The AI pilot is the easy part.
A convincing demonstration can now be built in an afternoon. Summarise a document. Draft an email. Analyse a spreadsheet. Put a chatbot on a handful of PDFs.
Then the difficult phase begins: productive operation.
A recent Swiss survey makes the gap visible. Fifty five per cent of companies already use AI deliberately in at least one business area. Thirty one per cent remain in the pilot phase. Only 9 per cent report that AI has changed their business model. Swiss SME Portal
The obstacle is no longer access to technology. It is the organisation around it.
Which data are reliable? Which processes may be automated? Where must a person approve the result? Who monitors quality? Which systems are connected? Which roles need new skills? How will economic value be measured?
Infrastructure comes next. According to the same survey, data quality and data silos are among the biggest barriers. Security and data protection concerns follow close behind.
That may make 2026 the year after the first wave of AI enthusiasm. The technology now has to enter real processes, and that costs money. Sixty four per cent of Swiss CFOs surveyed by Deloitte accordingly expect technology and IT spending to rise. Swiss SME Portal
A pilot answers whether the technology can do something. Scaling answers a harder question: Can our company use it reliably, securely and economically?
Category: Innovation
We are recruiting skilled workers for tasks we could have eliminated long ago.
The debate about skills shortages focuses almost entirely on recruitment: advertisements, salaries, employer branding, training and immigration. All correct, necessary and well documented. [Evidence: current skilled-labour shortage index or SECO labour-market data]
The other half of the answer receives much less attention. When people are scarce, what matters is not only how many the company finds, but how much of their time genuinely creates value and how much is lost to searching, waiting, duplicate entry, setup processes, rework and unplanned journeys.
In many companies, a double-digit percentage of working time is lost this way. That is not employee negligence but the result of processes and tools that evolved separately and were never considered together. Each friction point is too small to fix on its own. Together, they amount to a position the company would not need to fill.
Modern equipment, digital scheduling, properly equipped vehicles and ergonomic tools therefore work twice. They increase output per employee without requiring anyone to work faster, an important distinction often lost in the productivity debate. And they change how the workplace feels.
That creates an uncomfortable budget issue: investment in tools is investment in employer quality. We simply do not account for it that way. It therefore competes with other capital expenditure rather than with the cost of an unfilled position. That cost is substantial: overtime, declined orders, pressure on the existing team and months of recruitment effort.
Combining the two calculations produces different decisions. Financing also helps: a monthly instalment can be compared directly with the capacity gained. A one-off investment, by contrast, enters a budget pot where it loses against everything else.
An installation company leases modern vehicles and digital scheduling tools so its technicians spend less time in the office and more time with customers.
How many hours a week does your best team lose to avoidable friction, and what would one fewer position be worth?
The most interesting AI question may soon no longer be which model we use. It may be where our data resides.
A survey published on the Swiss SME Portal in July 2026 reveals a telling tension. Eighty nine per cent of respondents already use AI professionally. Fifty one per cent consider compliance with Swiss or European data protection standards important. And 29 per cent may use AI through private accounts. Swiss SME Portal
That shows how quickly technology spreads and how slowly governance can follow.
An executive team can no longer stop at purchasing AI licences. It must decide on a data architecture.
Which information may enter external systems? Which information must remain internal? How are confidential customer data handled? Who controls accounts and permissions? Which models may access which corporate knowledge? What happens to prompts and outputs?
Those questions sound technical. They concern intellectual property, customer protection, compliance and competitiveness.
They also affect investment. A company may need a different cloud architecture, a private AI environment, new interfaces, stronger access controls or a structured knowledge system that gives employees secure access to information.
Data sovereignty is becoming infrastructure. Infrastructure belongs on the investment agenda.
It should not arrive there only after the first incident demonstrates what was missing.
Most AI projects do not fail because of AI. They fail because of Excel.
In almost every discussion about artificial intelligence, someone eventually asks which data the model is actually supposed to access. Then the room goes quiet.
The uncomfortable truth for many mid-sized businesses is that production data is on paper or in the shift supervisor’s head. Inventory is recorded in a spreadsheet maintained by exactly one person. The machines provide no data because they date from a time when nobody wanted any. And the ERP system was last seriously addressed when the predecessor’s predecessor was still in charge.
