GGenerali CentralExecutive Cockpit

Tech Overview

The architecture under the cockpit — how data flows from many disconnected line, channel and claims systems into one governed truth, how that truth becomes a decision, and how two 360s still deliver over real data gaps.

Generali Central Insurance Company Limited · FY25 (Mar'25, audited)
Mid-tier private multiline general insurer — Top-10 private (rank ~10)
2,644 employees · 167 branches · 21,000+ agents
Under the hood

One governed brain over
many disconnected systems.

No big-bang migration. The platform federates each line, channel and claims system, resolves it to one ontology, and serves a single trusted number — then turns that number into a decision, and the decision into an owner's action.

10/12
Sources fresh
3,836,503
Records governed
4/10
Lines on BaNCS grain
56%
Premium at branch grain
Technical architecture

The governed stack — eight tiers, federated not centralized

Each line, channel and claims system stays where it is. The platform layers ingestion, master-data resolution, a shared ontology and a semantic layer on top, then serves one governed truth to the apps and AI. Data flows top → bottom.

Sources
12 systems of record
TCS BaNCS core · line ledgersPolicy Admin · claims systemsIRIS agent platform · CBS bridgeCRMHRIS · payrollIRDAI / IIB filings · news
Ingestion
adapters · lineage · SLA
Source adaptersCDC & batch loadsFreshness / SLA monitorLineage capture
Store
raw → curated
Raw landing zoneCurated storeVersioned snapshots
MDM · Resolution
many codes → one node
Entity resolutionGolden recordsSurvivorship rulesDedup · term-conflict
Ontology
the knowledge graph
T-Box · 10 classesA-Box · instancesTyped predicatesBranch = keystone
Semantic
defined once, federated
Metric definitionsGrain tagsFederation engineAllocation + confidence
Serving
governed access
Governed metrics APIQuery layerReconciliation tests
Consumption
apps + intelligence
360 views · Next.jsExec briefs · deterministicAzure OpenAIAgentsweb-grounding
Data flow

How one record travels — source to served truth

A single transaction's journey through the stack. A confidence flag and a reconciliation tie-out ride along with it the whole way.

1
Extract

Adapters pull each line, channel & claims system's events on schedule / CDC.

2
Land

Raw records stored verbatim, with lineage + timestamp.

🧩
3
Resolve

Codes matched to one canonical entity.

🧬
4
Model

Mapped onto the ontology — classes & relationships.

📐
5
Define

Native fields → governed metrics; estimates flagged.

🔌
6
Serve

One metrics API; reconciliation gates the numbers.

🔭
7
Consume

360 views, exec briefs & AI read one truth.

🏷 A confidence flag (Actuals / Allocated / Region-only) and a reconciliation tie-out travel with every value — so a number is never shown without knowing how bankable it is.
From data to action · the decision flow

How one number becomes a decision

The governed truth doesn't sit in a warehouse — it routes itself to the right view, the right action, and the right owner.

The agentic layer

An agent on every value pillar

The four value-creation pillars don't just have dashboards — each has a standing agent that reads its governed data products and recommends the next move. Same ground truth, automated.

🏥Profitable Health & Motor Growth
Watches

the ~99% health loss ratio, motor OD/TP economics & GWP vs plan

Grounds on
business_unit · service_line · signal · kpi
Acts in Profitable Health & Motor Growth
📉Combined-Ratio Turnaround (below 110%)
Watches

the combined ratio, loss ratio, EOM cap & the investment-income offset

Grounds on
kpi · vcp · synergy_prog · signal
Acts in Combined-Ratio Turnaround (below 110%)
🏦Bancassurance & Distribution (Central Bank)
Watches

the Central Bank bancassurance ramp, channel mix & policy retention

Grounds on
customer · customer_source · site · kpi
Acts in Bancassurance & Distribution (Central Bank)
🤖Digital, AI & Fraud Control
Watches

claims leakage (8–10%), digital issuance & AI adjudication vs target

Grounds on
data_source · ops_metric · project · kpi
Acts in Digital, AI & Fraud Control
The hard part

Two 360s that work before the data is clean

Some lines and channels aren't fully on the common BaNCS grain yet, so the product→line mapping and branch-grain detail are incomplete. These views still answer — by resolving, allocating-and-flagging, then reconciling. The estimate is labelled, never hidden.

🗂Org Roll-up 360
Open →
The gap
6 of 10lines not yet branch-grain

Some channels report at their own grain — so the product → line → legal-entity rollup is partly missing or inferred.

How the platform bridges it
1
Resolve. AI maps each legacy product / channel / leader code to one canonical org node.
2
Allocate + flag. Where a line isn't mapped, premium is disaggregated from its zone on learned drivers — and marked an estimate.
3
Reconcile. Allocated parts must foot back to the line total; breaks are surfaced, not hidden.
~56%grain coverage

of premium already at true branch grain; the rest labelled & closing as lines move onto BaNCS

📍Branch & Distribution 360
Open →
The gap
10 clustersallocated or region-only

For branches not yet on the common grain, provider-network and premium detail isn't available at branch grain — so the branch twin would otherwise be blank.

How the platform bridges it
1
Estimate. Branch figures are modelled to ~89% coverage from zone totals and network signals.
2
Flag confidence. Every estimated branch carries an Actuals / Allocated / Region-only badge and a coverage %.
3
Flip to actuals. As each cluster cuts over to BaNCS, its branch grain rises and estimates become ledger actuals.
~89%grain coverage

avg branch-grain coverage today — transparent where it's modelled

The harness catches the gaps
12 of 14 governed identities tie out to the cent

Reconciliation runs live on the data. The 2 known breaks below are the plant-grain gap surfaced on purpose — exactly what a CFO or auditor wants flagged, not buried.

Open Data Health →
⚠ flagged
Network nodes = Σ cluster providers
⚠ flagged
NEP = Σ branch-cluster NEP
12
tie to the cent