Legacy code analysis & migration
164 evidence items
AI that analyses legacy systems to document behaviour, identify dependencies, and assist migration to modern platforms. Includes COBOL-to-Java migration and mainframe modernisation; distinct from code refactoring which improves existing code within its current platform.
Overview
Legacy code analysis and migration uses AI to read ageing systems, recover the behaviour buried in them, map dependencies and help move them onto modern platforms. Anyone carrying a mainframe estate should care, and the practice is a leading-edge practice, steady: the tooling and analyst attention are real, but a clear path is not. AI reliably speeds up discovery and documentation; producing code that provably behaves like the original is another matter. Accuracy falls away at enterprise scale, and failures often pass visible tests unnoticed. The successes so far pair the tooling with heavy bespoke validation and human review — an engineering feat, not yet something a competent team can simply adopt.
Current Landscape
IBM has made its agentic assistant Bob the centre of its modernisation offer. Analytics Insight reports that IBM Bob reached general availability on 28 April 2026, covering planning, coding, testing, deployment and modernisation. More than 80,000 IBM employees now use it, self-reporting a 45% average productivity gain. At Blue Pearl, a Java upgrade that once took 30 days finished in three. IBM has also shipped Bob into IBM i systems via VS Code. Fujitsu pairs its conversion tooling with Bob for COBOL-to-Java work in APAC. IBM's Q2 2026 earnings put mainframe AI adoption at 50% of Z17 customers.
AWS Transform is the most widely documented rival at scale. Its throughput is reported at 4.5B LOC, with named deployments at ADP, CSL and Signaturit. Bridgestone moved 1.2M lines of z/OS COBOL/JCL to Java in 7 months with it. Thoughtworks combined its own tooling with AWS Transform for a mainframe exit with 80% timeline acceleration. AWS frames its Reimagine path around business-rule extraction with full traceability. Microsoft was named a Leader in the 2026 Gartner Magic Quadrant for AI-Augmented Code Modernization Tools.
Services firms and specialist vendors report COBOL-to-Java delivery at volume. Software Mind describes Cotality modernising 73 legacy COBOL services to Java/Spring Boot with 90% automation. CLPS Incorporation reports completing an AI-driven COBOL-to-Java migration at scale. GFT says its Wynxx platform has modernised 100M+ LOC across 113 enterprise customers. Zengines describes a Fortune 100 financial institution modernising 60M+ LOC with AI-assisted data lineage. Kanerika draws its delivery outcomes from 161 real engagements between 2015 and 2026. These figures are self-reported by the vendors involved.
Software-intelligence mapping is drawing consulting alliances. Deloitte India and Cast announced an alliance on 23 September 2026 to map application estates across Asia-Pacific and Japan. The offering documents systems and identifies the code-level blockers that decide whether an application is rehosted, refactored or rebuilt. Cast's Marc Zablit says clients have reported 2x increases in agent accuracy once AI is given deterministic context from Cast, without disclosing a baseline or method. AWS has added AI Acceleration Insights to CAST Highlight for portfolio-scale agentic readiness assessment.
Newer agentic platforms lead with codebase-wide context. Blitzy launched a free Sandbox that ingests up to 1 million lines of code and returns up to 25,000 lines of tested code as pull requests. It cites COBOL-to-Java migrations and a customer that saved 4,300 hours of engineering time on a 500,000-line Java output. Unblocked reports a Meta case in which raising context coverage from 5% to 100% reduced AI tool calls by 40%. Both claims are vendor-reported, with no independent benchmarks.
Japanese enterprises are applying generative AI to comprehension and maintenance before conversion. A vendor roundup reports that Meiji Yasuda Life cut COBOL rework man-hours by roughly 25% using generative AI, with humans reviewing quality. The same roundup says Toyota Systems opened a Legacy Code Lab with IBM Japan support in October 2025, so that junior engineers could understand COBOL and PL/I. It also describes AI analysis of more than 50,000 Progress2 programs previously judged impossible to analyse. Claude Code has been applied to digital archaeology at the National Museum of Computing.