Nobody likes to discuss this because it sounds like the company has fallen behind. Yet it is simply the normal state of an economy whose substance lies in machinery and expertise rather than data architecture.
AI is not software that you simply install. It is the top layer of an infrastructure that must work beneath it: ERP, sensors, connected equipment, data quality and defined processes. Without that foundation, even the best model produces precise answers based on incorrect figures, and nobody notices because the answer is well written.
That is the real investment, and it is not glamorous. Nobody gives a keynote about barcode scanners. That is precisely why it is postponed: it is difficult to market as innovation and easy to classify as “later”.
The practical answer lies in sequencing, not abstention. These preliminary investments can be staged. Each stage must create value in its own right. Telematics saves kilometres before any algorithm joins the calculation. Financing with predictable instalments makes this staging budgetable in the first place because it avoids a new capital request every year.
A logistics SME leases telematics systems and modern scanners. Only then does AI-assisted route planning become a realistic option rather than a presentation slide.
An honest question for the room: how many AI workshops have you completed, and how many interfaces have you built in the same period?
Most AI projects do not fail because of AI. They fail because of Excel.
In almost every discussion about artificial intelligence, someone eventually asks which data the model is actually supposed to access. Then the room goes quiet.
The uncomfortable truth for many mid-sized businesses is that production data is on paper or in the shift supervisor’s head. Inventory is recorded in a spreadsheet maintained by exactly one person. The machines provide no data because they date from a time when nobody wanted any. And the ERP system was last seriously addressed when the predecessor’s predecessor was still in charge.
Nobody likes to discuss this because it sounds like the company has fallen behind. Yet it is simply the normal state of an economy whose substance lies in machinery and expertise rather than data architecture.
AI is not software that you simply install. It is the top layer of an infrastructure that must work beneath it: ERP, sensors, connected equipment, data quality and defined processes. Without that foundation, even the best model produces precise answers based on incorrect figures, and nobody notices because the answer is well written.
That is the real investment, and it is not glamorous. Nobody gives a keynote about barcode scanners. That is precisely why it is postponed: it is difficult to market as innovation and easy to classify as “later”.
The practical answer lies in sequencing, not abstention. These preliminary investments can be staged. Each stage must create value in its own right. Telematics saves kilometres before any algorithm joins the calculation. Financing with predictable instalments makes this staging budgetable in the first place because it avoids a new capital request every year.
A logistics SME leases telematics systems and modern scanners. Only then does AI-assisted route planning become a realistic option rather than a presentation slide.
An honest question for the room: how many AI workshops have you completed, and how many interfaces have you built in the same period?
A weak process can become faster without becoming better. Automating it may increase the speed at which its weaknesses affect customers. The case for artificial intelligence therefore begins with understanding the task, rather than celebrating the tool.
For leasing businesses, potential applications include document handling, data checks and analysis. Whether those applications create value depends on the quality of the information, the suitability of the task and the ability to verify results. Time saved is only part of the calculation.
FINMA addressed governance and risk management for artificial intelligence in its supervisory guidance 08/2024. The leadership question is straightforward: who owns an application when its recommendation is wrong, and who has the authority to respond?
Putting a person at the end of a process may not provide effective oversight. That person needs information, competence and time to challenge the output. If the workflow encourages automatic acceptance, a formal review step may offer less protection than it appears to provide.
I would start with a problem. Define the desired result, the errors that matter and the conditions requiring intervention. Compare the application with an existing process. A useful test should reveal weaknesses before they become repeated decisions affecting customers or the balance sheet.
External dependencies deserve attention too. What happens if the provider changes the model, cannot explain a result or becomes unavailable? Swiss financial businesses should understand how operations continue and which responsibilities remain inside their own organisation.
As a CEO, I can delegate tasks. I still need to ensure that the process has accountable owners and a workable response to failure. A purchase agreement with a technology provider does not resolve those management questions by itself.
The strongest case for adoption includes an account of its limits.
Which AI supported decision could you explain and correct today if challenged?
Leave a Comment