Thinking is getting cheap. What that means for companies that sell or buy brainwork.
An answer from a language model now costs a fraction of a cent. That shifts what a company does in-house, what it buys and what gets done at all. This article explains the mechanics without forecasts and says where they do not apply.
Fabian Weiss Key Takeaways
- The cost of one unit of brainwork from a machine has been falling for years. Anyone who sells thinking at a premium needs to check their pricing.
- Cheaper brainwork rarely means less work. It makes tasks affordable that used to sit in the backlog.
- When tasks can be assigned and ended in seconds, permanent employment loses its cost advantage for routine work. That does not apply to judgment and responsibility.
Two cost blocks have limited every company so far
The first is energy. It is limited by physics and raw materials. The second is brainwork. It used to be limited by people, by their training, their working hours and their number.
Industrialisation solved the first block. Muscle power and productivity have been decoupled ever since. Language models are solving the second one right now. Brainwork, at least the simple kind, no longer depends on the number of people in the office.
It is March 2026. An answer from a language model, say reading and summarising a document, costs a fraction of a cent. Providers have cut their prices per request several times since 2023 and the price lists are public. The same task done by an employee costs minutes of their time, so euros instead of fractions of a cent. The gap is so wide that the exact number does not matter.
What it means for you depends on how much of your work consists of reading, sorting, comparing and forwarding. In projects we often see that this share is larger than management estimates, because it is spread across many roles.
The point in one sentence
When simple brainwork costs almost nothing, less thinking is not the result. More is. The only question is who gets the tasks.
Why companies exist and what is changing about it
The economist Ronald Coase asked in 1937 why companies exist at all. Why not hire someone on the open market for every task? His answer was transaction costs. It is too much effort to find, brief and supervise a new bookkeeper every day. So we hire one permanently and trade flexibility for less coordination.
Exactly those costs are now shrinking for certain tasks. A program that talks to a language model needs no job ad. It needs no onboarding if it has access to the company’s records. It starts in seconds and stops when the task is done. For well-defined, recurring work the reason to create a position for it disappears.
That explicitly does not apply to everything. Whatever needs judgment, carries responsibility or talks to customers stays with people. The company of the coming years will therefore not be empty. It consists of a smaller core of people who decide. In between, tools handle the routine. What that looks like day to day is in our article on growth without new hires.
Cheaper means more work, not less
The obvious worry goes: if the machine thinks, there is nothing left for people to do. Economic history has a counterexample. William Stanley Jevons observed in 1865 that coal consumption rose when steam engines became more efficient. Steam got so cheap that it was used for things nobody would have paid for before. Today that is called the Jevons paradox.
The same happens with brainwork. An example from sales: a few years ago a personal proposal for every single prospect was too expensive. So everyone got the same standard email. Today a personal proposal, a personal follow-up and a personal summary of the last call cost almost nothing. So they get done.
In projects we therefore rarely see waves of layoffs. We see lists of tasks that sat untouched for years because they were too expensive. Now they are being worked through. What that looks like in candidate screening and in sales automation is described in separate articles.
How the cost structure shifts
Billing per task instead of per month or hour
Subscription software costs per month and user. Human labour costs per hour. A language model costs per request. If little happens in August, the automation costs almost nothing in August. In September it grows with the business without anyone having to be hired. The cost risk moves from fixed to variable.
Corrections stay in the system
An employee learns and takes what they learned along when they leave. In a well-built system every correction by an employee is stored and used as an example for the next cases. That does not happen by itself. Someone has to maintain those examples. But they are not lost when a colleague leaves. Why that matters for a company’s knowledge is in the article on company knowledge.
Many tasks at once
A person handles one task after another. A program handles as many at once as you allow. A typical case: checking freight documents. Before, the department’s daily capacity was the limit and peaks in the month meant overtime. Now the limit is how many documents arrive. The waiting time disappears, not the check.
| Question | Permanent staff | Language model with integration |
|---|---|---|
| How it is billed | Per month, fixed | Per request, variable |
| What happens with more work | New position, weeks of lead time | More requests, immediately |
| What happens on departure | Knowledge leaves | Examples and rules stay |
| Who carries responsibility | The employee | Still a person |
What this means for your decisions
The old question was: do core competence in-house, buy the rest. The new question is: what is routine and what needs judgment? Routine where someone reads, checks and forwards data can be automated. Judgment stays with people who know the context. The bookkeeper who only books gets competition. The bookkeeper who tells management what the numbers mean becomes more important.
Three steps you can take this quarter:
- Map the routine. Write down every process where someone reads, interprets and forwards data. Those are the candidates. How to do that without becoming dependent on one vendor is in Automation Without Handcuffs.
- Check licences before you renew them. For every piece of software, ask whether its function is reachable through an interface without user seats. Why old software becomes a burden is in Legacy Is Liability.
- Watch revenue per employee. This figure tells you whether the automation reaches the result or only shifts effort around. It is not an end in itself. Some companies stay the same size and do more with the same team.
Honestly, the effect depends heavily on your starting point. Whoever already has clean data and clear processes benefits quickly. Whoever has to create order first takes longer and the first gain is the order itself.
What you can check now
- Count on a normal day how often someone reads a document only to forward it or enter it somewhere. That is your share of simple brainwork.
- Take the list of tasks that have sat for years because they were too expensive. Check which of them are affordable now.
- Look at your software contracts for user seats you pay for work no person would need to do any more.
- If you want to work through an example from your own day, send us the process in three sentences. We will tell you honestly whether it is a fit.
Research for technical decisions
New reports, benchmarks and technical analyses on SaaS economics, AI engineering and owned infrastructure.
Original research Public sources No sales mail
By subscribing you receive new analyses and updates from FW Delta by email. You can withdraw your consent at any time. Further information is available in the privacy policy.