Pilot · Agriculture and manufacturing

An enterprise assistant separated advisory answers from protected transactions

BluePi built a usable demonstration that routes advisory questions to approved knowledge and protected business requests to authorised enterprise tools. The pilot shows the interaction and control pattern. Draft metrics and testimonials are excluded.

Two user groups needed one clear entry point

Field users need crop and product guidance in a simple conversational form. Sales and operations users need current order status and credit-limit information from enterprise systems. Before the pilot, these questions could require a field specialist, static documents, enterprise screens, email, or messaging groups.

The design goal was one interface with separate capability and permission boundaries. A shared conversation must not give every user access to every backend.

Demonstrated scope

BluePi designed and built the conversational interface, intent routing, modular tool pattern, Google Cloud deployment, Gemini integration, and SAP OData connectivity support. The router identifies an advisory request, order-status request, credit-limit request, or need for clarification.

Advisory questions use approved knowledge. Transactional requests use a narrow tool only when the user, endpoint, and credentials are authorised. The modular pattern can support more tools later, but future tools are not part of the demonstrated result.

One conversational entry point
Routing for advisory and transactional intents
Approved knowledge for advisory answers
Authorised order-status and credit-limit lookups
Google Cloud deployment and enterprise connectivity support

Security and review boundaries

The user interface does not grant backend access. The application checks the user and requested capability before it invokes a protected tool. Advisory content cannot call an enterprise endpoint. A transactional tool receives only the approved inputs for its function.

A rollout must confirm user groups, roles, audit requirements, retention, approved knowledge owners, and escalation. These controls were marked for customer validation in the source. This story therefore describes the pilot pattern, not an approved production security model.

Supported result and missing evidence

The source supports a usable demonstration of one assistant that separates advisory knowledge from authorised enterprise lookups. It does not support a public claim for users served, languages, response time, manual escalations, or time saved.

The source also contains a sample field vignette, outcome placeholders, and draft testimonials. None are used here. A next pilot should measure answer grounding, correct routing, permission denials, lookup accuracy, user completion, escalation, and reviewer corrections.