Before talking about models, agents or integrations, we need to know if there is a project that is worth building. These ten questions help us move from a broad intuition to a concrete and responsible hypothesis.
What problem do we want to solve?
We are looking for an observable situation, not a generic aspiration. What is slow, inconsistent, hard to find, expensive or impossible to scale today?
How is it currently resolved?
We describe people, steps, tools, expectations and exceptions. The actual process tends to be different from the written procedure and is where the best opportunities appear.
Who knows what it means to do well?
We identify the experts and the criteria they use. Without a definition of quality, we will not be able to train, test or review the system.
What sources do you need?
We locate documents, data, applications and tacit knowledge. We ask what source has authority, how often it changes and who keeps it.
What data should I not use?
We define sensitive information, permissions, contractual limitations and uses incompatible with the purpose. The data boundary is part of the design, not a later configuration.
What can the AI do and what should a person validate?
We separate preparation, recommendation, decision and execution. We place Human Gates where impact, uncertainty or irreversibility demand judgment.
What happens when the case is not normal?
We look for contradictions, absence of information, out-of-scope requests and tool errors. A good system must know how to stop, explain the limit and climb to the right person.
How will we measure if it works?
We define a baseline and indicators related to the problem: time, quality, coverage, risk, adoption or economic impact. We also decide when to review the results.
Who will be the owner after the pilot?
Someone must maintain sources, permissions, tests, and criteria. Without a process owner, the project depends on the provider or degrades when the context changes.
What is the first useful scope?
We choose a boundary small enough to learn and real enough to demonstrate value. It can be a phase, a type of document, a team or a channel. We don't try to solve the whole organization at the same time.
The answers may also indicate that it is not yet necessary to build
Sometimes these questions reveal that it is first necessary to sort documents, define a policy, form the team or clarify the process. It is not a deviation: it is the work necessary for a future solution to make sense.
At Karmina AI Studio we use this framework to decide the next proportional step. A good start is not what incorporates more technology, but what makes the challenge a capacity that can be tested, governed and improved.