Independent benchmarks show accuracy falling sharply with scale. RepoMod-Bench records a performance cliff from 91.3% to 15.3% as codebases grow. AlphaCorp cites the May 2026 "Articulate but Wrong" study, which found silent behaviour drift in 39.7% of 1,980 code modernisation attempts across 11 production LLMs. In that study, the same model endorsed 31.7% of those errors as correct. DEPBENCH, a peer-reviewed benchmark of agents on legacy dependency upgrades, documents hidden breakage that passes visible tests. NTT DATA identifies three structural barriers to AI COBOL modernisation using April 2026 benchmark data.
Research is narrowing parts of the gap. AlphaCorp reports that IBM Research's SANER 2026 summary-augmentation method improved COBOL-to-Java translation outcomes on 36% of eligible benchmark samples. On the hardest enterprise cases the improvement reached up to 50%. Domain-adapted COBOL-Coder models significantly outperform general models on COBOL translation. The Locksmith Loop paper proposes an agentic method for deterministic validation of legacy code migration, so that equivalence evidence is generated rather than assembled by hand.
Practitioners argue that verification, not translation, decides whether output is usable. Kodebaze's Claus Villumsen writes that most AI refactoring tools capture no pre-transformation behaviour, log no reasoning and produce no evidence of equivalence. In his view, that makes them unsuitable for regulated environments. He estimates that a behavioural baseline takes two to four weeks for a medium-complexity system. Teams that skip it spend three to six months diagnosing regressions. Kyndryl's Surasa Mukherjee argues that AI discovery should provide evidence for investigation, not an unquestioned answer.
Analysts expect many programmes to disappoint. Gartner forecasts that over 70% of mainframe exit projects will fail due to AI capability overestimation. Kearney's 2026 Transformation Study found only 29% of transformations achieve intended value. WatchZ reports that AI compresses syntax work by 60% but that 95% of pilots fail to reach production, stalling on architecture. CMC Global cites MIT NANDA data showing 67% success with specialised vendors against 33% for in-house efforts.
Code quality and security remain material risks in transformed output. Veracode reports that 45% of LLM outputs introduce OWASP Top 10 vulnerabilities. Software Improvement Group finds that agentic code carries 2x the security violations of human-written code. It also warns that agentic work can run €10-15M token bills. Faros links high AI adoption to bugs rising 54% and incidents rising 57.9%. A PREreview systematic review reports that 58% of AI coding studies show quality degradation on legacy code.
Agentic tooling with production access has already caused damage. Data Platform Advisory documents production incidents in which agentic AI executed destructive operations during data migrations. It argues that scoped credentials and approval gates outside the agent's reasoning are structural requirements. AgentPatterns proposes a three-phase archaeology workflow that keeps AI assistance bounded. Precision Federal catalogues AI failure modes in legacy financial systems where the code is the only documentation.
Many enterprises are extending legacy estates rather than exiting them. Ensono's 2026 survey of 500 US and UK IT leaders found that 78% regard legacy systems as more critical than they did two years ago. More than half are optimising and extending legacy systems in place rather than replacing them: 57% in the UK and 48% in the US. The most-cited barriers were difficulty integrating AI into existing workflows (33%) and infrastructure limitations (28%). Zan Digital likewise frames integration, not model limitations, as the structural blocker.
What blocks broader adoption is behavioural validation and lost institutional knowledge, not translation speed. Specira argues that AI transpilers translate syntax rather than intent, leaving undocumented business rules invisible until production. OpenLegacy's CTO places AI's durable value in the discovery phase rather than in end-to-end automation. Organisations reporting results keep human review, parallel validation and approval gates in front of cutover. That keeps the slowest, most expert-dependent work in human hands.
Tier History
Evidence (164)
— Japanese named cases: Meiji Yasuda cut COBOL rework ~25% with GenAI plus human review, Toyota Systems opened a Legacy Code Lab, and AI analysed 50,000+ Progress2 programs. Vendor-sourced.
— Deloitte India and Cast form an APJ alliance for AI-led legacy estate mapping. Cast claims a 2x gain in agent accuracy on brownfield code from deterministic context, with no baseline given.
