AI & Modernization

IBM watsonx Code Assistant for Z (WCA4Z) achieves a median structural quality score above 80% and functional correctness above 75% across enterprise COBOL-to-Java deployments — the most documented results available. A 2025 academic study found that summary augmentation (adding natural language descriptions to source code before translation) improves 36% of benchmark samples and 50% of low-scoring enterprise samples. In direct comparisons, WCA4Z substantially outperforms GitHub Copilot for mainframe code: Copilot produces 'plain class skeletons' lacking business logic, while WCA4Z delivers context-aware translations with proper database handling and CICS integration. Manual refinement effort is substantially reduced but not eliminated — expect 20–30% human review on any enterprise COBOL migration.

Published data is sparse, but: broader LLM translation research (October 2025) shows GPT-4-class models exhibit 33–60% hallucination rates in general translation tasks, adding or removing content not present in source material. The best-performing specialized models (2026 Vectara benchmarks) achieve sub-1% hallucination rates. Most enterprise-grade tools fall into a 'medium hallucination group' with 2–5% error rates. For production code, this means automated evaluation is non-negotiable — manual review of AI-translated code by engineers who understand both the source language semantics and target language idioms is required. Do not deploy AI-translated code without a regression test suite.

Yes, with documented results. UnitTenX (an open-source AI multi-agent system, 2025) achieved 100% line coverage on a real-world C codebase starting from 0% coverage, covering 186 of 199 functions (93.5%) automatically. Salesforce's production deployment using Cursor AI cut legacy code coverage time by 85% for repositories with less than 10% initial coverage. The Salesforce approach combined AI generation with mandatory human oversight: engineers reviewed all generated code, checking test intentions through AI-generated JavaDoc comments, with SonarQube as the final validation gate. Where AI-generated tests fail: business logic validation in undocumented legacy systems — the AI can cover code paths but cannot verify that the business rules encoded in those paths are correct.

Data fragmentation is the #1 barrier by a significant margin: 82% of enterprises report fragmented data blocks AI accessibility (Cisco global study, 2025), 69% report poor data quality limits decisions, and 45% cite fragmented/unstructured data as the top AI adoption blocker. Legacy system integration is the #2 barrier: 60% of AI leaders identify legacy system integration as their primary barrier to agentic AI adoption (Deloitte Tech Trends 2026), and Gartner predicts 40% of agentic AI projects will be cancelled by end of 2027 due to infrastructure constraints. The failure pattern: AI pilots succeed in isolated environments with clean test data, then fail when connecting to legacy systems without APIs or real-time data capabilities.

AI-ready enterprises deploy three core integration layers: API integration layers (80% adoption among AI-ready companies), data lake/lakehouse architectures (77%), and enterprise data warehouses (72%). Beyond technology, AI-ready infrastructure requires: data governance with operational data contracts, freshness SLAs, and error budgets; real-time event-driven pipelines that activate instantly on business events; federated data mesh with centralized guardrails where regional teams publish data products under shared contracts; and policy-as-code embedding consent, residency, and retention rules. Legacy infrastructure fails on all of these: siloed CRM, POS, ecommerce, and service data creating blind spots, no lineage or quality controls, and 'one-off' integrations lacking reusable contracts.

BCG research (October 2025) quantifies the gap: AI future-built companies achieve 5x the revenue increases and 3x the cost reductions compared to AI laggards. However, only 5% of firms worldwide are 'AI future-built,' while 60% report negligible gains despite substantial investment. McKinsey (June 2025) found that over 80% of companies embracing AI see no real productivity gains — most remain in 'copilot mode' rather than deploying agentic systems. Accenture data: competitors successfully leveraging AI are cutting costs by 30% and boosting productivity by 50%. The window for catching up is closing: in a world where competitors condense a month's work into a single day using agentic systems, the cost of remaining in pilot mode may soon exceed the cost of doing nothing.

The typical AI readiness modernization path has three phases: Phase 1 (3–6 months) — data audit, master data governance, API layer over legacy systems, cloud rebalancing to get workloads on infrastructure that supports AI inference; Phase 2 (6–12 months) — data lakehouse implementation, real-time event pipelines, MLOps CI/CD pipeline build, model monitoring infrastructure; Phase 3 (12–24 months) — federated data mesh, policy-as-code, retraining infrastructure on proprietary data, agentic system deployment. The timeline compresses when data quality is already high (rare) and extends when ERP/CRM systems are on-premise with no APIs (common). Only one-third of enterprises are currently fully equipped to deploy AI at scale (NetApp, October 2025).

A 2025 consolidated MLOps lifecycle framework identifies the most common organizational gaps: lack of defined roles and skill requirements for MLOps (teams have data scientists but no ML engineers), insufficient CI/CD pipelines for ML model deployment (models are trained but deployment is manual), inadequate monitoring and observability for production models (no alerting when model performance degrades), absence of data governance frameworks (models trained on stale or biased data), and no clear resource allocation for MLOps at various maturity levels. The enterprise AI readiness framework requires an 80%+ data quality score, executive sponsorship, and defined ROI metrics before model deployment — criteria that most organizations pursuing AI pilots cannot meet.