AI agents: what can they really do?

From chatbots to agents

Where classical chatbots answer individual questions, agents work more autonomously: they can make decisions, execute actions, and combine multiple steps to reach a goal.

The fundamental difference can be summarized like this:

Chatbot	Agent
Reacts to input	Actively pursues goals
Single responses	Multi-step workflows
Needs precise instructions	Interprets intent
No actions	Executes real actions

An example: ask a chatbot “Book me a flight to London” and it responds with information. An agent, by contrast, searches for flights, compares prices, picks the best option, and books it - with your confirmation.

How agents work technically

Behind every AI agent sits an architecture made up of several components:

The agent stack

1. 1

   User

   Starts the task in natural language, sets goal and constraints.
2. ↓
3. 2

   Orchestration layer

   Decomposes the goal, plans steps, fires tool calls, checks intermediate results.
4. ↓
5. 3

   LLM (reasoning engine)

   GPT-5.x, Claude 4.x, Gemini 3 - make the actual reasoning decisions.
6. ↓
7. 4

   Tool layer

   Browser, email, calendar, databases, APIs - this is where the agent touches the real world.
8. ↓
9. 5

   Memory / context

   Long-term memory, prior interactions, vector stores - keeps the agent consistent across turns.

Data flow: request on top, tool execution at the bottom. Memory feeds back into orchestration.

The four layers in detail

1. Developer layer: tools such as GitHub Copilot and Claude Code help with writing and testing code
2. Knowledge-worker layer: agents support research, writing, and report generation
3. Workflow layer: platforms automate multi-step business processes
4. Control layer: systems provide security, monitoring, and access management

Adoption and market potential

Industry reports (e.g. McKinsey State of AI, IBM Institute for Business Value) show a clearly elevated adoption level for agentic AI workflows in 2025 - the exact numbers vary by methodology and industry. The drivers are the same across all studies: skills shortages, rising wage costs, and the increasing capability of the underlying LLMs.

Concrete market indicators for SMEs:

• The majority of large LLM providers (OpenAI, Anthropic, Microsoft, Google) launched agent frameworks in 2024-2025.
• No-code platforms (OpenAI Agent Builder, Microsoft Copilot Studio) significantly lower the entry barrier.
• Specialized agent SaaS fills niches (sales, support, research).

What are agents used for?

Software development: the front-runner

In software development, the productivity gains are best studied: a controlled study with over 4,800 developers shows around 26% more completed tasks (individual vendor figures run higher but depend heavily on the setup). Concrete applications:

• Code generation: descriptions become functioning code
• Bug fixing: automated identification and remediation of defects
• Code reviews: quality checks and improvement suggestions
• Test generation: automated unit and integration tests
• Documentation: code is documented automatically

Tools such as Claude Code or GitHub Copilot are standard in many developer teams today.

Business applications

Agents also provide support outside of IT:

Research and analysis:

• Create competitive analyses
• Identify market trends
• Summarize and compare documents

Communication:

• Draft and answer emails
• Generate meeting summaries
• Produce reports automatically

Organization:

• Coordinate appointments
• Plan and book travel
• Prioritize tasks

Customer service:

• Categorize and answer requests
• Detect escalations
• Automate follow-ups

The most important agent platforms in 2025

Platform	Strength	Ideal for
Manus (Butterfly Effect)	Fully autonomous tasks	Enterprise automation
Claude Computer Use	Desktop control	Complex UI tasks
OpenAI Operator	Browser automation	Web-based workflows
Devin	Software development	Coding teams
AutoGPT	Open source, flexible	Experimentation

Opportunities and limits

What agents do well

• Repetitive tasks: anything that follows clear rules
• Data processing: gathering, structuring, analyzing
• Multi-tool workflows: coordinating different systems
• 24/7 availability: no breaks, no vacation
• Consistency: the same quality on every run

Where agents (still) fall short

• Creative decisions: real innovation needs humans
• Emotional intelligence: complex interpersonal situations
• Unknown situations: when no rules exist
• Ethical judgments: value decisions stay with humans
• Long-term planning: strategic directional calls

Critical success factors

AI agents are only as good as:

• The data they access
• The rules they act under
• The goals they are given
• The controls humans apply

Governance for Swiss SMEs

Before an agent goes into production, four Swiss specifics must be clarified:

• Art. 21 DSG (Swiss Data Protection Act) - for automated individual decisions involving personal data (applications, pricing, complaints) a human review option is mandatory. A “silent” pre-sorting without an escalation path is not permissible.
• Art. 22 DSG - for sensitive data (health, religion, biometric data) a data protection impact assessment is required before productive use.
• DPA with the platform provider - a data-processing agreement under Art. 9 DSG, with clear training-use and deletion clauses.
• FINMA implications - banks, insurers, and asset managers with a FINMA license must review recording and outsourcing obligations against agent use. A pilot without compliance alignment is not enough here.

Covered in depth in the DSG guide for Swiss SMEs.

How to get started with agents

Phase 1: Understand (1-2 weeks)

1. Test free tools (ChatGPT, Claude Free)
2. Identify repetitive tasks in your daily work
3. Document: what does this task cost today?

Phase 2: Pilot (4-8 weeks)

1. Choose ONE process with a clear ROI
2. Define success criteria (time, quality, cost)
3. Start with a small team
4. Measure continuously

Phase 3: Scale (ongoing)

1. Document best practices
2. Identify additional use cases
3. Train the team
4. Build governance

Conclusion

AI agents open up enormous productivity gains - when deployed correctly. The key lies in:

• Clear goals: what should the agent achieve?
• Fitting processes: not everything lends itself to automation
• Human oversight: agents as assistants, not replacements

In our trainings we highlight both the possibilities and the limits of this technology - and show you how to start concretely.

---

Sources:

• Index.dev: State of AI Agents Report (2025)
• McKinsey: AI Agents in Business (2025)
• GitHub: Copilot Impact Report (2025)
• Gartner: AI Agent Market Analysis (2025)