/ 成功案例 /
Insurance: MAPFRE industrialises its software factory with MIND, built with Claude Code
- Insurance
- Software delivery
- AI governance
- Regulated industries
MAPFRE and Thinkia built MIND, the insurance group's governed software delivery platform. The methodology ran with Claude Code as its orchestrator while the platform itself was being built. In production: 37-42% less development effort than the traditional method, with quality held constant.
- 37-42% Less development effort than the traditional method, quality held constant
- 137 Commits in 7 days building the platform itself
- 567 User stories analysed one by one in the first line of business
- 100% Artefacts signed by a named person—zero gates self-signed by an agent
Context
MAPFRE needed to industrialise how it builds software, not merely to speed it up. Requirements lived across thousands of user stories with no traceable link to the code implementing them; specification, technical design and coding were manual and serial; and the core systems cannot be rewritten from scratch.
Operating under DORA and EU AI Act obligations, the group could not adopt a generic coding copilot: everything generated by AI had to be auditable, attributable and reproducible, with human accountability preserved at every step. The question was never “how much code does the AI write?” but “who answers for it?”.
Collaboration
MIND is what the two organisations built together. Thinkia brought its AI-SDLC methodology and its Pulse platform; MAPFRE brought its own delivery know-how — engineering standards, corporate stack, regulatory framework and decades of insurance systems. The result is not a product installed on top of the customer: it is the group’s own platform, sharing Pulse’s core and catalogue but composed for its estate. A governed software factory, not an assistant.
The methodology decomposes into version-pinned skills exposed as MCP tools, so any MCP host can run the pipelines natively. Claude Code is the first-class host and, throughout the build, acted as the orchestrator of the methodology: the platform was built by applying its own process to itself, before any automatic runner capable of doing so existed.
Work is routed by a declared model tier — reflex, reasoning or deliberation — naming the cognitive effort a step demands rather than any one vendor’s model. A second, independent axis is component criticality, which escalates the effective tier. In the most critical zone no person writes code: the AI generates under maximum scrutiny, the tests are derived from the specification by a different skill that never sees the generated code, and a person reviews line by line and signs.
Agents transcribe, they do not interpret. If a business decision is not in the Build Spec, the agent stops and asks. Agents never sign: every quality gate is signed by a named person, and a semantic graph ties each requirement to each function and each data entity, so the impact of a change is calculated rather than estimated.
Outcome
Delivery efficiency, in production: a 37-42% reduction in development effort against the traditional method, with quality held constant. Delivered in staggered waves: Definition (6 weeks, 4 pilot projects) → Development (4 weeks) → operationalisation in DevOps.
Building the platform, as evidence of the method (MIND, June 2026): 137 commits in 7 days; 240 specification artefacts, 24 architecture artefacts and 67 review artefacts, produced with Claude Code orchestrating the method end to end. The first substrate band was specified and built from specification to code with build-validate green — architecture lint, type checking and test gates passing — with zero gates self-signed by an agent.
In the Claims line (wave 0): 567 user stories analysed one by one, 133 classified into the water-damage MVP, some 90-100 effective operations, and 41 in the first wave.
Governance results: 100% of artefacts signed by people, with navigable requirement → function → data traceability. Every run is budgeted and recorded in an immutable ledger, with the model actually used noted in an AI-BOM.
Targets in flight, which are not results: ≥60% reduction in lead time from idea to production without relaxing gates, and 60-80% reduction in the pre-signature validation cycle through multi-perspective deliberation.
Real status, stated plainly: MIND is an MVP with real projects on it, not a consolidated platform nor a deployment at scale. The platform routes model calls through MAPFRE’s corporate gateway and does not itself select a model; the Claude angle of this story is Claude Code as the orchestrator of the methodology, which is what is verified.
Case metrics correspond to specific projects and do not constitute a promise of results; each context produces its own range.