The State of Play

AI-assisted legacy code migration has solidified at leading-edge maturity: vendors are shipping competitive agentic platforms, consulting firms are scaling AI practices at industrial scale (EPAM certifying 1,300+ architects with 10,000+ target), and deployments span multiple vertical markets. The practice uses AI to analyse, document, and transform systems written in older languages—primarily COBOL on mainframes—into modern platforms like Java and cloud-native microservices. Its urgency is demographic and economic. With 10% of COBOL developers retiring annually, 43% of US banking still COBOL-reliant, and the modernization services market projected to grow from $22.1B (2026) to $50.7B (2033) at 12.6% CAGR, the economics are forcing action. Yet production deployment remains at ~13-14% (per 2025 surveys), and the binding constraints are organisational—semantic validation expertise, behavioral equivalence assurance, and change management—not technical capability. The tools work. Real-world evidence now shows AI accelerates the discovery phase by 2-3x and handles 30-60% of migration work, but the final 40-70% (business logic validation, regulatory compliance, zero-trust testing) remains human-intensive. Scaling them means solving the human problem first.

IBM dominates the vendor landscape through watsonx Code Assistant for Z, which reached v2.8.0 in December 2025 with agentic capabilities that orchestrate multi-step analysis and transformation across mainframe codebases. Its Project Bob initiative consolidates RPG and COBOL assistants into a single platform. That dominance faces intensifying competitive pressure: Anthropic's February 2026 announcement of Claude for COBOL modernisation triggered a 13.2% single-day drop in IBM stock. AWS launched Transform service (GA April 2026) with agentic AI for code analysis and PL/I modernization, demonstrating real-world deployment velocity—ADP, Itaú Unibanco, and Western Union achieved 90%+ reverse-engineering time reductions and 80%+ forward-engineering acceleration on business-logic externalization. AWS Transform's June 2026 Assess and Reimagine workflows add deterministic business rule extraction with full source-code traceability, enabling enterprise-scale auditable modernization. AWS added portfolio-level AI readiness assessment (CAST Highlight AI Acceleration Insights, May 2026) enabling enterprises to systematically identify candidates for agentic intervention. BMC shifts to agentic architecture, capturing institutional knowledge from historical resolutions to produce AI-analyzed application narratives. CAST Imaging and OpenLegacy round out the ecosystem with documented results: 110M COBOL lines analyzed in four weeks by a leading insurer; Thoughtworks delivering financial services firms 4M lines of COBOL/HLASM modernization in four weeks using agentic AI with 80% code comprehension; Meliá Hotels migrated 20+ year mainframe reservation system to AWS microservices (60% cost reduction, 75% time-to-market gain); Bridgestone completed 1.2M-line z/OS COBOL migration in 7 months using AWS Transform with 90% efficiency gains. Heirloom Computing's GA mainframe platform adds credible vendor options beyond IBM, recognized by ISG as leading multi-platform solution across diverse verticals (education, insurance, payments). Market scale is accelerating: global modernization market reached $12.7B in 2025 and is projected to grow to $27.9B by 2034 at 9.4% CAGR, driven by 67% of CIOs citing mainframe skills scarcity as their #1 infrastructure risk. Real-world deployment evidence from July 2026 confirms acceleration: Thoughtworks case study documents a 10-sport data platform migrated from 2-3 year baseline to 3-4 weeks using AI-assisted business logic extraction with per-sport onboarding compressed from 10-15 weeks to under 1 day; consulting firm Kanerika reports 50-60% effort reduction and 90-day delivery on 24+ month baselines across 161 enterprise engagements (2015-2026), with 76% achieving 50%+ metric improvement through foundation-first approaches. Domain-specialized models prove superior: COBOL-Coder domain-adapted LLM reaches 73.95% compilation success vs. GPT-4o's 41.8%, validating necessity of specialized tooling for legacy language translation.

