Chris Skinner
There is a rather awkward conversation beginning in boardrooms about artificial intelligence. For the last few years, the question has been what AI can do, how quickly companies can deploy it, and how many people it might replace, but the question now landing on the CFO’s desk is much simpler: how much is this stuff costing us?
The answer is becoming uncomfortable.
One of the more striking examples comes from McKinsey, a firm making substantial revenues advising corporations about AI transformation.
Releasing their latest survey about AI usage across industries, they find quite a few anomalies:
- 20% of organisations say AI operating costs, including tokens, are already constraining AI usage.
- 28% are spending more than 10% of their entire technology budget on AI, yet 60% still expect AI investment to increase over the next year.
- Nearly nine in ten organisations use AI, and 44% are now scaling it across the enterprise, but only 37% report a positive EBIT contribution.
- Crucially, only 14% actually reduced their workforce because of AI last year, despite 32% having expected reductions.
- It’s well worth a read.
What I found particularly interesting is that McKinsey’s very own CFO Eric Kutcher has warned that the rate at which the firm’s AI expenditure is increasing cannot continue, saying there is “no way” the company can afford another year of growth at the same rate.
That is McKinsey’s people’s usage of AI to assist in reporting to clients.
McKinsey reportedly uses AI for as much as 30% of its tasks, which means the problem is not failed adoption but successful adoption.
The bottom line is that the more people who use AI, the bigger the bill becomes.
That reveals one of the fundamental differences between AI and the previous generation of enterprise software.
Traditional software was largely predictable. Buy 10,000 Microsoft licences and the CFO knew roughly what the annual bill would be, whereas AI increasingly operates like electricity, cloud computing or mobile data because you pay according to consumption.
Every prompt consumes tokens, every answer consumes tokens and, critically, every AI agent working autonomously in the background can consume thousands or millions of tokens without the employee who started the process seeing what is happening.
Agentic AI makes the economics even harder because one instruction from a human can trigger dozens of machine-to-machine actions, queries and decisions.
Goldman Sachs forecasts cited in the reporting suggest token consumption could increase 24-fold between 2026 and 2030, meaning companies are heading towards an extraordinary expansion in AI consumption even as the price of individual tokens continues to decline.
This creates the great AI paradox: AI is becoming cheaper and AI is becoming more expensive at the same time.
The unit cost of intelligence is collapsing, but the quantity of intelligence companies consume is exploding.
The experience of Uber illustrates the problem.
Uber exhausted its entire 2026 AI coding budget in four months, despite achieving extraordinary adoption among its engineers. By March, 84% of engineers were using Claude Code and around 70% of committed code originated with AI, but Uber executives found that token consumption did not translate directly into useful features delivered to customers.
Microsoft has faced similar concerns over AI coding costs, while Nvidia’s vice-president of applied deep learning has said that his team’s compute costs now exceed the cost of the employees using that compute.
AI costs more than people!
That turns the original AI employment argument upside down.
There is also a deeper problem because the AI invoice is only one part of the bill. The true enterprise cost includes preparing and cleaning data, rebuilding data architectures, cloud infrastructure, GPUs, specialist engineers, security, compliance, model monitoring, integration with existing systems and continuous maintenance.
Riseup Labs estimates that data preparation alone often consumes 30% to 50% of an AI project’s budget, while ongoing maintenance can consume another 15% to 30% of the original implementation cost every year. Production infrastructure can easily reach tens of thousands of dollars per month before the organisation begins counting the people required to operate it.
Then there is integration.
AI sitting in a browser is cheap. AI embedded inside the operational nervous system of JPMorgan, HSBC and Deutsche Bank is not. It has to understand customers, products, permissions, regulation, workflows, data, legacy systems and decades of accumulated corporate complexity, which means that the expensive part of enterprise AI is often not artificial intelligence at all but making the enterprise intelligible to artificial intelligence.
Banks will experience this more intensely than almost any other industry because their technology estates contain decades of legacy infrastructure combined with enormous regulatory, security and data obligations. Putting a clever chatbot on top of that architecture is easy. Rebuilding the bank so that intelligent agents can safely operate inside it is a completely different proposition.
There is another uncomfortable factor hidden beneath today’s prices because the AI industry is still subsidising adoption.