— IBM Bob reached GA on 28 April 2026 with modernisation in scope. It has 80,000+ internal users (45% self-reported gain), and a Blue Pearl Java upgrade fell from 30 days to 3. All figures are vendor-sourced.
— Critical practitioner view: AI legacy-transformation tools give no behavioural baseline or evidence of equivalence. Skipping a 2–4 week baseline costs 3–6 months of regression diagnosis.
— Curates two-sided 2026 research: IBM SANER COBOL-to-Java gains on 36% of samples versus 'Articulate but Wrong' silent behaviour drift in 39.7% of 1,980 modernisation attempts.
159 more · latest 2026-09-16 →
— Ensono survey of 500 US/UK IT leaders: most extend legacy in place rather than replace, and 33% cite difficulty integrating AI into workflows. A barrier signal against wholesale migration.
— Kyndryl architect sets out AI's role in mainframe discovery and dependency mapping. Warns AI output is evidence for investigation, not an answer, since source code alone does not describe production behaviour.
— Blitzy Sandbox ingests up to 1M LOC for autonomous legacy modernisation, citing COBOL-to-Java migrations and 4,300 hours saved on a 500,000-line Java output. Vendor-claimed.
— Multi-org deployment scale: AWS Transform processed 4.5B LOC over 12 months (ADP supporting 1.1M customers, CSL, Signaturit) with documented efficiency gains (97% speed improvement, 6-8 weeks vs 6-8 months, 60% cost reduction).
— Fortune 100 custody accounting migrated 60M+ LOC COBOL, 80K+ modules from mainframe to cloud SaaS; reconciliation breaks reduced 40%→0.4% using AI data lineage; discovery phase compressed from 6 analysts/9 months to minutes, enabling business-led redesign.
— Critical documentation of two 2026 production incidents (PocketOS database deletion, Supabase table drops) where agents given broad credentials executed destructive operations—demonstrates system prompts are probabilistic inputs, not enforced boundaries, requiring governance controls.
— Methodology for AI-assisted legacy comprehension via forensic audit (adversarial prompting), verifiable baseline (running system), incremental understanding lift against oracle—prevents confabulation and enables traceable legacy modernization assistance.
— Named manufacturer migrated mainframe JCL batch jobs to Python/AWS Batch and Db2 to PostgreSQL using behavior-first testing; achieved 80% timeline acceleration (18 months → 5 months), hundreds of Java classes and 1000+ SQL queries modernized, warranty operations uninterrupted.
— Practitioner guidance (ex-IBM CTO) maps high-fit AI use cases (code understanding, dependency discovery, test generation) versus low-fit failures (behavioral validation, architecture decisions, tribal knowledge recovery)—clarifies realistic AI role not full autonomy.
— Independent engineering firm documents AI failure on semantic understanding (loop invariants, aliasing, arithmetic precision) with concrete risk for financial systems; behavioral recovery identified as 31% of effort, the binding constraint not code writing.
— Synthesis of peer-reviewed studies documenting 45% of AI samples introduce OWASP Top 10 vulnerabilities, Java 72% failure rate, 19.7% hallucinate dependencies—critical risk signal for production legacy modernization deployments.
— Named customer (Bupa, 7M Asia-Pacific users) modernized My Bupa mobile app from Xamarin to native Swift/Kotlin with AI-assisted reverse/forward engineering; achieved 60% timeline reduction, app rating 3.7→4.7 stars, crash rate -24pt Android/-8pt iOS.
— Monash/Microsoft benchmark found 51.2% agent success rate on dependency upgrades; critically, 61.8% of non-passes satisfied visible tests but failed hidden integration tests—signals systematic overconfidence from incomplete test feedback.
— CTO analysis emphasizing AI excels at discovery phase but fails on full migration due to context window and test coverage gaps; cites Gartner prediction of 70%+ project failure from AI capability overestimation.
— Survey analysis showing 35% of large enterprises cite data readiness and integration as top barrier to scaling agentic AI; only 17% of IT leaders confident infrastructure supports mission-critical agents, establishing integration as binding constraint.
