Choose a process you understand
Look for repeated work using existing information: finding internal instructions, comparing documents or preparing a review. Appoint an owner who understands a correct result and can judge mistakes.
Avoid a first project that replaces a core system, reorganises the entire archive and automates critical decisions at once. Separate dependencies. Better search or data structure may sometimes solve the problem without a generative model.
Test Bulgarian in the actual task
Fluent Bulgarian conversation does not prove accuracy with contracts, manuals or internal abbreviations. Prepare realistic questions, including mixed Bulgarian and English terminology, numbers, negation and cross-document references.
Evaluate source retrieval separately from answer formulation. Some errors begin in text extraction or embeddings. A subject-matter reviewer should assess the result, not merely its linguistic polish.
Prepare data before budgeting for hardware
Identify where documents live, who owns them and which people may use them. Check for scans, duplicate versions and missing permissions. Source preparation often shapes scope more than the model choice does.
Use synthetic or pre-approved material for the first demonstration. Do not upload customer agreements or personal data into a public tool merely to see whether it works.
Separate technical and organisational work
IT evaluates integrations, access and operations. The process owner defines usefulness. Responsible legal and regulatory specialists review applicable conditions. None of those roles is replaced by an “EU-hosted” label.
Agree who accepts the pilot, how users will be trained and who maintains sources after launch. This makes proposals comparable and exposes hidden dependencies.
What to put in the first brief
A short brief can describe one workflow: who asks the question, what answer they need and where the correct information comes from. List permitted documents, approximate user numbers, deployment constraints and the person able to judge the result. Do not attach confidential material to a public inquiry.
Add several difficult cases: an outdated document, a missing answer, mixed terminology and a user without the required access. These are a more useful starting point than a general request for “AI in the company”. They support a practical discussion of pilot scope, acceptance criteria and dependencies that need preparation first.
End the assessment with a decision
A useful outcome is an evidenced “yes”, “no” or “after preparation”. Record measured quality, cost assumptions, risks and the next bounded step. A pilot does not need to become a deployment to prevent an unsuitable investment.