AI agents are already performing work that people used to handle at some companies: writing software code, checking documents, processing support cases, and preparing contracts. These agents are Artificial Intelligence (AI) systems that pursue a goal and carry out several related steps with limited supervision. The shift particularly affects people entering the workforce, but it also matters to small and midsize businesses (SMBs) whose lean teams have to manage large amounts of routine work.
Tasks rather than entire professions
The available reports point less to the full automation of professions than to a redistribution of individual tasks. Traditional job-task analysis breaks a role into specific activities and asks whether each one belongs with an AI agent, conventional software, or a person. Josh Bersin’s assessment of task analysis offers a straightforward example: If updating a customer record in Salesforce is repetitive and unambiguous, an agent can take care of it.
The approach becomes less reliable when a task only makes sense within a broader role. A human resources specialist, for example, does not simply complete a collection of isolated steps but must account for company context, relationships, and unusual situations. The central issue is therefore not just whether a task can be automated, but whether it is structured, verifiable, and sufficiently independent of human judgment.
Replit provides a practical operating model. When a support request arrives, an agent first examines error reports, software information, and internal tickets. A person receives the case with that added context and, if they decide a fix is needed, can launch another agent to prepare the change. Replit’s description of a partly self-driving company illustrates a workflow in which agents perform much of the preparation while a person retains the decision in the middle.
Why junior roles are under pressure
Anthropic offers the clearest available view of how entry-level roles may be affected. According to an internal study, the share of junior employees in its technical departments declined from 21% to 14% after the beginning of 2025. Their share of new technical hires fell even more sharply, from 27% to 9%, while open junior roles decreased particularly strongly after December 2025.
The company attributes this development to its intensive use of the Claude AI system. In August 2026, AI agents accounted for six times as many work hours as human employees, according to Anthropic. That does not mean one agent was six times as productive as one person: The figure compares total recorded hours and, by itself, says nothing about quality or economic value.
Anthropic nevertheless reports substantial increases in output. Engineers produced 5.1 times as many software changes, data scientists submitted 3.1 times as many queries, and the legal department processed 3.1 times as many contracts. Sales, recruiting, and accounting also handled larger volumes, as the report on Anthropic’s internal research summarizes.
Those figures come with clear limitations. Anthropic is a rapidly growing software company that has organized its workflows around unusually intensive AI use, and the study itself warns against applying its findings directly to the broader economy. More software code also does not automatically mean better code or greater business impact, although Anthropic says it has not found evidence that code quality is declining.
The remaining human work does not necessarily become easier. Employees reported pressure to use Claude across as much work as possible, while 20% said in an internal survey that their job had shifted from performing work to managing multiple Claude instances. TechCrunch describes human attention and judgment as a new bottleneck: People must read, organize, and assess a flood of AI-generated material while also spending more time in meetings.
Practical value for SMBs
For SMBs, the most immediate benefit lies in recurring administrative and information work. A YouGov survey commissioned by Alibaba.com and the German E-Commerce and Mail Order Association found that about one-third of surveyed SMBs saw administrative overload as a concrete problem. Around 40% identified order processing and customer service as particularly time-consuming, while more than half of businesses with at least ten employees saw potential to save at least 10% of operational routine work through AI agents.
These figures appeared in a sponsored article about the commercial Accio platform, so they should not be treated as an independent effectiveness study. The use cases it lists are still concrete: Agents can combine information from multiple sources, compare data, pre-structure documents, and prepare product research or quotes. The article about AI agents for midsize companies also names order processing, customer service, regulatory documentation, and international sourcing.
Online retail offers an even more granular division of labor. German startup Shop Agentic says it wants to deploy as many as 60 specialized agents for one online store. They would prepare product information, monitor inventory, and process orders, according to the profile of Shop Agentic. The source does not independently verify whether such a large group of agents produces better everyday results than a smaller number of more flexible systems.
The next stage extends beyond internal routine work. Anthropic is reportedly consulting law firms about rules for agents that could negotiate prices, resolve complaints, or improve contract offers. The company’s experiments suggested that more capable models negotiate more effectively, potentially putting weaker market participants at a disadvantage. The work on negotiation agents therefore enters an area where fairness, accountability, and difficult-to-measure quality matter more than speed alone.
The seven reports do not provide enough information to compare the cost of these tools. They lack consistent figures for licenses, setup, supervision, and the cost of errors. For an SMB, the number of available agents is therefore less meaningful than the total effort required to complete a process reliably.
Pros and Cons of AI agents as coworkers
Pros:
- Less routine work – Agents can handle repetitive data entry, document checks, and information searches.
- Faster preparation – Support cases, quotes, and orders can be structured before a person makes a decision.
- More output per team – Anthropic’s internal figures show much higher processing volumes across several departments, albeit in an unusual company.
- Capacity for smaller teams – SMBs can manage specialized workflows without completing every intermediate step manually.
Cons:
- Fewer entry-level roles – Standardized tasks that traditionally help junior employees build experience are among the first to be automated.
- New supervision demands – Large volumes of automatically generated material can turn human attention into the limiting factor.
- Uncertain quality – More code, contracts, or responses do not automatically produce better outcomes or greater business value.
- Unequal bargaining power – More capable agents could gain an advantage in price and contract negotiations before fair rules are established.
What this means for you and Swiss companies
If you are getting started, choose a limited and frequently repeated process with an outcome you can verify. Sorting incoming support requests or compiling product information is a more suitable first test than allowing an agent to handle an important contract autonomously. Record how long the process currently takes, which errors occur, and where a person must explicitly make the decision.
If you are an advanced user, examine an end-to-end process and assign its tasks selectively. Some steps may suit an AI agent, some belong with conventional automation, and others should remain human because they depend on context or responsibility. Greater value comes not from maximizing the number of agents, but from clear handoffs, verifiable outputs, and deliberate limits on what each agent may do.
The sources provide no Switzerland-specific information about availability, supported national languages, data locations, or privacy terms. International examples therefore do not establish that a tool will fit a Swiss business without modification. Companies would need to assess the relevant product terms and their own requirements separately when handling customer information, employee data, or contract documents; the reports did not study outcomes specific to Switzerland.
For now, AI agents appear more likely to replace bundles of tasks than neatly defined professions, but that is still enough to reshape job descriptions and career paths. Their value is most plausible in structured, repetitive processes, while judgment, supervision, and context become more central to human work. The unresolved risk is how companies will develop experienced specialists if they automate the junior tasks through which that experience was traditionally acquired.
Sources
- Anthropic stellt kaum noch Nachwuchskräfte ein, weil KI-Agenten sechsmal so viel wie Menschen arbeiten – Unknown, 2026-10-18
- What Is Task or Work Intelligence And Can We Decompose Every Job For AI? – Unknown, 2026-10-19
- Warum der Mittelstand bei KI-Agenten jetzt nicht warten sollte – Unknown, 2026-10-18
- Anthropic berät sich mit Kanzleien, damit KI-Agenten künftig Preise und Verträge aushandeln – Unknown, 2026-10-18
- Anthropic says its workforce has become less junior and is spending more time in meetings – Unknown, 2026-10-17
- Replit’s CEO says the future of work is the ‘self-driving company’ – Unknown, 2026-10-16
- 60 KI-Agenten für einen Webshop: Shop Agentic will den E-Commerce umkrempeln – Unknown, 2026-10-16
Image: Christina Morillo via Pexels