However, mounting evidence documents critical governance and quality barriers, even as late-July and early-August 2026 data reveal progress on semantic validation. Peer-reviewed research (PREreview systematic review, July 2026, PRISMA 2020, 34 empirical studies across 2,847 developers) finds 58% report significant quality degradation when using AI assistants; on legacy code specifically, defect density increased median 23% in AI-heavy modules, and 54% of teams saw no net throughput gain after 6 months—quality gates (test coverage >80%, cyclomatic complexity <10) prevent 67% of debt insertion. A 2026 survey of 200 enterprise SRE/DevOps leaders found 43% of AI-generated code requires manual debugging in production—developers spend 38% of weekly time fixing AI output. Gartner's June 2026 forecast predicts 70% of mainframe exit projects initiated in 2026 will fail due to overestimation of AI capability, with 75% of modernization vendors expected to pivot or exit by 2030. Software Improvement Group's analysis of 30,000 systems (400B+ lines of code) documents that autonomous AI agents produce structurally unmaintainable code carrying 2x the security violation rate of human-written code, with productivity gains vanishing entirely beyond 100K-line codebases—token costs alone can reach €10-15M for a single week's transformation work. Specira's analysis (July 2026) identifies a fundamental failure mode: transpilers translate syntax but not intent; 220 billion COBOL lines carry decades of undocumented business rules and edge-case behaviors that survive in production for years, invisible to transpilers. Speed masks incompleteness—a module passes all tests and reaches production, then fails week 3 on untested behavior that was never discovered by the conversion process. Veracode's GenAI security testing (45% of LLM outputs introduce OWASP Top 10 vulnerabilities, Java reaching 70% failure rate) documents material risk that AI transformation can introduce security defects absent from original systems. The Stack Overflow survey (49K+ developers) shows 80% AI adoption but only 29% confidence in accuracy; 66% spend extra time fixing near-correct output.

Yet August 2026 evidence signals incremental progress on the semantic validation constraint: a peer-reviewed agentic test-synthesis methodology (Locksmith Loop) achieved 91.90% branch coverage on production-grade COBOL programs with deterministic parity checks, advancing beyond question 'can AI translate syntax' to answering 'can we validate correctness at scale.' Vendor convergence accelerates as structural boundaries clarify: Fujitsu and IBM's strategic partnership pairs deterministic conversion (PROGRESSION) with agentic refinement (Bob), signaling market consensus that 1:1 transpilation requires human-guided validation, not autonomous automation. IBM's Q2 2026 earnings reveal 50% of Z17 customers actively deploying Spire Accelerator AI capabilities and watsonx Code Assistant users growing MIPS capacity 3x faster than non-users—quantified adoption momentum that contradicts early-stage characterization. Discovery capability validation accelerates: Claude Code's application to systematic reverse engineering (National Museum of Computing digital archaeology projects) and Thoughtworks' high-velocity business-logic extraction frameworks demonstrate AI's proven role in the 40% of modernization work that is comprehension-intensive. OpenLegacy's CTO analysis refines market positioning: AI excels at discovery phase (time-to-comprehension scaling from months to days), but full automation stalls on context windows, regulatory requirements, and test coverage—the 80% of remaining cost that is organizational, not technical.

Independent testing confirms all mainstream AI tools produce semantically incorrect COBOL-to-Java transformations without expert validation. Practitioner consensus (Infosprint, tokai-is, Unblocked) identifies context debt—missing institutional knowledge about the why behind 40-year-old business rules—as the binding constraint, not code transformation capability. Successful organizations solve the comprehension problem first (building context through hybrid human-AI analysis), then apply deterministic or guided transformation with full behavioral equivalence testing. The UK government's experience illustrates organisational headwinds: legacy systems consume £2.3 billion of £4.7 billion IT budget, with high-risk systems growing 26% annually despite modernization remediation efforts.

• 2023-H1: Vendor deployments (SoftRoad 950+ migrations, CAST 50%+ time savings) establish proof-of-concept; IBM announces Watson GA; adoption barriers identified as organizational (strategy clarity, skills gaps) and technical (LLM semantic limitations).

• 2023-H2: IBM watsonx Code Assistant and CAST Advisors reach GA; however, Gartner notes lack of customer case studies for watsonx validation. Developer skepticism persists around correctness and non-technical barriers (organizational alignment, skills crisis). Real-world adoption remains incremental and risk-averse.