The largest AI providers are spending extraordinary amounts on infrastructure while competing aggressively for customers, meaning today’s enterprise AI price does not necessarily represent the long-term economic cost of providing the service.
Forbes argues that normalisation of AI pricing could push enterprise bills substantially higher, particularly as providers move heavy users away from flat-rate subscriptions towards consumption-based charging.
In other words, corporations are worrying about the AI bill while AI remains comparatively subsidised.
That should concentrate the CFO’s mind.
But it leads to the more important question … is it worth it?
Yes.
The mistake is believing that the business case for AI is simply replacing people.
If a bank employs 100 people costing £10 million and replaces them with AI costing £12 million, then, viewed through the traditional cost-reduction spreadsheet, AI has failed. Yet that calculation ignores what happens if those AI systems perform ten times as much work, operate twenty-four hours a day, respond instantly to customers, analyse every transaction simultaneously, detect fraud continuously, write software faster, personalise millions of interactions and allow the remaining humans to concentrate on decisions where human judgement matters.
The denominator has changed.
We should not measure the economics of AI simply as AI cost versus people cost. We need to measure AI input versus economic output.
This is where many organisations are getting the argument wrong because they are adding AI to existing organisations rather than redesigning organisations around AI. Giving every employee an expensive frontier model and encouraging them to use as many tokens as possible is not transformation. It is adding another technology expense to the P&L.
The answer is architecture.
Companies will route simple work to small, cheap models, complex reasoning to powerful models, deterministic processes to conventional software, and human judgement to humans.
There is no point in using Einstein to summarise an email.
AI economics will increasingly depend upon selecting the cheapest intelligence capable of completing each task rather than throwing the world’s most powerful model at everything.
This is the same journey computing has taken repeatedly.
We did not abandon cloud computing because early cloud bills were unpredictable. We created FinOps, monitoring, workload optimisation, and sophisticated infrastructure management. AI will go through the same process, except this time we will develop something closer to Intelligence Economics, where companies measure the cost of machine intelligence against the economic value generated by that intelligence.
This is where the CFO is focused most.
A CFO wants a chief of staff who prepares board materials, reads documents, summarises meetings, remembers actions, prepares daily briefings and keeps track of decisions. Hiring another senior employee creates salary, benefits, recruitment and management costs, whereas an AI system connected to the relevant corporate information can perform large parts of that function continuously.
Multiply that concept across an organisation and the economics become far more interesting.
The future bank will have AI credit analysts, AI compliance assistants, AI fraud investigators, AI software engineers, AI customer-service agents, AI treasury assistants and AI relationship-management copilots working alongside humans. Eventually those agents will begin communicating and transacting with other agents, creating an organisation whose productive capacity is no longer constrained by its human headcount.
That is where the productivity revolution lies.
The problem today is that many companies are measuring AI using twentieth-century management accounting.
They are asking how many jobs AI eliminates when they should be asking how much additional economic activity, revenue, and results it creates.
There is also an important historical parallel.
The internet produced one of the greatest investment bubbles in history because companies correctly understood that the internet would transform the world but wildly misunderstood how quickly the economics would work. The dot-com crash did not prove the internet was wrong. It proved that price, timing, and business models mattered.
AI is heading towards the same reckoning.
There will be wasted billions, abandoned projects, spectacular corporate mistakes and companies discovering that their shiny AI strategy consists of an enormous monthly token bill and some slightly better PowerPoint presentations. There will also be companies that redesign their operations around machine intelligence and achieve productivity levels their competitors cannot match.
That creates the dividing line.
The AI winners will not be the companies that spend the most on AI. They will be the companies that extract the greatest economic value from every unit of intelligence they buy.
For banks, this is particularly important because AI will not simply reduce the cost of banking. It will change the productive capacity of the bank itself.
A bank employing 100,000 people today will not need 100,000 people to produce tomorrow’s level of economic output, but that does not mean the objective is to replace 50,000 employees with 50,000 digital workers. The objective is to create an institution capable of doing several times more with a radically different combination of human and machine intelligence.
That is why the CFO is right to worry about the AI bill and the CEO is right to keep investing.
AI 1.0 was about adoption. AI 2.0 is about economics.
The winners will discover that the most important question was never about how much does AI cost, but how much is intelligence worth?