European businesses are adding chatbots, document assistants, recommendation systems, content generation, and automated decision support to ordinary software. The EU AI Act makes one question unavoidable: what role does the company play in the system, what risk does the use create, and what information must be given to the people affected?
This article is a technical and product-planning guide, not legal advice. Classification and obligations depend on the specific system, sector, deployment, and contractual relationships. The useful engineering lesson is that governance cannot be added the night before launch.
Begin with the use case, not the model name
The same language model can support a low-risk writing assistant or participate in a sensitive employment, education, credit, health, or public-service decision. Document the intended purpose, users, inputs, outputs, affected people, and consequences of error. That description is more useful for risk analysis than a list of model features.
Make the AI interaction visible
The European Commission identifies transparency requirements for certain AI systems, including informing people when they interact with a chatbot and labelling specified synthetic content. Product teams should design this into the interface: a clear identity for the assistant, explanations of what it can and cannot do, and an obvious route to a person when the interaction matters.
Know your role in the supply chain
A company may use a third-party model while still designing and deploying the customer-facing AI system. Responsibilities can be distributed among model providers, system providers, deployers, integration partners, and customers. Contracts and technical documentation should identify who chooses the model, controls data, evaluates the use case, handles incidents, and communicates material changes.
Keep evidence of how the system was evaluated
Evaluation should cover the actual business task, languages, user groups, and foreseeable misuse. Record test datasets, success criteria, known limitations, human-review rules, and changes between releases. A demonstration with ten friendly questions is not evidence that a system will behave reliably with real customers or employees.
Build human oversight as a real control
A human approval step is useful only when the reviewer receives enough context, has time and authority to disagree, and can understand the proposed action. Interfaces should show evidence, uncertainty, and consequences. Repeatedly asking a person to click approve on opaque output creates ceremony, not oversight.
Control data and retention
Map what information enters prompts, retrieval indexes, logs, evaluation datasets, and third-party services. Apply data minimization, access control, retention periods, and deletion processes. Do not assume a provider setting alone resolves privacy, confidentiality, intellectual-property, or cross-border data questions.
Prepare for change
Models, prompts, tools, and source documents change frequently. Version the important components, monitor performance, and define when a change requires renewed evaluation or user communication. Incident reporting and rollback should be designed before the AI feature becomes part of a critical workflow.
A practical launch checklist
Before release, record the intended purpose and responsible owner; assess the use and affected people; disclose the AI interaction where required; verify provider and data terms; test accuracy, misuse, accessibility, and security; establish human escalation; log important actions; and schedule ongoing review. Legal counsel should validate the final classification and obligations for the specific deployment.