• 2024-Q1: INAIL case study shows 700-app portfolio migration with CAST, proving scale viability. Academic research (ICSE 2024) advances COBOL-to-Java LLM translation. CAST and IBM expand product capabilities (advanced search, new migration advisors). Persistent adoption barriers: programmer shortage (now including C/C++), organizational inertia, need for human expertise in semantic translation.

• 2024-Q3: Enterprise deployments expand: insurance company achieves 80% faster code understanding with IBM watsonx. Independent survey (Kyndryl, 500 IT leaders) reports 86% planning GenAI deployment for mainframe modernization with 114-225% ROI, but 43% lack skills to operationalize. IBM expands on-premises deployment options. Skills shortage intensifies as critical adoption constraint.

• 2024-Q4: IBM releases watsonx Code Assistant GA with multi-language support; AWS endorses CAST Highlight in official guidance. CAST serves hundreds of modernization clients via Google Cloud partnership. Software AG discontinues legacy Unix support, creating platform rehosting mandate. NTT DATA documents persistent GenAI challenges (hallucinations, semantic complexity); category enters mature production phase but remains constrained by organizational and human factors rather than technical capability.

• 2025-Q1: Vendors advance (IBM watsonx v2.x, OpenLegacy on AWS Marketplace), but adoption gap widens: 2025 Arcati survey shows only 13% in production (vs. 86% planning in 2024). Case studies confirm value (110M COBOL in 4 weeks, 66% reverse engineering speedup), but technical validation remains critical: independent testing shows all major AI tools fail semantic correctness in COBOL-to-Java conversion. Market drivers persist (platform EOL urgency, proven ROI), but semantic correctness and organizational skills remain binding constraints.

• 2025-Q2: Vendors mature tooling through Q2: IBM releases watsonx v2.6 (June 27) with AI agents for autonomous COBOL generation and expanded language support; IBM and CAST deepen partnership for enhanced application discovery; OpenLegacy documents 353% ROI in independent Forrester study. Academic and industry research intensifies: FSE 2025 industry paper presents automated testing framework for semantic equivalence validation in COBOL-to-Java translation, addressing production reliability. Generative models' capability for reverse engineering of legacy systems gains technical validation. Category remains in production deployment phase with mature vendor capabilities but sustained adoption constraint from semantic correctness validation requirements and organizational change management complexity.

• 2025-Q3: Vendor ecosystem expands across platforms: IBM releases watsonx v2.7 (August) with business rule discovery and natural language COBOL generation; previews watsonx Code Assistant for i (RPG modernization) with 10M+ LOC training. Academic research accelerates quality assurance: ASE 2025 submissions present automated evaluation systems for semantic correctness in COBOL-to-Java translation. Practitioner and market evidence: Thoughtworks case study documents one-month assessment of 2.2M-LOC system; Stack Overflow survey reveals trust paradox (80% adoption but 29% accuracy confidence, 66% fixing "almost-right" code). Adoption remains nascent (13% in production per Arcati survey), confirming organizational and validation barriers persist despite mature technical capabilities and documented ROI.

• 2025-Q4: IBM advances vendor platform consolidation: watsonx v2.8.0 (December) introduces agentic chat enabling multi-step orchestration across mainframe transformation tasks; Project Bob announcement unifies RPG and COBOL assistants into single platform. Q4 saw primarily vendor announcements and feature releases rather than new deployment case studies; agentic capabilities signal movement toward higher-order automation in legacy analysis workflows. Adoption barriers (semantic validation expertise, domain knowledge, organizational change management) remain unchanged; category remains at leading-edge production phase with mature capabilities but nascent organizational adoption.

• 2026-Jan: Vendor momentum and real-world deployment signal confirmed at end-Q4 2025: IBM CEO highlighted watsonx COBOL-to-Java refactoring driving highest z17 revenue in 20 years (Q4 growth 48% YoY). Market data shows 75%+ organizations now using AI for legacy modernization, with market projected to grow from $24.98B (2025) to $56.87B (2030). Practical deployment evidence from insurance sector and engineering case studies documents both AI-assisted approaches (IBM watsonx reference deployments) and deterministic alternatives (AlfaStrakhovanie ANTLR-based migration). Organizational barriers remain binding: UK government report shows legacy systems consuming half of IT budgets (£2.3B of £4.7B in 2025) with high-risk systems increasing 26% despite remediation efforts, indicating systemic failure to scale modernization. Category consolidated at leading-edge with mature vendor ecosystem, proven economics (353% ROI documented), and demonstrated real-world deployment, but constrained by human validation expertise and organizational change management rather than technical capability.

