Looking back at 2025. Why companies with a high personnel cost share cannot sustain a price war.
In 2025 insolvency numbers kept rising and so did wages. At the same time it became much cheaper to have routine work done by programs. A look back at the year in which the personnel cost share became the decisive metric, with an honest worked example and three things you can check for 2026.
Fabian Weiss Key Takeaways
- A company with a high personnel cost share cannot follow a competitor's price cut. That is not a technology problem but a question of cost structure.
- Many home-built automations from 2023 and 2024 failed in 2025 for lack of version control, error handling and monitoring.
- The human role is shifting from approving every single transaction to reviewing the exceptions. That is the lever, not the AI itself.
What shifted in 2025
December 2025. The number of business insolvencies in Germany has been rising since 2023 and in 2025 it was again above the previous year. In November the Federal Statistical Office reported 5.7 percent more regular insolvency filings than in November 2024. Wages kept rising and skilled staff remained scarce.
At the same time it became much cheaper this year to have routine work done by programs. Read text, sort it, move it into another system. In 2023 that was still an experiment for the marketing department. In 2025 it became a normal tool.
Together these two movements changed something. Efficiency was an advantage for a long time. For companies with thin margins it became a condition in 2025. I call this the great filter. The term is mine and not from a study. It describes that the question is no longer whether you automate but whether your competitor does it before you.
The point in one sentence
A company with a high personnel cost share cannot follow a competitor's price cut and in 2025 it became much easier for that competitor to lower its share.
Why cost structure decides the price war
A worked example. It is constructed and simplified but it shows the mechanism. Two companies, both with 10 million euros in annual revenue.
| Item | Company A | Company B |
|---|---|---|
| Personnel costs | 6.0 million euros (60 percent) | 4.0 million euros (40 percent) |
| Technology and automation | 0 | 0.5 million euros (5 percent) |
| Other costs | 3.5 million euros | 3.5 million euros |
| Earnings before interest, tax and depreciation | 0.5 million euros (5 percent) | 2.0 million euros (20 percent) |
Company B cuts its prices by 10 percent to win market share. Its revenue falls to 9 million euros while costs stay at 8 million. It still earns 1 million euros, around 11 percent. Company A follows. Its revenue falls to 9 million while costs stay at 9.5 million. It makes a loss.
Company A does not have a technology problem. It has a cost structure problem that only becomes visible when someone else sets the prices. The example is exaggerated. Whether a company can lower its personnel cost share from 60 to 40 percent depends heavily on the industry. A trades business cannot. A company where a large part of the work consists of moving, screening and forwarding data often can. Why the cost of that kind of work has fallen so far is explained in the article on cognitive deflation.
Why many home-built automations failed in 2025
A pattern we kept meeting in projects in 2025: automations that were quickly assembled in 2023 or 2024 with a click tool like Zapier. That worked as long as the processes stayed simple. With every additional step the structure became more fragile.
A typical case: a vendor changes its interface and a process stops without anyone noticing. An employee deletes the Google Sheet that served as a database. The orders in it are gone. There is no version control, so no way to restore an earlier state. There is no error handling, so no plan for a step failing. And there is no monitoring, so nobody sees the standstill.
We call this shadow automation, after shadow IT. A considerable part of our work in 2025 consisted of replacing such processes with software that is versioned, monitored and run on a server of your own. A tool like n8n in your own environment is often a good start, as the guide to self-hosting n8n shows.
The takeaway for 2026 is unspectacular: automation is software development. It needs the same rules. Treat it like a hobby project and you get a hobby project that eventually shows up in the month-end close. Why a chatbot without integration is the same problem in different packaging is covered in a separate article.
How the human role is shifting
The second shift is quieter but more important. Until 2024, automation with AI mostly looked like this: the program makes a suggestion and a person clicks approve. The person was part of the assembly line and the line was no faster than they were.
In 2025 that turned around in many projects. The program handles the transaction itself and only presents the cases where it is not sure. The person reviews exceptions instead of routine cases. A typical case is the incoming invoice: the process reads sender, amount and order number, matches them against the order and posts it. Only when something does not fit does the invoice land with a person.
The lever is not the AI but the threshold. How many cases a program may decide on its own is a management decision, not a decision of the model. In projects we start with a low threshold and raise it once the logs show the decisions are correct. Why revenue per employee is the more honest metric than headcount is explained in the article on revenue per employee.
What the AI Act actually required in 2025
The AI Act, the EU regulation on artificial intelligence, has been in force since August 2024 and applies in stages. In 2025 the first obligations applied: since February the bans on certain applications and the duty to ensure that staff using AI are trained for it. Since August the rules for providers of large foundation models apply. Most further obligations follow from August 2026 and in November 2025 the European Commission proposed postponing part of them.
For a company that uses AI only in internal processes, 2025 was therefore not a year of panic. It was a year in which one question paid off: can I trace why every automated decision was made the way it was?
We decided early on not to use language models as a black box but with source references. The program answers based on your documents and states where the answer comes from. If it finds nothing, it says so instead of guessing. That is not a trick but architecture. And it is the part that counts in an audit. Why the human still remains the most common attack path is covered in the article The firewall is me.
What you can check for 2026
Three things that can be checked without a consultant and without a forecast.
- What is your personnel cost share and which part of it goes to moving, screening and forwarding data? That is the part a competitor can lower.
- Which automations run in your company without version control, without error handling and without monitoring? Each of them is a silent outage that has not happened yet. Why old systems become a burden is explained in the article Legacy is liability.
- Which software licences expire in 2026 and which of them do you pay per head for work that a program now does? Before renewing, it is worth checking whether the function is reachable through an interface.
How hard this hits you depends on where you start. A company with a 20 percent margin has time. One with 5 percent does not. If you want to know where your biggest lever is, talk to us.
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.