MDM in Insurance 2026: From Legacy Silos to Generative-AI-Ready Data
In 2020, the challenge was centralising master data. In 2025, the challenge is for that data to feed autonomous agents, corporate RAGs and language models without hallucinating non-existent policies. 78% of Spanish insurers admit that their biggest brake on scaling generative AI is not the model —it is the quality and governance of their master data (Source: ICEA & Mapfre Economics, Informe Sectorial 2024). It is the starting point of any BAOSS IT consulting project for insurers.
At BAOSS we have guided four of the ten largest Iberian insurers through that transition. This article sells no smoke: it describes the real problem we see in 2026-2027, the architecture that works today and metrics from completed projects (anonymised, verifiable).
The Real 2026 Problem: “We have MDM, but AI doesn’t trust it”
Most insurers already deployed classic MDM (IBM InfoSphere, Informatica, Talend) between 2010-2018. The result: technically correct golden records, but semantically incomplete for AI. Three patterns repeat in every audit we run:
- Missing attributes for LLM/RAG: The MDM knows “name, tax ID, policies”, but not “claims history with free-text narrative”, “channel preferences”, “fraud propensity score updated weekly” or “vector embeddings of incoming correspondence”.
- Unacceptable latency for autonomous agents: A CrewAI agent resolving claims in real time needs to resolve identity (is this customer the same one from the 2023 claim?) in <200ms. Legacy MDMs answer in 1.2-3s over synchronous SOAP/REST. Before accelerating, it is worth controlling the AI agents your company has already deployed.
- Governance disconnected from MCP (Model Context Protocol): Data catalogs (Collibra, Alation) do not expose metadata in MCP format. Agents cannot “ask the catalog” what
poliza_vida_riesgomeans or what transformations thebeneficiario_principalfield underwent in 2022. This mismatch is exactly what the Hugging Face case lesson on AI agent governance revealed.
Real case (anonymised): Insurer “NorteVida” (Top-5 Spain, 2.3M policies). Their IBM MDM v11.6 was running “fine” by internal KPIs (duplicates <1.2%, completeness 94%). When piloting an AutoGen agent to automate life-risk claims, 34% of decisions required human intervention because the agent could not resolve identity between policyholder, insured and beneficiary in policies with complex structures (reinsurance, overlapping coverages, notarial amendments). The MDM had the data, but not the versioned temporal relationships or the semantic embeddings the agent needed.
BAOSS IT consulting architecture 2026-2027: AI-Ready MDM
The solution is not to replace the MDM —it is to extend it with a semantic and operational layer ready for agents. The pattern we deploy at BAOSS (4 projects in production, 2 in advanced stages):
1. Real-Time Entity Resolution Layer (Sub-200ms)
We deploy vLLM + Ollama serving fine-tuned embedding models (BGE-M3 adapted to Spanish insurance terminology: “claim”, “accident report”, “expert assessment”, “excess”, “waiting period”) on A100/H100 GPUs in Kubernetes. Entity resolution goes from querying MDM tables to ANN vector search (HNSW) over enriched entity embeddings + deterministic rules for regulatory compliance (GDPR, LSSI, Solvency II Directive).
NorteVida metric: P99 latency 147ms (vs 2.8s before), entity resolution accuracy 98.7% (vs 91.2% pure rules), inference cost €0.00018/resolution.
2. Insurance Knowledge Graph
We build a Property Graph (Neo4j / TigerGraph) synchronised via CDC (Debezium) from the MDM and operational sources (life core, claims, CRM, document management). Nodes: Person, Policy, Claim, InsuredAsset, Intermediary, Coverage. Relationships versioned with valid_from/valid_to and source_system.
This graph exposes MCP endpoints so LangGraph/CrewAI agents can query: “What coverages did policy P-2023-045891 have on the date of claim S-2024-11203?” without SQL, without a rigid API, in natural language translated to Cypher/Gremlin by a Text-to-Cypher agent (GPT-4o + curated few-shots).
3. Continuous Quality Pipeline with Validator Agents
We replace nightly batch jobs (DataStage, SSIS) with AutoGen multi-agent systems that monitor quality in streaming:
- Agent Profiler: Detects drift in distributions of key fields (e.g.
codigo_postalstops matching INE v2024). - Agent Reconciler: Proposes entity merges/splits using LLM + rules, presents them to the data steward via Slack/Teams with explainability (SHAP values + natural language).
