The Latin phrase "Da mihi factum, dabo tibi ius" ("Give me the facts, and I will give you the right") implies that facts are readily available, complete, and current. However, this assumption is rarely tested when artificial intelligence systems provide answers that cite their sources correctly. In a previous discussion titled The signature at the bottom of the page, the focus was on hallucinations, source verification, and the responsibility of professionals using AI-generated content. The need for reliability remains, as even an accurate rule might not apply to a specific company. This discussion focuses not on the compliance of AI systems, which is governed by Regulation (EU) 2024/1689, but on how these systems are used to apply the law to individual companies. A legal department in France oversees a dozen companies. The human resources department asks whether a subsidiary, which now has 54 employees, is subject to the obligation to participate in results. An AI tool can identify the relevant articles of the Labor Code, correctly cite the threshold of 50 employees, and explain the applicable regime. However, the conclusion might not be accurate. The obligation begins after the five-year civil period mentioned in the Social Security Code, and the threshold must be met or exceeded over five consecutive years. A decrease in the number of employees resets the count. The relevant workforce is not the current number but the average over the previous civil year. Therefore, the number of 54 employees alone is not sufficient to answer the question. Historical data is needed, and this data is often spread across multiple systems, such as human resources and payroll. Retrieval-Augmented Generation (RAG) techniques, which use documents retrieved from a corpus to back up a model's answers, reduce the risk of citing non-existent sources. However, they do not inform the model about the actual situation of the company. For example, the documents can confirm a threshold but not the legally relevant workforce. This highlights a key aspect of legal reasoning: facts are interpreted in light of the rule, and the rule is selected based on the facts. In works published in 2026, Eljas and Tuula Linna examine the issue of legal reasoning by AI using the framework of Issue, Rule, Application, Conclusion. They emphasize the importance of selecting the right legal framework, applying the rule, and evaluating evidence. The challenge remains in identifying what facts the system has, how reliable they are, and what to do when the necessary facts are missing. Legal professionals often deal with ongoing operations, disputes, or questions about a specific date. Legal departments manage organizations with constantly changing characteristics: workforce numbers, opening or closing locations, acquisitions, and new activities abroad. Laws also evolve: new texts come into force, thresholds change, and legal interpretations evolve. A conclusion that was accurate at the time may need to be revised later without the initial reasoning being wrong. At the scale of a group with many companies, or for professionals managing hundreds of companies, it is essential to identify which entities a change might affect and what facts are needed to determine that. This requires knowing where the information is, who ensures its accuracy, and what changes would justify reevaluating a conclusion. As "agentive" systems, which use the organization's data and tools, become more common, these challenges will grow more complex. The absence of information is not the same as having negative information. The lack of data showing a company exceeds a threshold does not prove it is below it. An activity not recorded in a system may still exist, and data from different systems may contradict each other. A data point that was once accurate may no longer be. Therefore, a system designed to assist in applying the law should distinguish at least three scenarios: the rule applies, it does not, or the information is insufficient to conclude. The third scenario is often undervalued. An answer stating that a conclusion is not possible without the subsidiary's workforce history may be less impressive than a quick response, but it is often more useful to the lawyer who must address the issue. Some works emphasize auditible decision chains, uncertainty evaluation mechanisms, and hybrid architectures, while acknowledging their limitations. Knowing when not to conclude is a method: identifying what is missing and leaving the professional to act accordingly.