AI on the payroll line, not in the software budget
How much does AI cost and which budget does it belong to? If AI does work that used to be done by a human, its true point of comparison is not a software license but the payroll. A new “wage category” has emerged in the cost structure — one whose price flexes with usage, is investor-subsidized today, and requires the same monitoring as the payroll.
When we ask executive teams how much AI costs, the answer is almost always found in the IT or software budget. That is understandable, but it is also wrong.
We see the same “classification error” repeatedly, regardless of the organization. Most AI is currently bought in Finnish organizations like software: as a fixed monthly fee per user, from the IT budget, filed alongside email and antivirus. But unlike antivirus, AI does not support work – it does the work. It writes, analyzes, codes, and drafts decisions. When technology starts doing work, its price stops being an IT question and becomes a cost-structure question: which part of the price of work will be fixed salary going forward, and which a variable usage fee.
This is not merely a question of cost centers, nor an artificial exaggeration. It means that AI as a workforce changes your cost structure. Alongside traditional labor and software investments, organizations now need to manage a new kind of productive capacity whose cost is increasingly tied to usage rather than headcount. Its development must be tracked with the same seriousness as the payroll. This is a genuinely new situation brought about by AI.
Computers automated routine cognition – bookkeeping, calculation, filing – decades ago, as Autor et al. showed in their classic study (2003). What is new is that AI reaches non-routine, open-ended knowledge work: writing, analysis, coding, reasoning. Things that used to be beyond the reach of automation. Economists Agrawal, Gans and Goldfarb describe the same phenomenon as a collapse in the price of prediction: when the price of an input falls steeply, its use spreads to tasks where it did not pay off before. At the same time, the value of its complements – such as human judgment – increases.
How does the price of AI differ from a salary?
The price of AI differs from a salary in four ways: it flexes with usage, it is investor-subsidized today, no employer’s side costs are paid on it, and it is task-specific. The first difference is flexibility. A salary is fixed, negotiated and predictable; it does not flex downward in a quiet month without layoffs, nor scale upward without recruitment and onboarding. The price of AI does both. The cost moves with usage in both directions. The new wage category is already making the price of certain expert tasks resemble spot-market electricity more than a fixed hourly wage – and with electricity, the contract model is usually a deliberate choice.
The second difference is that the price may surprise you within just a few years. Today it is exceptionally low, and the reason is not production costs. According to leaked financial statements verified by the Financial Times, OpenAI’s operating loss in 2025 was about 21 billion dollars – on roughly 13 billion in revenue – and Sequoia’s David Cahn has publicly asked where the returns on the industry’s hundreds of billions in annual investment will ultimately come from. Token prices are indeed falling fast, but that is a price, not a cost. Our reading of the market is fairly blunt: we are in a phase where market share is being bought with investors’ money, and once capital starts demanding returns, the gap between price and true cost will close. This may not show up as rising token prices but as rising total cost, driven by growing usage and value-based pricing. An operating model built on today’s price may have been built on a temporary discount.
The third difference only becomes visible when you look at a salary through the employer’s eyes. A euro of salary is not the full personnel cost of work: on top of it come statutory pension and social insurance contributions, occupational health care, employee benefits and holiday pay. A common rule of thumb is to multiply the salary by 1.35 for a quick estimate of the personnel cost, and by 1.6 to know what an employee truly costs the company in total. For work done by AI, no multipliers are paid – the usage fee is in practice the entire cost. It has costs of its own, too – deployment, oversight and governance. But those are investments and management work, not a percentage multiplier on every euro. This is the change in cost structure at its most concrete. When you compare the price of a machine and a human in a task, the right benchmark is the full labor cost including side costs, not the nominal salary.
The fourth difference is the most important. The price is task-specific. In many tasks the machine is already cheaper than a human – but not in all of them. Moreover, the boundary moves constantly, because both price and capability keep changing. That is why the question “what does the machine do, what does the human do” is not a one-off outsourcing decision but a recurring assessment, much like a salary round. After all, no one thinks salaries are negotiated once and for all.
“Same money, different line – why does it matter where it’s booked?”
We hear this often, and it is a fair question. The answer builds on the fact that where a cost is booked determines who looks at it, at what rhythm, and above all what questions they ask of it. The payroll has an entire management machinery: budgeting, forecasts, comparisons, negotiation rounds, workforce planning. The software line has its machinery too. License counts get trimmed and prices negotiated, often quite skillfully. But the machinery of software management manages the price of a tool, not of work. It asks how many seats are needed and what a seat costs – not what work is being done with the spend, what is worth having done, and in which tasks a human would do it more cheaply.
Another quick conclusion deserves an equally direct answer: if the machine is cheaper, why not replace people with machines everywhere possible? The comparison lives at the level of tasks, not people. A job is a bundle of tasks, of which the machine takes a share and leaves the part requiring judgment to the human. Whoever prices their people out at today’s subsidized price makes a permanent decision at a temporary price. Lasting competitive advantage lies in people and in how quickly they learn – including how to direct and manage AI agents. Not to mention human skills.
Three questions for finance leaders
We encourage finance leaders to ask these right away:
- Who tracks the price of AI, and at what rhythm? The payroll is forecast, benchmarked and negotiated regularly. If the new wage category has no owner and no rhythm, it is not under control – it is adrift.
- In which tasks have you compared the unit cost of machine and human – and when will you compare again? Compare against the full labor cost including side costs, not the nominal salary. A one-off comparison goes stale, because both price and capability keep moving. Give the assessment a recurring slot, for example alongside the budget round.
- What happens to your operating model if the price of this wage category doubles when the subsidy ends? If the answer is “we don’t know”, the plan has been built on a discount.
Cost structure is one of the four reasons why every strategy question is ultimately an AI question – alongside skills, workflows and governance. And beneath them all lies the same foundation: lasting advantage does not come from the cheapest labor but from people, and from how quickly they learn to decide which work belongs to the machine and which to the human.
Astu Labs helps executive teams model AI as a workforce: what the new wage category costs now, how its price behaves, and in which tasks the machine beats the human – and where it does not. If you would like to see what your numbers look like from the payroll line, get in touch – we respond within one business day.
Sources
- Agrawal, A., Gans, J. & Goldfarb, A. (2018). Prediction Machines: The Simple Economics of Artificial Intelligence. Harvard Business Review Press; and “The Simple Economics of Machine Intelligence”, Harvard Business Review, November 17, 2016.
- Autor, D., Levy, F. & Murnane, R. (2003). “The Skill Content of Recent Technological Change: An Empirical Exploration.” Quarterly Journal of Economics 118(4), 1279–1333.
- OpenAI’s financial statements for 2025: leaked, audited documents whose figures the Financial Times has verified (reported in June 2026). Operating loss approx. $20.9 billion, revenue approx. $13.1 billion.
- Cahn, D. (2024). “AI’s $600B Question.” Sequoia Capital, June 2024.
- Epoch AI (2025). Trends in language-model token prices.
