Useful AI starts with a well-defined task.

A public brochure, an internal procedure and an identity document require different handling. Before choosing a tool, the project needs to identify who may read each source, why the information is needed and who will receive the result. These responsibilities remain in place when a language model becomes part of the process. A useful task might be narrower than the original request for an assistant.

For example, finding the current version of an approved procedure may require reliable search rather than generated prose. Preparing a comparison may benefit from AI, provided that the differences can be checked against both sources. The proposed scope should name the authorised inputs and the situations in which the system should give no answer. It should also explain how a person will resolve that uncertainty.

An initial evaluation should use synthetic examples with expected answers defined before the test. It needs incomplete pages, conflicting dates and questions that the material cannot answer, alongside straightforward cases. Choosing only attractive demonstrations would conceal the work still needed. A model that produces a convincing sentence is not necessarily producing a usable answer to the business question.

The assessment should show correct results, incorrect results and abstentions separately. Review time also matters: checking a plausible but unsupported answer may take longer than completing the original task. Changing a model or its instructions calls for another comparison against the same reference cases. A construction schedule can then reflect the findings rather than a standard promise about how many weeks AI takes.

Possible applications include a source-linked summary, a draft based on an approved letter, or a comparison between a declaration and its supporting documents. Each needs a named reviewer and a clear destination. An internal search must respect the reader's access to the underlying documents; returning a summary does not remove that requirement. Corrections should remain understandable when the source or the system changes later.

For due diligence, Vedetta is being developed for all professions subject to the relevant obligations in Monaco. Its assistance functions and public self-service journeys remain work to implement and verify. The intended principle is human judgement: suggestions prepare an examination, while an authorised professional decides. A demonstration should never imply that an organisation or a client has been approved by software.

An AI project needs a clear account of where documents go, which services can read them and how long intermediate results remain available. That includes the model, technical logs and backups, as well as the application itself. We address those choices with the organisation before introducing real material. A first evaluation can use fictional examples that reproduce the task, allowing the team to assess usefulness and errors without opening its client files.

The economic assessment should include document preparation, model calls, failed attempts and the remaining review. A price per request is not the cost of a usable case. Quality and delivery speed both matter. Training a specialised model is an option to justify through evidence, not an assumption or an implied permission to reuse the organisation's documents.