— Peer-reviewed benchmark shows dramatic performance collapse as codebase size grows—91.3% pass rate on <10K LOC vs 15.3% on >50K LOC—establishing structural limit of probabilistic LLM code generation independent of model quality.
— Legacy modernization platform modernized 100M+ lines of code across 113 corporate customers with 1,900 active users in 12 countries; reports ~50% reduction in software development time with expanding agentic AI integration.
— Real estate intelligence company migrated 500K+ lines of COBOL across 3,500 files to Java with 90% automation, 55% effort reduction vs standard approaches, validated functional parity through shadowing, 4-6 hours per-service processing vs ~36 hours manual.
— Named enterprise (CLPS, Nasdaq-listed) completed full production migration: 985k LOC converted, 1,124 database tables migrated, delivered in 16 months with 20 developers vs 80-person 5-year estimate, with AI accuracy improving from 80-90% to 98% through iterative optimization.
— April 2026 Factory AI Legacy-Bench found state-of-the-art AI models at 70%+ on general benchmarks achieve only ~15% success on legacy code migration; identifies silent-failure and conversion≠correctness gaps as binding constraints.
— Vendor selection guide grounded in third-party research: MIT NANDA finds specialized vendors succeed 67% of time vs 33% in-house; Veracode documents 45% of AI-generated code introduces OWASP Top 10 vulnerabilities despite 95% syntax pass rates.
— Survey of 102 senior transformation executives: only 29% consistently achieve intended value; 80%+ report <50% ROI on AI initiatives; resistance to change tops implementation barriers, establishing organizational factors as primary constraint.
— Japanese systems integrator adopting IBM ALSEA platform for AI-driven enterprise development; scaled team from ~70 to 200 engineers by end FY2027, with projects already contributing revenue in production phase after PoC transition.
— Battery distribution company case: 15+ legacy applications modernized with 53% timeline reduction (8.5 months to 4 months); durable outcome is institutional knowledge preservation—AI as archaeological tool uncovering business logic.
— Critical investment analysis: AI accelerates analysis phase but cannot compress hard constraints—testing, validation, regulatory approval, institutional caution dominate real migration timelines; adoption barriers remain organizational.
— Inaugural Gartner MQ for AI-Augmented Code Modernization Tools recognizes Microsoft as Leader, signaling analyst-validated mainstream category recognition for AI-assisted legacy system transformation.
— Peer-reviewed benchmark testing seven LLMs: 41% semantic-equivalence misclassification at zero context, 29% with minimal context—demonstrates core limitation across model families independent of vendor.
— Next Pathway deterministic semantic translation platform: 160+ enterprise modernizations completed, 1B+ lines transformed, 80% timeline compression with verified functional parity—alternative vendor approach to LLM-only transformation.
— DORA 2026 research cites Stanford finding: 35-40% productivity gains on greenfield but 'often 10% or less on complex legacy code'—quantifies realistic constraint for legacy-focused modernization ROI.
— IBM Bob Premium Package for i (GA June 24, 2026) enables native IBM i agentic workflows supporting RPG/COBOL modernization directly in VS Code—distinct platform from Z, expanding agentic capability to mid-range systems.
— IBM watsonx Code Assistant Enterprise Java extension (VS Code GA) with Java runtime modernization, version upgrades, code explanation, and test generation—direct evidence of IDE-integrated agentic Java modernization tooling.
— Scott Logic 10K-line Java-to-Rust migration case study: test traceability, architectural constraints, and mutation testing enable teams to trust generated code—operational techniques reducing risk of autonomous transformation.
— Gartner's critical risk assessment predicting 70%+ of 2026 mainframe exit projects will fail; by 2030, 75% of modernization vendors forecast to exit/pivot; GenAI overestimation named primary cause alongside complexity and undocumented business rules.
— Register columnist reports Claude Code enabling rapid reverse engineering of undocumented legacy systems (Transputer, Econet NFS); ROMs fully commented in a day; validates AI's transformative role in legacy system discovery and analysis at scale.