- Agent Enricher: Calls external sources (Cadastre, BOE, ICEA, corporate social networks) via MCP tools to complete missing attributes.
NorteVida result: 82% reduction in manual stewardship workload, 3.4-week early detection of quality issues that previously reached BI/AI.
4. Unified Feature Store for ML/AI (Feast + Iceberg)
Master attributes (demographic, contractual, behavioural, embeddings) are materialised in Feast on Apache Iceberg (S3/MinIO) with temporal versioning. The pricing, churn, fraud, next-best-action models consume features consistent between training (offline) and inference (online). The MDM stops being a “reporter” and becomes a first-class feature provider.
2026 Technology Stack (What we actually deploy)
| Layer | Technology | Role |
|---|---|---|
| MDM Core | IBM MDM / Informatica / Semarchy (existing) | Golden record, stewardship, compliance |
| Vector Search | vLLM + Ollama (fine-tuned BGE-M3) + Qdrant/Milvus | Sub-200ms semantic entity resolution |
| Knowledge Graph | Neo4j Aura / TigerGraph Cloud | Complex relationships, MCP endpoints, traceability |
| Agent Orchestration | LangGraph (production) / CrewAI (rapid prototyping) | Stateful multi-agent workflows, human-in-the-loop |
| LLM Gateway | LiteLLM / Portkey (router for GPT-4o, Claude 4 Sonnet, local Llama 3.3 70B) | Cost control, fallback, observability, PII masking |
| Feature Store | Feast + Iceberg + Trino | Consistent offline/online features, time-travel |
| Data Contracts / MCP | OpenLineage + custom MCP Servers (Python/FastAPI) | Versioned data contracts, agent discovery |
| Observability | LangSmith + Grafana + OpenTelemetry | End-to-end traces agent → MDM → Graph → LLM |
BAOSS Project Metrics 2026-2027 (Anonymised, Verifiable)
| KPI | Before (Legacy MDM) | After (AI-Ready MDM) | Projects (n=6) |
|---|---|---|---|
| Time to deploy a new AI use case (days) | 45-90 | 8-18 | -78% |
| Entity resolution latency (P99, ms) | 1,200-3,500 | 90-180 | -94% |
| Entity accuracy for AI (F1) | 0.87-0.92 | 0.97-0.99 | +8-12pp |
| Manual stewardship cost (FTE/year) | 3.2-5.8 | 0.6-1.1 | -82% |
| Project ROI (6 months) | — | 2.8x – 4.1x | Average 3.2x |
| “AI-ready” attribute coverage (% entities) | 12-28% | 89-96% | +3.5x |
The 3.2x ROI at 6 months comes from: (1) life/motor claims automation (+€1.4M/year), (2) reduction of undetected fraud (+€0.9M), (3) freed stewardship FTEs (+€0.4M), (4) new pricing/churn models in production (+€1.1M). Total platform cost (infrastructure + licences + BAOSS services): €1.1-1.7M depending on scope.
Case “Mediterráneo Seguros” (Top-8, 1.1M policies): From Pilot to Production in 11 Weeks
Challenge: Automate 60% of home claims (water damage, breakage, theft) with an autonomous agent. Blocker: master data for property, contents, repair contractor scattered across 7 systems, with no single identity and no vectorised history of expert assessments.
- Weeks 1-3: Data audit + definition of the Insurance Knowledge Graph schema (12 node types, 18 relationships).
- Weeks 4-6: CDC Debezium → Neo4j + BGE-M3 fine-tuning with 45k anonymised historical assessments. vLLM/Ollama deployment on AKS.
- Weeks 7-9: LangGraph agent
ClaimsResolver(3 sub-agents: EntityResolver, CoverageChecker, PayoutCalculator) + graph MCP server + Feast feature store. - Weeks 10-11: Shadow mode (agent proposes, human decides) → A/B test on 5k claims → progressive go-live.
Result at week 14: 67% full automation (vs 60% target), claimant NPS +23pts, average resolution time 4.2h → 47min, unit cost -61%. The legacy MDM is still there —it now feeds the graph and exposes MCP. Nobody touched it. That modernisation without tearing everything down is the approach described in BAOSS, the operating system for business agents.
Mistakes We Avoid (and that you see in RFPs)
Prefer to talk to us directly? Contact the BAOSS team and tell us about your challenge.