• 2026-Feb: Competitive disruption enters market: Anthropic announces Claude AI for COBOL modernization (February 23-24, 2026), triggering IBM stock decline 13.2% in single day—largest move in 25+ years—signaling investor concern about competitive encroachment on IBM's mainframe modernization franchise. Market context: 220B lines of COBOL running in banking, government, healthcare; 10% of COBOL programmer base retiring annually. Real-world deployment accelerates: food wholesaler modernization case (Keyhole Software) documents 20-30% development acceleration via AI tools; practitioner analysis shows iterative agentic workflows (engineer-agent refactoring, critic-agent validation) enable traceable, testable modernization. Analyst competitive analysis (Futurum) notes Claude targets discovery/analysis/documentation phases while full modernization encompasses broader architectural transformation. Assessment tool market broadens: Replay's visual reverse engineering approach reduces timelines from 18 months to weeks, addressing 70% legacy rewrite failure rate. Vendor ecosystem remains IBM-dominant but competitive pressure signals market acceleration and feature consolidation across platforms.

• 2026-Mar: Deployment evidence and practitioner analysis refine understanding of AI's actual role. Thoughtworks documents 66% reverse engineering acceleration (6 weeks to 2 weeks per 10K lines) on automotive 15M-line COBOL codebase, validating AI's discovery-phase impact. Practitioner consensus emerges: Thoughtworks (18+ months delivery) describes real workflow as layered (chunking, summarization, relationship inference with external scaffolding), not single-prompt automation; LLMs augment but cannot simultaneously process tens of millions of lines. Heirloom (PHEAA—largest mainframe-to-AWS migration) argues LLMs structurally unsuited for 100% correctness requirement; deterministic compilation produces byte-for-byte equivalence verifiable against tests. Independent adoption survey (Arcati 2026 mainframe practitioners): only 49% expect significant AI impact over 3-5 years; 8% expect major transformation. Top AI use cases remain narrow (anomaly detection 29%, security monitoring 26%), signaling augmentation not replacement posture. Market delineation: Indium analysis clarifies IBM's stock reaction—AI genuinely solves code analysis/documentation but cannot solve business logic extraction, behavioral equivalence validation, organizational change (estimated 80% of total cost). New vendors entering: CLPS Incorporation (20+ years banking domain) completed PoC with major Hong Kong bank, demonstrating commercial viability of AI-assisted transformation in regulated financial services. Bob 1.0 release (March 2026) consolidates RPG/COBOL assistants into multi-model platform (Anthropic Claude, Meta Llama, Mistral, IBM Granite) with unified pricing tiers ($20-200/month). Category remains leading-edge: vendor capabilities mature and competitive, deployment evidence accumulating across financial services and engineering, but adoption gap persists (13% production per Q1 2025 survey) due to organizational and validation barriers rather than technical capability. Risk-balanced evidence now confirms AI's discovery-phase impact while documenting persistent limitations in semantic validation and regulatory compliance—key factors governing adoption velocity.