— Novel agentic test-synthesis method for COBOL-to-Java migration validation; empirical tests on 3 case studies reached 91.90% branch coverage with deterministic parity checks, addressing production correctness validation gap.
— Expert CTO analysis positioning AI for discovery phase (where it excels) rather than full migration (where context windows and test coverage gaps cause 70%+ project failures); frames adoption barriers as organizational, not technical.
— Strategic vendor convergence: Fujitsu PROGRESSION (deterministic conversion) paired with IBM Bob (agentic AI) for COBOL-to-Java migration in APAC, signaling integrated platform-agnostic approach over single-vendor solutions.
— AWS Transform and Accenture partnership combining agentic reverse engineering, business rule extraction, and forward engineering for mainframe modernization; claims 30-40% timeline reduction.
— IBM Q2 earnings reveal 50% of Z17 customer base investing in AI (Spire Accelerator) and watsonx Code Assistant users growing MIPS capacity 3x faster than non-users, validating measurable business impact of AI-assisted mainframe modernization.
— Speed masks incompleteness: transpilers faithfully reproduce visible logic but discard invisible business reasoning. Case study: module passed all tests, broke week 3 on untested behavior. Demonstrates why recovery of documented requirements precedes conversion.
— Named deployment: AI-assisted business logic extraction accelerated 10-sport migration from 2-3 years to 3-4 weeks; per-sport onboarding compressed from 10-15 weeks to under 1 day using reusable framework and SME-directed validation.
— PRISMA 2020 review of 34 empirical studies (2,847 developers): 58% report quality degradation; legacy code defect density +23% in AI-heavy modules; 54% of teams saw no net throughput gain after 6 months. Mitigation via quality gates prevents 67% of debt insertion.
— Market context (220B COBOL LOC, $25B market) with critical barrier: Gartner predicts 70% of 2026 mainframe exit projects will fail due to GenAI overestimation. Success pattern: disciplined enterprises achieve 40-50% faster delivery and 70% cost reduction with architect-governed AI acceleration.
— Real consulting outcomes: 50-60% effort reduction, 90-day delivery on 24+ month baselines, 75% licensing cost reduction; 76% of 161 engagements improved by 50%+ on at least one metric. Foundation-first modernization produces documented results.
— Domain-specialization proven: COBOL-Coder achieves 73.95% compilation success vs GPT-4o's 41.8% on code generation; 34.93 Pass-1 on Java-to-COBOL translation vs near-zero for general-purpose models. Peer-reviewed validation of specialized approach necessity.
— FinTech 40% effort cut on 20K LOC; insurer 50%+ efficiency gains; healthcare $12M savings + 85% defect reduction. Demonstrates AI-augmented modernization delivering measurable outcomes across sectors.
— Meta 4,100+ modules case: context debt (missing institutional knowledge) is core AI bottleneck. 100% context coverage enables 40% tool-call reduction. Frames modernization constraint as knowledge, not code.
— New GA workflows: deterministic business rule extraction with line-level source traceability, semantic enrichment, automated validation testing. Enables enterprise-scale transformation with auditability.
— ADP (1-in-6 US payrolls), Itaú Unibanco (70M customers, largest Latin American bank), Western Union: 90%+ reverse-engineering speedup, 80%+ forward-engineering acceleration, zero-downtime business logic externalization.
— Tier-2 analyst forecast documenting gap between marketed AI capability and real-world performance; 70% project failure rate; 75% of mainframe vendors expected to pivot/exit by 2030. Critical negative signal.
— Market sizing: 71% of large enterprises run mainframe workloads; 60% have <5 staff with deep expertise. Skills bottleneck (67% CIOs cite as #1 risk) drives sustained multi-year investment cycle.
— Comparative analysis of four major vendor approaches: all converge on business-domain-first strangler-fig pattern with phased validation, not bulk AI conversion. Signals market-wide strategic consensus.
— Practitioner-validated failure analysis paired with 6-phase hybrid framework (comprehension before code). Teams following approach report 30-50% faster modernization, 40% lower QA costs, 12-24 month completion.
