An AI answer can be clear and well phrased, yet still fail to answer the question exactly. At work, the right approach is therefore not just to choose a tool: it is to decide what task to give it, what information to share and how to check what it produces.
Luca Barnabo, a Luxembourgish founder of consultancy Lux Digital Solutions, summed up this gap in an interview published by L’Essentiel on 12 November: a natural-sounding conversation can inspire confidence without guaranteeing accuracy. His advice is to treat AI as an aid for a defined task, rather than as an authority that validates its own results.
Start with the task, not the tool
Proofreading a text, organising ideas or preparing a summary from public information are tasks whose results can be reviewed and corrected. In a hypothetical example, a team could ask an assistant to group the published features of several products in a working table. The tool then does some of the comparison and formatting; a person still needs to check the data and decide whether the comparison is relevant.
This distinction helps avoid confusing production with validation. An answer can be useful as a first draft, even if it is not a reliable conclusion. The reader can edit the text, rephrase the request or check the points that matter. The more clearly the task is defined, the easier it is to see what the tool was meant to do and where human oversight begins.
Before assigning a task, it is therefore useful to ask: what result are we expecting, what information should it be based on, and how will we know whether it is correct? If these questions do not have concrete answers, a vague instruction is more likely to produce a result that is difficult to assess.
The information shared is part of the decision
The second decision concerns data. A text to be summarised does not necessarily need to include names, contact details or internal information. If a public version is enough for the task, it reduces the information shared with the service. This principle is reflected in guidance from the National Commission for Data Protection (CNPD): define the purpose of processing, limit personal data to what is necessary and proportionate, explain how it will be used and carefully control access.
In a hypothetical example, an employee who wants to prepare questions about a contract could use a public document or a general description of the problem. If the analysis requires confidential information, they must first determine whether the tool is authorised for this use and whether the safeguards in place are appropriate. Removing a name is not always enough to make a document anonymous: the combination of a job title, date and particular situation may still make it possible to identify someone.
The price of a service does not, in itself, settle the question of confidentiality. Barnabo makes this point in his interview with L’Essentiel: a paid subscription does not automatically guarantee that conversations will remain confidential. Before sharing data about clients, employees or other people, it is therefore necessary to check the terms of use and the framework adopted by the organisation. Passwords, meanwhile, should not be copied into a conversation with an assistant.
Check an answer by tracing it back to its source
To check an answer, start by opening the references it provides. A source may genuinely exist without supporting the sentence that the AI links to it. Check who is speaking, when, in what context and exactly what the document says. A summary may omit a condition or turn a possibility into a certainty.
BEE SECURE, Luxembourg’s information and awareness service for online safety, recommends asking where information comes from, whether its source is reliable and how other media are covering the subject. These checks help assess a claim, but do not by themselves guarantee that it is true. Comparing several reports of the same press release does not amount to having several independent confirmations either.
The level of checking should match the consequences of an error. For a draft or a list of ideas, careful proofreading may be enough. For an important decision involving law, health or money, a chatbot answer is no substitute for authoritative documents or, where necessary, professional advice. Barnabo recommends basing the level of checking on the potential impact of an error.
When AI does more than answer
Some tools can connect to other software or carry out actions. It is then important to distinguish between a proposal written by AI and an operation actually carried out: sending a message, changing a file or sharing data does not have the same consequences as producing a draft.
Barnabo believes that an agent’s access should be limited to the task envisaged and that important actions should require human approval. A written instruction is not enough to guarantee these limits: the permissions and available tools must also be appropriate. For work with significant consequences, it is better for the person concerned to be able to review the action before it is carried out.
In Luxembourg, as elsewhere, the professional challenge is therefore not to delegate judgement wholesale. It is to assign a task precisely enough that its result can be reviewed, choose only the information that is genuinely necessary and maintain oversight proportionate to what is at stake. Knowledge of the job remains crucial: it makes it possible to spot an answer that seems plausible but does not fit the case.
References
- L’Essentiel, “When and how to use artificial intelligence?”
- CNPD, “AI obligations: which GDPR provisions must be respected?”
- BEE SECURE, “Getting information online by questioning it”
Working at a laptop. Illustrative photograph taken in July 2021.
Alabama Extension / Devontae Lindsey — Wikimedia Commons — CC0 1.0.
