Artificial Intelligence (AI) is changing not only individual tasks but also how people enter technical and commercial careers. The shift is particularly visible at AI company Anthropic, which is hiring fewer junior employees while assigning more work to AI agents. For people starting their careers, this creates a difficult gap: the routine tasks that once provided practical experience are precisely the ones being automated.
Why entry-level opportunities are narrowing
According to an internal study conducted by Anthropic, junior roles in its technical departments fell from 21 percent to 14 percent since the beginning of 2025. Among new hires, the share dropped even more sharply, from 27 percent to 9 percent. Open junior positions declined particularly strongly after December 2025.
Anthropic links this development to AI agents. These are systems that can handle multistep assignments with some autonomy rather than merely responding to a single text request. In August 2026, the company says its agents logged six times as many working hours as its human employees. That does not mean an agent was six times more productive in every hour; the figure initially describes recorded work volume.
Tasks such as writing code and finding errors were increasingly assigned to Claude. Junior employment fell most sharply in jobs where the system was used for particularly complex work. Anthropic itself acknowledges that it is an unusual case: a fast-growing software company with exceptionally intensive AI use.
The broader labor-market risk is not necessarily the disappearance of entire professions. It is the loss of the traditional learning ladder. A junior developer who rarely searches for smaller errors independently, or a young data specialist who mainly reviews generated queries, gains different experience from previous generations. Companies may become more efficient in the short term, but they still need a way to turn beginners into experienced professionals.
How AI agents are changing professional roles
The shift extends beyond software development. Anthropic reports that its legal department processed 3.1 times as many contracts, sales staff closed 1.7 times as many deals, and accounting processed 1.4 times as many invoices. These are provider-reported figures and do not establish that other organizations would see the same effects.
At the same time, Anthropic wants to entrust agents with more demanding assignments. The company is consulting law firms about negotiation standards that could allow agents to negotiate prices, resolve complaints, or improve contract offers. According to Anthropic, its experiments also reveal a concern: stronger models negotiate more successfully and could therefore reinforce existing inequalities.
This creates new kinds of work instead of simply making expertise unnecessary. Anthropic is recruiting legal professionals and other specialists to test Claude on judgment-based questions in law, philosophy, economics, and related fields. The company lists compensation of $100,000 to $167,000 for a four-month fellowship. Such unusually well-paid positions do not, however, replace the broad base of entry-level jobs that may be declining.
The underlying problem is evaluation. In mathematics or software code, it is often relatively straightforward to check whether an answer is correct. Legal reasoning, political judgment, and contract negotiations depend more heavily on context and human interpretation. Some professionals may therefore move from directly completing work to defining quality standards, supervising systems, and deciding difficult edge cases.
What the productivity figures really show
Anthropic reports substantial internal gains: engineers produced 5.1 times as many recorded code changes, data specialists submitted 3.1 times as many queries, and recruiters hired 1.4 times as many employees. Yet output counts do not automatically reveal whether results improved or how much time was required for review and correction.
There is also a conflict between Anthropic’s company data and the broader picture. While Anthropic reports large increases, t3n points to a software-development study that raises doubts about AI productivity gains. Anthropic also notes that the wider economy has not yet demonstrated a comparable improvement.
One reason is that jobs are not merely collections of independent activities. Task intelligence, meaning the systematic breakdown of a job into individual activities, can reveal suitable opportunities for automation. However, Josh Bersin cautions against an overly simple task model: individual activities are connected to business processes, professional roles, and situational knowledge.
Two practical examples illustrate the distinction. An agent can reliably complete a two-minute update in a customer-management system after a sales conversation. In a safety-related setting, an automated system can remind an airline pilot to lower the landing gear at the correct speed and time. Whether AI, conventional automation, or a person is best suited to either process still requires a separate assessment.