— Critical assessment: AI shows speedup on isolated tasks but 19% slowdown on mature 1M+ LOC codebases (legacy context). Security risk: commits 3-4x faster, security findings jumped 10x. Signals debt swap risk.
— 22,000-developer telemetry: output velocity jumps but quality deteriorates sharply. Root cause: architectural governance gap. AI code lacks guardrails against system-level violations. Deployment context: production.
— 30,000 systems analysis (400B+ LOC): autonomous agents produce unmaintainable code; security violations 2x human code; productivity gains vanish beyond 100K LOC. Documents structural governance gap.
— Named case (NOSI): AI excels at code understanding/documentation but lacks business context expertise. 94% metric validates AI's discovery-phase impact; reinforces human-AI hybrid model necessity.
— Critical quality risk in AI-assisted modernization: 45% of LLM code generation tasks introduce OWASP vulnerabilities; Java shows 70% security failure rate. Negative signal documenting material risks in generated code not present in legacy systems.
— Sophisticated practitioner analysis distinguishing syntactic translation from operational contract reconstruction; surveys three institutional approaches (IBM watsonx, AWS Transform, open-source Reversa) for 250B COBOL lines with context on 13% IBM stock drop market validation.
— GA mainframe modernization platform recognized by ISG as leading multi-platform solution; clients span diverse industries (PHEAA education, insurance, payment processing); delivers 65% operational savings and months-vs-years timelines with high-confidence refactoring.
— Named global hotel chain (380+ hotels, $3B revenue) completed 2-year migration of 20+ year legacy COBOL mainframe to AWS: 60% compute cost savings, 75% time-to-market improvement, 99.99% availability, 234ms→160ms response time.
— Named automotive manufacturer replatformed 1.2M lines of legacy z/OS COBOL/JCL to Java in 7 months using agentic AI with human supervision; achieved 90% efficiency gains and balanced automation with human oversight to preserve business logic.
— AWS and CAST announce portfolio-level AI assessment capability: evaluates applications for agentic AI readiness (API maturity, security, observability, resilience patterns) across legacy portfolios, enabling target prioritization for modernization.
— Named credit union eliminated 40+ hours/week of manual compliance mapping, achieved 100% pre-cutover data validation across five layers, executed zero-downtime cutover without rewriting legacy COBOL via agentic data transformation framework.
— AWS details AI-assisted core banking modernization for COBOL/mainframe systems, with specific tools (AWS Transform, Kiro) and methodology for business logic extraction and code generation.
— Direct case study: Morgan Stanley's 280k developer hours in COBOL, DevGen.AI reduced migration work ~50%. Includes modernization sequence and regulatory drivers with practical tooling examples.
— Flagship AWS product page with comprehensive 1-year deployment metrics: 4.5B lines of code processed, 1.6M hours saved, hundreds of thousands of VMs migrated, demonstrating enterprise-scale adoption.
— AWS Transform product documentation with embedded CSL case study showing 10x faster application discovery and 10.5 weeks saved in wave planning for multi-data-center exit.
— Longitudinal survey of 270+ organizations across 5 quarterly waves (Mar 2025–Mar 2026) tracking legacy refactoring adoption maturity, revealing slow progression despite perceived high value.
— Reports METR controlled research showing AI tools are 19% slower on mature codebases vs. greenfield, and experienced developers already possess repository context that AI lacks. Negative signal for legacy code work.
— Practitioner at scale: agentic AI system extracting business knowledge from 2,000 COBOL batch jobs; solves economically infeasible manual reverse-engineering; demonstrates practical deployment pattern.
— Named global services firm (FPT) with 300+ systems and 200M+ LOC transformed; 30% effort reduction in assessment phase for major steel manufacturer case, documenting real deployment economics.
— Market sizing: $22.1B (2026) → $50.7B (2033) at 12.6% CAGR, with application modernization 34% of market, BFSI dominant, broad vertical adoption signaling category maturation.
— AWS and Anthropic official documentation demonstrating integrated workflow for legacy mainframe modernization using reverse engineering and agentic code generation.
