in artificial intelligence to make their operations more efficient. Now they’re building teams to track what their AI agents are spending.
A November Reuters report describes banks dealing with the rising costs of AI agents. Agents work around the clock, exchanging information and carrying out tasks. Each instance consumes computing resources. McKinsey estimates that, in some scenarios, the tokens powering these systems could cost as much as 20 percent of labor expenses.
At almost the same time, researchers at Epoch AI reported something extraordinary. Since 2023, the cost of achieving comparable performance across selected AI benchmarks has fallen approximately 47 percent every quarter. That translates to a thirteenfold decline annually.
The price of measured AI performance is collapsing. The bill for using it can still be enormous.
When Expertise Becomes Abundant
For generations, law firms, consulting companies, and accounting practices have organized themselves around expertise that takes years to acquire. Experienced professionals command higher fees. Junior employees perform much of the research, analysis, and preparation. Firms build reputations around knowledge clients cannot easily reproduce.
AI is beginning to change the economics of that work. Industry research, financial analysis, market comparisons, and sophisticated presentations can increasingly be produced with fewer people.
For a small consulting practice, that could open opportunities previously beyond its reach.
Competing against a global firm is another matter. The global firm already has relationships with client leadership, access to proprietary information, and a reputation that may reassure a board of directors. It can mobilize resources across countries and absorb the financial risks of a complicated engagement.
The smaller practice may now produce comparable analysis. Winning the assignment is another matter.
The cost of AI performance can fall dramatically within months. Building client relationships, establishing a reputation, and gaining access to markets can take years.
An established firm may retain its market position long after the expertise that helped create its advantage becomes widely available.
The Value of Difference
Kevin Kelly recently asked a provocative question. If everyone had an AI assistant that could make almost anything they wanted, why would anyone need to buy something someone else had made?
In his recent essay, New Units of Knowledge, Kelly explores what happens when creation becomes abundant. He suggests that meaningful differences will stay valuable.
Lower production costs have historically opened markets to new participants and created opportunities that were difficult to imagine when the underlying technology was expensive. AI could enable individuals and small organizations to develop products, services, and specialized capabilities that once needed considerable investment.
Kelly also considers how specialized AI systems might collaborate and exchange knowledge, opening up possibilities beyond what any one system could accomplish.
An entrepreneur might build sophisticated software without assembling a conventional engineering organization. The software still needs customers, infrastructure, and access to the platforms it runs on.
Some of these barriers may also become less expensive through AI. Others may become more consequential as more people gain the ability to produce competing offerings.
Where the Value Goes
Amazon opened a national marketplace to independent sellers who could never have afforded a comparable retail network. By 2025, independent businesses accounted for more than 60 percent of sales in its store.
Reaching those customers, however, came at a price. Sellers paid commissions and fulfillment fees, and many spent additional money on advertising to make their products visible. In a 2023 antitrust complaint, the Federal Trade Commission alleged that Amazon’s practices increased costs for sellers and restricted competition.
Amazon’s own financial statements reveal another side of the relationship. In 2025, the company generated more than $172 billion in revenue from services provided to third-party sellers.
AI may produce a different outcome. Competition among providers could prevent durable concentrations of power. Open-weight models and falling computing costs could reduce dependence on today’s dominant companies.
Yet the infrastructure supporting advanced AI demands considerable investment in computing facilities, specialized chips, electricity, and networks.
The companies providing that infrastructure may gain influence even as the capabilities themselves become cheaper.
More people may be able to create valuable products and services, while a relatively small number of organizations control the systems through which they reach markets.
The Cost Beyond the Answer
An AI agent may perform an individual analytical task at very low cost. But a bank rarely needs only the answer.
The bank must still meet regulatory requirements, protect customer information, maintain records, and account for errors.
In McKinsey’s analysis of certain banking customer-service workflows, tokens account for roughly 20 to 25 percent of an AI agent’s variable operating costs. Human monitoring accounts for 70 to 75 percent.
A medical practice may have access to sophisticated analytical tools at a fraction of their former cost, but it still needs clinicians to evaluate the results, meet regulatory requirements, and earn patients’ trust. A small business can now produce professional-quality marketing materials for very little money. Getting those materials in front of potential customers is another matter.
The cost of producing an answer is only one part of the cost of putting that answer to work.
A Different Distribution of Power
In Powershift, Alvin and Heidi Toffler examined how knowledge was changing the relationship between wealth and power. Unlike money or force, knowledge could be shared and reproduced without the original holder having to surrender it.
But spreading knowledge and redistributing power were never the same thing.
Some established economic advantages may weaken as AI spreads. New competitors may emerge. But regulations, established relationships, and control over essential infrastructure can keep shaping who enters a market long after the technology itself has changed.
A person who once needed a team of specialists may soon be able to accomplish much of that work alone.
Yet creating something and finding a place for it in the economy remain different challenges.
We may be making intelligence extraordinarily inexpensive while increasing the economic importance of the relationships, infrastructure, and market access needed to put it to work.
The cost of creating may fall much faster than the cost of participating.




