Give the AI a task. Then a screen.
Is the starting point of an artificial intelligence project model selection or business need? Our approach through document analysis and corporate assistant examples.
This article contains notes from SARBO's approach to digital product development.
Define the task before choosing technology
When talking about AI integration, the first question is often which model to use. At SARBO, our starting question is which business to support. A support team has different requirements for accessing product information than an operations team for extracting space from documentation.
Describing the task in one sentence makes the scope visible: such as drafting answers to questions based on permitted documents. This recipe also determines in which examples we will evaluate success.
Determine data source and access limit
Which documents the assistant will use, how these documents will be updated, and which user will access which information should be addressed within the scope of the project. Information available to one user may not be available to another user.
It is also necessary to evaluate separately whether the solution is intended to only produce a response or perform an action in the system. Submitting information, drafting, and amending records require different authorizations.
Learn with a little pilot
Choosing a narrow use case for the initial release makes it easier to make decisions during the development process. By agreeing on sample questions, expected outcomes, and acceptance criteria, we can evaluate the pilot. This evaluation includes the sources on which the answers are based and in which cases human approval is required.
For example, instead of connecting a product assistant to the entire company data, we can start it with the documents of a specific product group. In this way, we see the content quality, question diversity and user expectations more clearly.
Connect to workflow
After the pilot, we establish the screen where the solution will be used and its connection with existing systems. A request form, admin dashboard or internal information display can be the right starting point. What records will be kept and how usage will be tracked are also determined at this stage.
We do not think of AI integration as a stand-alone chat box. We aim to develop an assistant that is adapted to the company's processes, has defined powers, and whose results can be evaluated.
The starting point is “which model?” before the question “Which mission do we support?” is the question.
Ask this question before adding AI to your business.
Where should you start with artificial intelligence? We first talk about the task, not the model. Which process repeats? What information is difficult to access? For example, an assistant's job may not be to manage the entire company, but to prepare responses from permitted product documentation. First, choose a small scenario. Identify data sources. Evaluate the result with sample questions. Then connect the solution to the actual workflow.