raises request
builds data model
written & tested
dashboard
cycle
receives insight
Even the simplest analytical question — "what is the claims approval rate this month by product line?" — required raising a formal request, waiting for a data engineer, and then an analyst before any answer appeared.
The data engineering team spent the majority of their capacity building ad-hoc intermediate data models for business requests — work that consumed skilled engineers and prevented investment in strategic data infrastructure.
By the time a fraud pattern report arrived, the fraudsters had moved on. By the time a claims trend dashboard was built, the operational window for intervention had closed. Stale insights drove reactive decisions.
A follow-up question — "now break that down by region" — triggered the entire cycle again. Four business teams, each needing iterative analysis, were multiplying the pipeline burden with every question they had.
Our fraud team was sitting on top of a gold mine of transactional data. But by the time we could actually see it, the patterns we needed to act on had already changed. We were always one week behind where we needed to be.
— Chief Data Officer, Insurance Enterprise (name withheld)
The business user's natural language question is classified by an LLM agent — determining whether it requires SQL/data retrieval, statistical analysis, metric calculation, or a general response. Prevents unnecessary database calls and routes each query to the optimal processing path.
LangChain · Claude 3.5 Sonnet · Amazon BedrockFusionReport loads the enterprise's database schema metadata — fact tables, dimension tables, foreign key relationships, column types. The LLM performs schema linking: identifying which tables and columns are relevant to the question and mapping the correct join paths through the star schema.
Database Metadata Only · No Raw Data Sent to LLMThe LLM decomposes complex questions into sub-questions, generates optimised SQL for each, and assembles the final query. Self-correction validates the SQL before execution — catching join errors, column mismatches, and aggregation mistakes. Supports PostgreSQL and MySQL dialects natively. Every query is read-only — data modification is blocked by design.
Multi-step SQL generation · Dialect-aware · Self-correctingThe validated SQL executes directly against the enterprise's database using an async query engine within their own AWS environment. Results are returned as a structured DataFrame. No data leaves the organisation's compute boundary at any point. The enterprise's existing database security, access controls, and network policies remain fully in effect.
AsyncPG · SQLAlchemy · Runs in your VPCThe resulting DataFrame is passed to a second LLM call on Amazon Bedrock, which generates production-ready Plotly visualisation code — selecting the optimal chart type (bar, line, scatter, pie, KPI card) for the data structure and question context. Charts render instantly in the browser with proper axis labels, titles, and colour schemes.
Plotly Express · Auto chart selection · Instant render-- FusionReport: Schema linked → fact_claims JOIN dim_agents JOIN dim_claim_types -- Sub-question decomposition: suspicious count + total count + ratio calculation WITH agent_claim_summary AS ( SELECT da.agent_id, da.agent_name, dct.claim_type_name, COUNT(*) AS total_claims, SUM(CASE WHEN fc.fraud_flag = 1 THEN 1 ELSE 0 END) AS suspicious_claims, ROUND( SUM(CASE WHEN fc.fraud_flag = 1 THEN 1 ELSE 0 END) * 100.0 / COUNT(*), 2 ) AS suspicious_pct FROM insurance.fact_claims fc JOIN insurance.dim_agents da ON fc.agent_key = da.agent_key JOIN insurance.dim_claim_types dct ON fc.claim_type_key = dct.claim_type_key JOIN insurance.dim_date dd ON fc.date_key = dd.date_key WHERE dd.full_date >= CURRENT_DATE - INTERVAL '90 days' GROUP BY da.agent_id, da.agent_name, dct.claim_type_name ) SELECT *, CASE WHEN suspicious_pct > 15 THEN '⚠ FLAGGED' ELSE 'OK' END AS risk_flag FROM agent_claim_summary ORDER BY suspicious_claims DESC LIMIT 10;
Real-time fraud pattern analysis — patterns identified same day, not next week. Suspicious agent flagging moved from weekly review to continuous monitoring.
Claims velocity dashboards that previously took 3 days to produce now refresh on demand. Adjuster performance analysis became a daily operational tool instead of a monthly report.
Actuarial teams gained self-service access to loss ratio and severity analysis — reducing their dependence on the data engineering team for routine analytical queries by an estimated 70%.
Regulatory reporting that required full data engineering cycles now runs on demand. Compliance exceptions are identified in real time — not discovered during quarterly audit cycles.
Fraud signals that previously surfaced in weekly review reports are now identified the same day the transactional data lands — giving the fraud team an operational window to act before patterns escalate.
Data engineering time previously consumed by ad-hoc business reporting requests has been redirected to strategic data infrastructure — an immediate uplift in engineering productivity and job satisfaction.
FusionReport operates entirely within the enterprise's AWS environment. Only database schema metadata is passed to the LLM — raw transactional data never leaves the organisation's compute boundary at any point.
Decision-makers across all four teams now enter meetings with data rather than waiting for it. The shift from reactive to proactive decision-making — driven by real-time transactional intelligence — was noted by the CDO within the first month.
The data engineering team used to spend most of their time building reports that should have been self-service. FusionReport gave four business teams the ability to answer their own questions — and gave our engineers their careers back.
— Chief Data Officer, Insurance Enterprise (name withheld)
AWS Generative AI Competency Partner. FusionReport is P3Fusion's Generative AI analytics platform for structured data — turning any star schema, data warehouse, or relational database into a self-service insights engine on Amazon Bedrock.
Ready to see FusionReport on your warehouse? Book a demo with the team.
