The Bottleneck Is No Longer Building. It's Deciding.
In short
The most important leadership question is no longer 'Can we build it?' It is 'What should we build, for whom, in what order?' Job market data shows AI agents are turning implementation into a commodity. Deciding is the new bottleneck. This analysis explains why Swiss leaders must own the what.
The most important leadership question is no longer 'Can we build it?' It is 'What should we build, for whom, and in what order?' The job market data in Lenny Rachitsky's 'State of the tech job market: late 2026' analysis, based on TrueUp and Live Data Technologies, shows a tectonic shift. Open product manager roles are at their highest level in nearly four years: 75% above the March 2023 trough and 56% above early 2023. Demand for classic implementation work is barely moving.
Engineering demand is plateauing. Open engineering roles grew 29% in 2024, 15% in 2025, and just 2.5% in the latest year. In the US, roughly 113,000 software engineers were hired over the past twelve months, less than half of the 239,000 peak in 2022. Product managers are the only one of the five core roles whose share of hires grew, from 16% to 21%. Building is becoming a commodity; deciding is becoming scarce.
+75%
Product manager roles above the March 2023 trough
2.5%
Growth in open engineering roles over the latest year
4.6x
More open roles for forward-deployed engineers
What the data reveals about the new value chain
Two more findings point in the same direction. LLM (large language model) is now the most frequently named skill in product manager job postings. Claude sits at number three, with Cursor, OpenAI and MCP in the top eight. People who decide on products and processes need to understand what models can do, not necessarily how to code them. For engineers, LLM is about to overtake Java as the most requested skill.
Forward-deployed engineers, builders who solve problems directly at a client site, are exploding: the number of open roles rose 4.6 times in a year to roughly 1,500. And the biggest AI hiring boom is not in classic software but in hardware, infrastructure and applied fields. The message for executives is simple: implementation can be bought, rented or outsourced. Direction cannot.
The new bottleneck
Once models, copilots and agents turn implementation into a commodity, the bottleneck shifts from building to deciding. That decision capability is the core competence of leadership. Hand it over, and you hand over the direction of your company.
For Swiss SMEs: the bottleneck is not technology
For a Swiss SME without its own AI team, this means the question is no longer whether you can execute, but whether you know which processes to automate first. Execution can be organized today through partners, agents and copilots, without building an engineering team of your own. What cannot be outsourced is the selection: which process contributes most to the business, which data is accessible, and which risk is acceptable.
This becomes particularly visible in integrations that are typical for Swiss SMEs: AI agents working in Abacus, Bexio or Microsoft 365 require clear decisions about data flows, access rights and process ownership. Add the revised Swiss Federal Act on Data Protection (revDSG) and Swiss hosting requirements, and you quickly see that these are leadership decisions, not technical details. A partner can handle the how. The what must stay with you.
Own the what, delegate the how
The operational consequence is clear: consistently hand over the how, and keep the what and why as strategic core competencies. In practice, this means building a decision structure that defines who approves an automation, which criteria drive prioritization, and how success is measured. As the Swiss podcast on AI agents as decision-makers argues, this is a management topic, not an IT topic.
- Build an in-house AI team
- Highest effort, longer time to value, full control over technology and data.
- Distribute chat licenses
- Quick but superficial: no integration with Abacus, Bexio or Microsoft 365, no real process automation.
- Outsource AI as an external division
- A partner handles implementation, decision and process ownership remain with the CEO, faster impact.
Four steps to dissolve the bottleneck
- Inventory: capture the processes that are repetitive, rule-based and data-rich.
- Evaluate: assess data quality, risk and measurable value for each candidate.
- Decide: determine which process runs first and who owns it.
- Review: check outcomes every quarter and adjust priorities.
What leaders should not delegate is the final word on risk and values. Even as agents take over more operational work, the limits of AI remain a leadership boundary. Your judgment decides which recommendations become decisions and which do not.
For mid-sized companies the message is encouraging: you do not need your own AI team to use AI agents effectively. You need a clear sense of which process to improve and what success looks like. At KI-Outsourcing.ch, we operate as an external AI division, taking over implementation while decision and process ownership stay with you. This is where real CEO relief comes from: delegate execution, gain decision space.
When implementation becomes a commodity, your advantage is no longer defined by the ability to build, but by the quality of your priorities.
The classic mistake
Delegating the decision itself to technology means losing direction. AI agents can execute processes, but they cannot define which process is valuable. That is why AI belongs in the C-suite, not on an IT project list. This is exactly what the Swiss podcast episode on making AI a C-level topic in mid-sized companies emphasizes.
The first step: a decision, not a project
Start with a single process. Define a success metric, identify the data source, and name an accountable owner. Hand implementation to a partner who can integrate it into your systems. What matters most is that you decide why this process comes first. If you do not set the order, someone else will. The bottleneck is not your ability to build; it is your willingness to choose.
Frequently asked questions
- Why is deciding the new bottleneck instead of building?
- Because models, copilots and AI agents are increasingly handling implementation. The job market data shows it: open engineering roles grow by only 2.5%, while product manager roles are at their highest level in nearly four years. What becomes scarce is not the ability to build, but the ability to decide what gets built.
- Does an SME need its own AI team to use AI agents?
- No. Implementation can be outsourced to partners who handle integration and operations. What you need is a clear decision about priorities, data and process ownership. For AI agents to work in systems like Abacus, Bexio or Microsoft 365, leadership decisions matter more than an internal engineering team.
- How do I identify which processes to automate first?
- Look for processes that are repetitive, rule-based, data-rich and measurable. For every candidate, assess data quality, risk and expected value. Start with one process, define a success metric, and hand implementation to a partner. You set the order, not the technology vendor.
- What data protection rules apply when using AI agents?
- The revised Swiss Federal Act on Data Protection (revDSG) requires clear responsibilities, data processing agreements and, where appropriate, Swiss hosting. Before deploying AI agents, you need to decide which data may be processed by which system. These are principle decisions for management, even when a partner handles implementation.
- What is the difference between chat licenses and real AI integration?
- Chat licenses give your team a generic tool that is not embedded in core processes or core applications. Real AI agents work inside your systems, such as Abacus, Bexio or Microsoft 365, and take over operational steps. The advantage comes not from the tool, but from deciding where and why it is used.
- How do I delegate implementation without losing control?
- By keeping decision authority. Define the what, the why and the success criteria, and hand the how to a partner. Keep a regular review rhythm and reserve the final word for risky decisions. Human judgment remains the boundary, even as AI agents take on more operational work.
Sources
Chris Jon Graf · AI Strategist & CEO
For over 15 years, Chris Jon Graf has built productive Swiss software platforms and advised executive teams on the strategic deployment of artificial intelligence. Would you like to evaluate this topic for your company?