Turning Fragmented Data Into Conversational Answers
A leading Australian enterprise needed its marketing team to answer complex questions spanning product, incident, and customer-interaction data without relying on static reports or a handful of specialists.
Skillfield designed, built and continued to evolve an agentic AI assistant that understands natural-language questions across multiple business domains, scaling the architecture as complexity grew while embedding permission-aware security throughout.
The Problem
Marketing teams in large, complex organisations often need answers that span many disconnected systems. In this case, status, incidents, outages, product, customer channels and interactions were each held in separate data sources. Getting a single answer typically meant relying on static report, manual database queries, spreadsheets and a small number of specialists who understood where the data lived and how it connected.
As the business began asking increasingly complex, cross-domain questions, this approach became difficult to scale. It slowed decision-making and left the organisation dependent on individual knowledge rather than a repeatable, self-service capability that could be used more broadly across the team.
The Solution
Skillfield designed and built an agentic AI assistant that provides the marketing department with a natural-language way to query data and receive structured responses through a purpose-built web interface.
As new data domains were introduced, Skillfield evolved the solution into a multi-agent architecture, with specialist agents coordinated by an orchestrator, and continually refined to improve accuracy, speed and efficiency.
A domain-agnostic semantic planner was introduced to support the addition of new domains without requiring the underlying system to be redesigned, while ambiguity-resolution capabilities helped the assistant interpret and respond to increasingly complex questions.
The introduction of a more advanced reasoning model also required careful prompt and system engineering. Skillfield refined the assistant’s handling of ambiguity to reduce unnecessary clarification questions, improve response efficiency, and optimise token usage. Further prompt engineering ensured specialist agents delegated tasks reliably, improving both the quality of the results and the efficiency of the overall workflow and token usage.
The Outcome
The assistant is now in production, giving the marketing team a single, conversational way to interrogate data that previously required multiple reports, manual queries and specialist knowledge.
Its multi-agent architecture allows new data domains to be added without a rebuild, while permission-aware prompting helps ensure the AI system only works with information each user is authorised to access. Ongoing tuning has improved response quality, reliability and latency as the solution has matured.
Adoption continues to grow across the team, with usage being monitored as the assistant becomes increasingly embedded in day-to-day decision-making and ways of working.