— Major consulting firm (EPAM) publicly commits to certifying 10,000+ architects on Claude with 1,300 already certified and 5,000 by Q3 2026; signals enterprise-scale consulting shift toward vendor-specialized practices.
— IBM SVP critical assessment: code translation ≠ modernization. Real work is system-level engineering (data architecture, runtime, transaction integrity). Includes three named customers with metrics.
— Market signal: IBM stock dropped 13.2% (worst day since 2000) after Anthropic announced AI-driven COBOL analysis, indicating investor perception that AI automates legacy discovery cost bottleneck.
— AWS-IBM hybrid cloud collaboration with specific named deployment (Toyota Motor NA): 40M+ LOC COBOL to Java in 50% less time with AI, demonstrating deployment velocity and ecosystem acceleration.
— Named case study (ZK Fiddle, 13-year-old app) with specific problem (lost institutional memory), solution (Claude Code + CaseFoundry knowledge base), outcome (unblocked 2-year stalled migration).
— Named organization (Novacomp) deployed Amazon Q Developer: upgraded 10,000 LOC Java 8 to Java 17 in 50 minutes vs. 3 weeks; 60% average technical debt reduction; improved security posture.
— IBM's agentic AI product for legacy mainframe modernization; combines watsonx Code Assistant with multi-step orchestration for COBOL/Z-system analysis and transformation.
— DORA 2026 research: 35-40% productivity gains on greenfield code but only 10% or less on legacy brownfield code, establishing realistic ROI expectations for AI-assisted legacy work.
— Practitioner analysis of AI code verification gap with audit methodology. Cites 43% production failure rate and field-tested verification checklist applicable to legacy migration workflows.
— AWS case study: enterprise software company migrated Control-M workflows to Apache Airflow in 2.5 weeks vs 12-week estimate; achieved 3-5x delivery acceleration, 10-20x effort reduction, 100% validation success across all workflows.
— ISG analyst report documents European enterprises operationalizing GenAI in legacy code analysis workflows; shows progression from pilots to production use with governance frameworks and human-in-the-loop oversight.
— Named financial services firm ($12B revenue) modernized 4M lines of COBOL/HLASM using agentic AI; delivered in 4 weeks vs 8 weeks planned, achieved 80% code comprehension accuracy with human-in-the-loop validation.
— Gartner analyst prediction that 70% of 2026 mainframe exit projects will fail—critical negative signal documenting persistent AI limitations and barriers to broader organizational adoption.
— Thoughtworks Technology Radar v34 identifies 'cognitive debt' as critical risk from AI-generated code; warns of semantic diffusion and need for human oversight and zero-trust controls—high-credibility independent assessment of AI limitations.
— Survey of 200 enterprise SRE/DevOps leaders: 43% of AI-generated code requires manual debugging in production post-deployment; developers spend 38% of weekly time fixing AI code, signaling substantial validation overhead.
— Sigma Software identifies critical flaw in naive AI test generation: AI-generated tests lock in incorrect behaviors (existing technical debt), producing untestable technical debt; proposes behavior-first approach reconstructing actual system behavior from specs, logs, and metrics before generating protective tests.
— AWS extends Transform to PL/I modernization (5% of legacy applications in finance/insurance/gov), replacing 24-30 month manual conversion with reverse-engineering analysis and Kiro agentic code generation, reducing project timelines and $15M+ conversion costs.
— AltexSoft analysis documents AI agent effectiveness in dependency discovery and architectural analysis (discovered 11 hidden cross-module dependencies vs. team's expectation of 5) while identifying critical limitations: context window constraints and persistence gaps in session-scoped AI tools.
— BMC shifts from GenAI assistance to agentic AI architecture: Knowledge Hub captures institutional knowledge from historical resolutions and documentation; zAdviser Enterprise generates AI-analyzed application narratives (behavior, risk, complexity) from telemetry and code analysis.
— Mimacom VP documents agentic engineering economics: 50-75% of legacy code translation effort automated, with 50% timeline and 30% cost reduction demonstrated across named deployments (Fujitsu, Thomson Reuters: 1.5M LOC/month modernization).