• 2026-Apr: Vendor ecosystem expands with AWS market entry: AWS Transform service (GA April 2026) delivers agentic AI for code analysis and migration, with a documented case migrating 12 weeks of Control-M workflows to Apache Airflow in 2.5 weeks (3-5x delivery acceleration, 100% validation success) and extending agentic support to PL/I, directly challenging IBM's mainframe-focused dominance. Thoughtworks documented a $12B revenue financial services firm modernizing 4M lines of COBOL/HLASM using agentic AI in 4 weeks (vs 8 weeks planned) with 80% code comprehension accuracy under human-in-the-loop validation—concrete evidence of agentic approaches reaching production speed targets. European enterprises progressed from pilots to production GenAI in mainframe workflows, with ISG analyst report documenting governance frameworks and human oversight requirements for scaled deployment. BMC software shifts from generative AI assistance to agentic architecture: Knowledge Hub captures institutional knowledge from historical resolutions; zAdviser Enterprise generates narrative application analysis combining code analysis with operational telemetry. Critical counter-evidence emerged: Gartner predicted 70% of 2026 mainframe exit projects will fail, and a survey of 200 enterprise SRE/DevOps leaders found 43% of AI-generated code requires manual production debugging with developers spending 38% of weekly time on fixes; Thoughtworks Technology Radar v34 identified 'cognitive debt' as a systemic risk from AI-generated code, warning of semantic diffusion and the need for zero-trust controls. Practitioner evidence accumulates: AltexSoft documents AI agent discovery effectiveness (revealed 11 hidden dependencies vs. team's expectation of 5) balanced with critical context-window and persistence limitations; Sigma Software identifies critical flaw in naive AI test generation (locks in incorrect behaviors as untestable technical debt) and proposes behavior-first testing approach. Vendor market survey (IT Jungle) documents competitive ecosystem maturity: IBM Bob (multi-model framework), Profound Logic CoderFlow, Fresche Solutions, ARCAD—signals market acceleration despite persistent adoption barriers. Category remains at leading-edge: vendor capabilities competitive and deployment evidence accumulating, but Gartner's failure prediction and persistent validation overhead confirm that organizational barriers (semantic correctness, expertise) remain the binding constraint.

• 2026-May: Consulting industry structural shift and market validation accelerate: EPAM Systems announces multi-year partnership with Anthropic to certify 10,000+ architects on Claude with 1,300 already certified as of May 6, 2026, signaling enterprise-scale industry shift toward vendor-specialized practices (vs. agnostic deck consulting). Named deployment evidence strengthens: ZK Fiddle (13-year-old legacy app, 2-year stalled migration) unblocked using Claude Code + institutional knowledge base; Novacomp completed Java 8→17 modernization (10K LOC) in 50 minutes vs. 3 weeks, 60% technical debt reduction; FPT deployed AI on 300+ systems with 200M+ LOC transformed and 30% assessment effort reduction. AWS Transform reached scale deployment—4.5B lines of code processed, 1.6M hours saved, hundreds of thousands of VMs migrated over its first year—with agentic core banking modernization (COBOL/mainframe, AWS Transform and Kiro tools) and a documented case (Morgan Stanley 280K developer hours in COBOL; DevGen.AI reduced migration work ~50%) reinforcing both the scale of the problem and the realistic productivity envelope AI delivers on brownfield work. Market analysts assess boundaries: DORA 2026 study shows 35-40% productivity gains on greenfield code vs. 10% or less on legacy brownfield, establishing AI's realistic contribution to legacy work. IBM SVP critical analysis articulates key limitation: code translation ≠ modernization—real work spans data architecture, runtime behavior, transaction integrity, and regulatory compliance, with business logic extraction and behavioral validation remaining human-intensive. Category consolidated at leading-edge with mature competitive vendor ecosystem and strong real-world deployment evidence; constraints remain organizational (semantic validation expertise, regulatory compliance) and human (verification overhead), not technical.

