Analysis note
Offers of “AI agents” aimed at small businesses are multiplying. Most of them rest on a language model framed by instructions and connected to a few tools. The whole follows a predefined sequence. The distinction between automated workflows and agents that decide on their own actions is not a debate about vocabulary. Indeed, it determines the risks, the level of control and the justified price.

Observation: what offers of AI agents cover
The term “agent” covers different realities. It may designate a conversational assistant equipped with instructions. It may also designate an automated sequence of tasks (reading an email, extracting information from it, drafting a reply). Finally, it may designate a system capable of choosing its own steps and tools to achieve an objective. Commercial offers, however, often use the same word for all three.
The EU AI Act defines an AI system as a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations or decisions that can influence physical or virtual environments (Article 3, point (1)). Autonomy is thus a matter of degree there, not a separate category.
Analysis: predefined workflows and agents
In a technical publication of December 2024, a model developer distinguishes workflows from agents. Workflows orchestrate models and tools through predefined code paths. In agents, on the other hand, the model directs its own process and tool usage. It recommends adopting the simplest solution that meets the need and adding autonomy only when simpler solutions fall short.
For an MSME, most of the needs expressed fall within predefined workflows. These include sorting mail, preparing quotes, answering frequently asked questions and posting on social media. This is not a weakness. In fact, a predefined workflow is more predictable, easier to verify and less costly than an autonomous agent.
Analysis: the risks specific to autonomous agents
The OWASP security framework for applications based on language models uses the term “excessive agency”. It designates the vulnerability that allows damaging actions in response to unexpected, ambiguous or manipulated outputs. It identifies three causes: excessive functionality, excessive permissions and excessive autonomy, in the absence of validation for high-impact actions.
An agent that reads emails or documents may receive, within that content, hidden instructions intended to divert it. Where it may send messages, modify files or place orders, the damage is no longer an erroneous answer. It then becomes an action carried out in the company’s name.
Implications: reading an offer of AI agents
Five questions make it possible to qualify an offer before any commitment:
- What actions can the system carry out without human validation, and on which systems?
- Is the sequence predefined, or does the model choose its own steps?
- What data are transmitted, to which model provider, and where are they processed?
- What records are kept of the actions and of the reasons for them?
- How is the system stopped, corrected or updated when the instructions or the model change?
An offer that does not make it possible to answer these questions therefore sells a promise rather than a service. The framing work (instructions, scope, controls, records) justifies the price, not the word “agent”.
Points of attention
- Require the list of actions that the system can carry out without human validation.
- Limit the system’s permissions to the necessary functions only.
- Subject to human validation any action with external effect: sending, payment, publication, deletion.
- Check the place of processing of the data transmitted to the model.
- Keep a record of the actions carried out and the instructions applied.
References
- Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence and amending Regulations (EC) No 300/2008, (EU) No 167/2013, (EU) No 168/2013, (EU) 2018/858, (EU) 2018/1139 and (EU) 2019/2144 and Directives 2014/90/EU, (EU) 2016/797 and (EU) 2020/1828 (Artificial Intelligence Act), OJ L, 2024/1689, 12.7.2024, Article 3, point (1).
- Anthropic, “Building effective agents”, 19 December 2024, https://www.anthropic.com/engineering/building-effective-agents (page consulted on 21 October 2026).
- OWASP, “LLM06:2025 Excessive Agency”, OWASP Top 10 for LLM Applications 2025, https://genai.owasp.org/llmrisk/llm062025-excessive-agency/ (page consulted on 21 October 2026).
See also: Guide – Configuring an AI assistant; Checklist – Security of AI uses; Grid – Assessing an AI provider. Information Systems & Digital Transformation practice.
General information document; it constitutes neither legal advice nor an opinion on a particular situation. References to legal texts should be checked against the version in force on the date of use.
Consulting firm not registered with any Bar; no representation before the courts and no act reserved to the regulated legal professions.