The subjective quality of work can change as well. Software expert Simon Willison describes coding agents as helpful for people who primarily care about building functioning systems. For programmers who value directly solving difficult logic problems as a central source of professional satisfaction, that same shift can feel like a genuine loss. Productivity and job satisfaction do not necessarily move in the same direction.
Pros and Cons of AI agents in the workplace
Pros:
- Less routine work – Agents can handle recurring data entry, document reviews, and information comparisons.
- Greater capacity – Teams may process more contracts, queries, or invoices when results can be reviewed reliably.
- New specialist roles – People are needed for quality evaluation, supervision, process design, and difficult edge cases.
- Better decision support – Agents can consolidate scattered information and prepare the next steps in a workflow.
Cons:
- Fewer entry-level roles – Automating simple assignments can remove learning opportunities and first professional experiences.
- Uncertain productivity – Producing more outputs does not automatically result in better or more economical work.
- Additional oversight – Errors, unsuitable decisions, and unclear accountability can create new review work.
- Unequal bargaining power – Stronger agents could give well-equipped companies and individuals an additional advantage.
A report commissioned by Alibaba.com and an industry association identifies order processing, customer service, regulatory compliance, product research, offer preparation, and international sourcing as potential uses for smaller businesses. In the YouGov survey cited by the report, roughly one-third of small and midsize businesses described administrative overload as a concrete problem. About 40 percent identified order processing and customer service as especially time-consuming, while more than half of companies with at least ten employees saw the potential to save at least 10 percent of operational routine work.
These figures appeared in a sponsored article about AI agents in midsize businesses and should be interpreted accordingly. They show perceived potential, not an independently verified productivity gain. The provided sources do not specify tool prices, implementation costs, or the expense of ongoing human review.
What this means for you and Switzerland
If you are at the beginning of your career, the most useful first step is not to hand every assignment to an agent as quickly as possible. Choose a clearly limited workflow, such as preparing product research or structuring a document, and compare the generated result with your own work. This helps you learn both the professional task and the tool’s typical limitations.
For advanced users, the larger benefit lies in designing complete workflows. You can classify activities by whether they follow clear rules, how easily their results can be verified, and how serious an error would be. Routine steps can be automated while approvals, quality criteria, and difficult decisions remain deliberately under human control.
In Switzerland, two developments are colliding. ETH Zurich and EPFL produce an unusually strong pool of highly trained AI specialists, according to Netzwoche’s report on the Swiss talent market, but American technology companies recruit some of those graduates. Google, Microsoft, Meta, Apple, OpenAI, and Anthropic have offices in Zurich or participate in the local research environment.
The report says Google donated 500,000 Swiss francs to the ETH AI Center in 2026 and provided three doctoral researchers with funding and mentors. Precise figures on how many graduates move to large technology companies are not available. Switzerland therefore faces a tension: it has a strong education and research base, but it needs attractive career paths if that knowledge is also to benefit local companies and the country’s digital autonomy.
AI is shifting the value of work away from execution alone and toward oversight, contextual understanding, and process design. It can reduce routine work and create highly qualified specialist roles while making the entry-level assignments that once led to those roles scarcer. The unresolved risk is whether employers will create durable new learning paths or exchange short-term efficiency for a longer-term shortage of experienced professionals.
Sources
- Anthropic Is Hiring Far Fewer Junior Employees as AI Agent Work Increases – The Decoder, 2026-10-18
- Anthropic Consults Law Firms About AI Agents Negotiating Prices and Contracts – The Decoder, 2026-10-18
- My obituary for programming – Simon Willison’s Weblog, 2026-10-19
- Five Things to Know This Week: Does AI Really Increase Productivity? – t3n, 2026-10-19
- What Is Task or Work Intelligence And Can We Decompose Every Job For AI? – Josh Bersin, 2026-10-19
- Big Tech Competes for Switzerland’s Best AI Talent – Netzwoche, 2026-10-19
- Why Midsize Businesses Should Not Wait to Adopt AI Agents – t3n, 2026-10-18
Image: Mert Coşkun via Pexels