What makes a useful AI application? Inside Ask BrisAI

Start with the job the software needs to do

A useful AI application connects a person’s goal with information and actions they can understand. The model is one component. The interface, source material, integrations and operational decisions determine whether the complete experience is useful.

Ask BrisAI is a studio example: a visitor asks about our work or a business problem, receives an answer grounded in an approved knowledge collection and can follow a relevant source or start a conversation. The task is specific enough to design and test.

Design the experience around the information

The application retrieves from a curated collection of studio information and practical AI guidance. Its custom Python service combines vector and full-text search to find relevant passages before generating an answer.

That creates an important maintenance responsibility. Changing a service or a price means updating the knowledge collection and ingesting it again. A polished interface cannot compensate for stale source material. Decide who owns that information before launch.

Give the model a structure the software can check

Ask BrisAI separates answer text, citations, opportunity references and handoff in a structured response. The service checks citation identifiers against retrieved material and catalogue references against allowed entries before the interface presents them.

This is application engineering as much as model prompting. The interface needs useful loading, error and retry states; the service needs predictable contracts. For a business tool, the same principle applies to permissions, records and any consequential action that follows an answer.

Evaluate the complete experience

The recorded 21 September 2026 evaluation ran 89 cases across business questions, general questions, scenarios, follow-ups and adversarial inputs. Citation identifiers resolved in all 89 cases; expected-source retrieval succeeded in the 54-case subset where an expected source was defined.

Those results answer particular test questions. They do not make every response correct. The same run recorded three failures, and its model-judge fully-grounded score was 74.2%. The linked project publishes the conditions and the complete dated report so readers can distinguish citation validation from answer quality.

Plan for the application after launch

The surrounding service includes durable rate and budget limits, redaction, feedback and tracing. These controls help make usage and failures visible. They belong in the project scope alongside the interface and AI behaviour.

For your project, agree the intended tasks, source ownership, acceptance criteria, running-cost assumptions and support responsibilities. Begin with a focused application, validate it against the work people need to do and extend it as the evidence supports the next step.

Sources and further reading

• Ask BrisAI project and dated evaluation
• BrisAI evaluation methodology