• 2026-Jun: Deployment evidence and competitive ecosystem mature across verticals. Named production cases include Meliá Hotels completing a 20+ year COBOL mainframe migration to AWS (60% cost reduction, 75% time-to-market improvement), Bridgestone migrating 1.2M lines of z/OS COBOL/JCL to Java in 7 months via AWS Transform (90% efficiency gains), and an unnamed $12B credit union achieving zero-downtime core banking cutover with agentic automation eliminating 40+ hours/week of manual compliance mapping. AWS extended Transform to PL/I and launched CAST Highlight AI Acceleration Insights for portfolio-level agentic readiness assessment; Heirloom Computing gained ISG recognition as a leading multi-cloud mainframe modernization platform. Against this momentum, Veracode's June 2026 security study found 45% of LLM code transformation outputs introduce OWASP Top 10 vulnerabilities (Java at 70% failure rate), and practitioner analysis emphasized that syntactic translation differs fundamentally from semantic contract reconstruction—250B lines of COBOL can fail in production when invisible 40-year-old business rules are not reconstructed alongside the code. Meliá Hotels completed 20+ year mainframe-to-AWS migration (60% cost reduction, 75% time-to-market, 99.99% availability); Bridgestone migrated 1.2M lines z/OS COBOL/JCL to Java in 7 months using AWS Transform with 90% efficiency gains; unnamed $12B credit union achieved zero-downtime core banking cutover with agentic automation eliminating 40+ hours/week manual mapping. AWS extends Transform to PL/I (May 2026) and announces portfolio-level AI readiness assessment (CAST Highlight AI Acceleration Insights) for systematic candidate identification at scale. Heirloom Computing's GA platform gains ISG recognition as leading multi-cloud mainframe solution, broadening vendor competition beyond IBM. However, security quality risks surface: Veracode GenAI security study (June 2026) documents 45% of LLM code outputs introduce OWASP Top 10 vulnerabilities, with Java reaching 70% failure rate—material evidence that AI transformation risks introducing security defects absent from original systems. Practitioner analysis (Invisible Codebase, June 2026) distinguishes syntactic translation from semantic contract reconstruction, documenting that 250B lines COBOL can fail in production when AI tools miss invisible business rules embedded in 40+ year old code. Category remains at leading-edge: deployment evidence accelerating across industry, vendor ecosystem maturing competitively, but security and semantic validation barriers now explicit constraints alongside organizational factors.

• 2026-Jul: AWS Transform expanded with Assess and Reimagine workflows reaching GA, adding deterministic business rule extraction with full line-level source traceability and automated validation testing—enabling auditable enterprise-scale transformation without sacrificing provenance. Multiple Fortune 500 deployments via AWS Transform documented, reinforcing that agentic mainframe migration has moved past proof-of-concept into repeatable production delivery. Cross-sector ROI evidence consolidated: FullStack Labs aggregated McKinsey cases documenting $12M healthcare savings, 50%+ insurer efficiency gains, and a FinTech achieving 40% effort reduction on a 20K LOC migration—providing sector-specific benchmarks for business-case construction. Practitioner analysis identified context debt (missing institutional knowledge about why 40-year-old business rules exist) as the binding bottleneck, not code transformation: Meta's 4,100+ module case documented that raising context coverage from 5% to 100% reduced AI tool calls 40%, directly validating the knowledge-first modernization approach emerging across the practitioner community. Mid-July evidence sharpened both sides of the ledger: Thoughtworks documented a 10-sport data platform migration compressed from a 2-3 year baseline to 3-4 weeks via AI-assisted business logic extraction, and Kanerika's analysis of 161 enterprise engagements (2015-2026) found 50-60% effort reduction with 76% achieving 50%+ improvement on at least one metric; a domain-adapted COBOL-Coder LLM reached 73.95% compilation success versus GPT-4o's 41.8%, validating specialized tooling over general-purpose models for legacy language translation. Countervailing evidence hardened in parallel: a PRISMA-based systematic review of 34 studies (2,847 developers) found 58% report quality degradation and 23% higher defect density in AI-heavy legacy modules, with 54% of teams seeing no net throughput gain after six months; Gartner's forecast that 70% of 2026 mainframe exit projects will fail from AI overestimation was reinforced by Specira's analysis that transpilers faithfully translate syntax but not the undocumented business rules that only surface in production weeks later.

• 2026-Aug: Vendor convergence continued with Fujitsu-IBM Bob's deterministic-plus-agentic COBOL-to-Java partnership in APAC and AWS Transform-Accenture's mainframe modernization offering (30-40% timeline reduction), while IBM's Q2 earnings showed 50% of Z17 customers now investing in AI-assisted modernization with watsonx users growing MIPS capacity 3x faster than non-users. Claude Code's reverse-engineering of undocumented legacy systems at the National Museum of Computing and a new agentic test-synthesis method (Locksmith Loop, 91.9% branch coverage on COBOL-to-Java validation) demonstrated AI's discovery-phase strength, even as Gartner reiterated that 70%+ of 2026 mainframe exit projects will fail from AI overestimation and an OpenLegacy CTO analysis argued AI belongs in discovery, not full migration.