Analysis
AI agents in business: what they are, what they can do and how to deploy them safely
An AI agent is a system in which a language model directs the steps of a task and its choice of tools itself, instead of following a predefined script. The length of the software tasks that leading models complete with a 50% success rate has doubled roughly every seven months since 2019. Agents work best where the task is well structured, the result can be checked and mistakes are reversible.132
Agent, workflow and chatbot: the difference in one sentence
In a workflow, the language model and the tools are linked by a predefined path; in an agent, the model directs its own steps and choice of tools1. The OECD describes agents as systems that, with a degree of autonomy, perceive their environment and act in it, using tools to achieve a goal2. In business practice the two approaches complement each other: where the steps are predictable, a workflow is cheaper and easier to control.
How fast agent capabilities are growing
The METR institute measures the length of the tasks that models complete with a 50% success rate3. Since 2019 this horizon has doubled roughly every seven months3, and according to a 2026 update the pace has accelerated since 2024 to a doubling about every 89 days4. METR cautions, however, that the measurement mainly concerns software tasks and does not directly tell you which tasks can be handed over to AI5.
Where agents prove their worth in companies
OECD interviews with organisations deploying agents show that, in the near term, agents work where a task is structured enough for the agent to plan it, the result can be checked against reliable data, and the cost of a mistake is limited or reversible2. This matches our experience: the agent prepares a report, sorts incoming requests or keeps a task register up to date, while the decision with real consequences stays with a person.
How many companies already use agents
| Study | Population | Uses agents | Plans to |
|---|---|---|---|
| Bitkom, Germany, 2026 (603 companies with 20 or more employees) | companies that use, plan or are discussing AI | 11% | 29%6 |
| BSI and TÜV, Germany, 2026 (505 companies) | companies using AI | 5% | 8%7 |
The results differ because the studies surveyed different groups of companies, which is why we report them separately rather than as a single figure. For Poland, we have not found a study with a methodology that measures agents; Statistics Poland (GUS) reports that in 2025, 8.7% of enterprises used AI at all8.
Risks that need to be addressed
OWASP lists excessive agency as one of the main risks of applications built on language models: too much functionality, too many permissions or too much autonomy allow harmful actions in response to an unexpected or manipulated model output9. The second risk is prompt injection through external content, for example web pages or files that the agent reads10.
For high-risk systems, the EU AI Act requires effective human oversight; following the 2026 amendments, these requirements apply to systems listed in Annex III from 2 December 202711.
How to deploy an agent with control
- Give the agent only the permissions it needs for the task.
- Leave actions with significant consequences, such as payments, messages to customers and changes to systems, for a human to approve.
- Treat external content (web pages, emails, files) as untrusted and keep it separate from instructions.
- Log every step the agent takes, so that you can reconstruct what it did and why.
- Measure quality on a test set built from real cases before the agent goes into production.
Frequently asked questions
How is an AI agent different from a chatbot?
A chatbot answers questions. An agent plans and carries out the steps of a task itself, using tools; for example, it retrieves data from a system, analyses it and saves the result.1
How many companies use AI agents?
In Germany, according to Bitkom, 11% of companies that use, plan or are discussing AI; according to a BSI and TÜV study, 5% of companies using AI. For Poland, official statistics so far only measure AI in general.678
Can an AI agent operate without supervision?
Not for actions with significant consequences. OWASP recommends that a human approves such actions before they are carried out. For high-risk systems, the EU AI Act requires effective human oversight from 2 December 2027.911
Will an agent handle any task that is shorter than its ‘time horizon’?
No. METR, which measures this horizon, states explicitly that a result of X hours does not mean that tasks shorter than X hours can be handed over to AI, and that the measurement mainly covers software tasks.5
Sources
- Anthropic: Building effective agents
- OECD: Agentic AI in organisations: Early insights from practitioner interviews (OECD AI Papers No. 65)
- METR / arXiv: Measuring AI Ability to Complete Long Software Tasks (Kwa et al., arXiv:2503.14499)
- METR: Time Horizon 1.1
- METR: Clarifying limitations of time horizon
- Bitkom: Erstmals nutzt die Mehrheit der Unternehmen KI
- BSI und TÜV-Verband (Feldarbeit: forsa): KI-Sicherheitsstudie 2026
- GUS: Społeczeństwo informacyjne w Polsce w 2025 r.
- OWASP GenAI Security Project: LLM06:2025 Excessive Agency
- OWASP GenAI Security Project: LLM01:2025 Prompt Injection
- EUR-Lex: Rozporządzenie (UE) 2026/1744 (Digital Omnibus on AI)