Untangling integration debt in a multi-site industrial group

Context

A multi-site industrial group operating across manufacturing, logistics, and distribution had grown over 15 years through acquisitions and organic expansion. Each site, each business unit, and each function had adopted its own systems: ERPs, MES platforms, WMS solutions, CRM tools, BI dashboards, and custom-built applications.

Over time, these systems had been connected to each other through direct, point-to-point integrations: database-to-database replication, file drops on shared drives, custom API calls without contracts, scheduled batch jobs with undocumented dependencies.

The result was a web of 120+ integration links between 43 systems, with no central visibility, no documentation, and no one who understood the full picture.

The problem

The integration debt had become the primary bottleneck for the entire organization:

• Every new feature was a risk: changing one system could trigger cascading failures in systems nobody expected to be affected
• No one owned integration: each link was "owned" by the team that built it, but most of those team members had left
• Data inconsistency was everywhere: the same customer, the same order, the same inventory level showed different values in different systems
• Incident resolution took days: when something broke, tracing the root cause across undocumented integrations required manual investigation by senior engineers
• New projects were blocked: a planned e-commerce platform could not launch because nobody could guarantee that order data would flow correctly through the existing integration web

Management had tried partial fixes: hiring an integration developer, buying middleware, documenting existing links. None of it worked because the problem was architectural, not technical.

What CMX delivered

Phase 1 - Integration landscape cartography (4 weeks)

We conducted a systematic mapping of the entire integration landscape:

• Identified and cataloged all 43 systems with their data ownership boundaries
• Traced all 120+ integration links: protocol, direction, frequency, data payload, error handling (or lack thereof)
• Classified each link by criticality (business impact of failure) and quality (contract clarity, monitoring, error recovery)
• Produced a dependency graph revealing 7 critical single-points-of-failure and 23 circular dependencies

This phase alone produced immediate value: the leadership team saw the full picture for the first time and could make informed decisions about where to intervene.

Phase 2 - Target integration architecture (3 weeks)

We designed the target state:

• Event-driven backbone: an event broker (Apache Kafka) as the central nervous system, replacing direct system-to-system connections with publish-subscribe patterns
• API contracts: standardized OpenAPI contracts for all synchronous interactions, with versioning and backward compatibility rules
• Data contracts: explicit schema definitions for every data exchange, with validation at the boundary
• Integration governance: clear ownership model (each domain owns its published events and APIs), review process for new integrations, monitoring standards

The target architecture was not theoretical. Each design decision was validated against the group's operational constraints: budget, team capabilities, migration risk tolerance, and business continuity requirements.

Phase 3 - Phased migration execution (16 weeks)

We executed the migration in priority order, starting with the highest-risk, highest-impact integration paths:

Wave 1 (4 weeks): Order-to-fulfillment chain: ERP → MES → WMS. Replaced 12 point-to-point links with event-driven flows. Immediate impact: order status became consistent across all systems within seconds instead of hours.

Wave 2 (4 weeks): Customer data synchronization: CRM → ERP → BI. Replaced file-based batch replication with event-sourced customer updates. Eliminated the "different customer data in every system" problem.

Wave 3 (4 weeks): Inventory and logistics: WMS → logistics partners → distribution. Replaced custom FTP-based integrations with API contracts and event notifications.

Wave 4 (4 weeks): Remaining systems: BI feeds, compliance reporting, HR data flows. Migrated remaining integrations and decommissioned legacy middleware.

Each wave followed the same discipline: map the current state, implement the new integration, run both in parallel for validation, cut over, decommission the old link.

Phase 4 - Governance handover (3 weeks)

We established the operational governance layer:

• Integration monitoring dashboard with real-time health status for all flows
• Alerting and incident routing based on integration criticality classification
• Integration review board process for approving new connections
• Documentation standards and integration catalog maintained as living artifacts
• Training for the integration operations team

Results

• 120+ point-to-point links replaced with governed, contract-based integrations
• Integration-related incidents dropped 74% in the first 3 months after completion
• Feature delivery time reduced by 60%: teams could ship changes to one system without weeks of impact analysis
• E-commerce platform launched within 8 weeks of integration completion - it had been blocked for 14 months
• Data consistency achieved across all critical business entities (orders, customers, inventory) for the first time in the organization's history

Why this matters

Integration debt is invisible until it becomes the bottleneck for everything. Unlike technical debt in application code, integration debt compounds across organizational boundaries - every new system, every acquisition, every process change adds links to a web that nobody governs.

The solution is not more middleware or more developers. It is architectural: defining clear boundaries, explicit contracts, and governance mechanisms that prevent the web from regrowing.

CMX brought the integration architecture discipline to treat this as the structural problem it was, not the series of tactical fixes the organization had been attempting for years.