— AWS official mainframe modernization hub featuring AWS Transform service with agentic AI capabilities for analyzing and documenting millions of lines of legacy code, identifying dependencies, and producing migration artifacts with estimated 70% cost reduction.
— IT Jungle market survey documents competitive vendor ecosystem for AI-powered legacy modernization: IBM Bob (multi-model framework), Profound Logic CoderFlow (5-10x productivity claims), Fresche Solutions (constrained hallucination), ARCAD (source analysis), Remain Ai Chat, Programmers.io—signals maturity and market acceleration.
— Sahaj Software documents systematic four-phase approach to legacy code analysis: unified codebase structure, business-domain context organization, AI-driven feature development, and incremental strategic modernization—addressing prerequisite conditions for AI-powered legacy system evolution.
— Journalist coverage of AWS Transform deployment outcomes: migration projects reduced from 18 months to 7-8 months via deterministic application analysis combined with agentic code reimagining; AI agents address testing automation bottleneck (historically 50% of migration timelines).
— IBM stock dropped 13.2% (~$31B market cap) on Anthropic announcement; critical boundaries mapped: AI solves code analysis but cannot solve business logic extraction, behavioral equivalence validation, org change (80% of cost).
— Automotive OEM achieved 66% reverse engineering time reduction (6 weeks to 2 weeks per 10K lines) on 15M-line COBOL mainframe using AI accelerators, projected 60,000 person-day savings across full codebase.
— Independent mainframe practitioner survey: 49% expect minor AI impact over 3-5 years; only 8% expect major change. Top AI use cases narrow: anomaly detection (29%), security (26%)—signals AI viewed as augmentation, not transformation.
— Vendor with 20+ years COBOL banking domain expertise completed successful PoC for major Hong Kong bank, demonstrating commercial viability of AI-assisted COBOL-to-Java transformation with core logic integrity maintained.
— Vendor with largest mainframe-to-AWS migration record (PHEAA student loan system) argues LLMs structurally unsuited for 100% correctness requirement; deterministic compilation produces byte-for-byte equivalence verifiable against test cases.
— Independent engineer balances deployment signals (BMW 75% test reduction, Fiserv 17-month compression) with critical risks: modernizing without mainframe runtime context (EBCDIC, numeric precision, middleware) risks silent data corruption.
— Major consulting firm (18+ months delivery experience) clarifies actual AI workflow: layered chunking, local summarization, relationship inference requiring external scaffolding; LLMs augment structural analysis but cannot handle tens of millions of lines simultaneously.
— Futurum analyst (Mitch Ashley) compares Anthropic Claude and IBM watsonx for COBOL modernization; notes Claude targets discovery, analysis, documentation phases while modernization is broader architectural challenge.
— Anthropic announces Claude AI capability for COBOL legacy code modernization; IBM stock drops 13.2% in response; market scale: 220B COBOL lines in banking/government/healthcare; workforce urgency: 43% US banking COBOL-reliant, 10% of COBOL developers retiring annually.
— Technical presentation on iterative COBOL-to-Java refactoring using AI agents; engineer-agent performs refactoring while critic-agent validates quality, enabling traceable, testable, reliable legacy modernization workflows.
— Leading food wholesaler deployed AI-optimized COBOL batch processing modernization to Spring Batch; achieved 20-30% acceleration in development timeline using AI-driven migration tools.
— Assessment tools benchmark shows Replay's video-based reverse engineering reduces modernization timelines from 18 months to weeks; 70% of legacy rewrites fail or exceed timelines, creating market demand for faster assessment.
— IBM Q4 earnings: CEO Arvind Krishna highlights watsonx Code Assistant for Z enabling COBOL-to-Java refactoring; IBM Z revenue up 48% for year with z17 launch driving mainframe modernization adoption.
— UK public sector spending shows legacy systems consume nearly half of IT budget (£2.3bn of £4.7bn); high-risk legacy systems increased 26% from 2023 to 2024; modernization efforts failing to address debt accumulation.
— Russian insurance company (AlfaStrakhovanie) deployed deterministic ANTLR-based COBOL-to-Java translator after finding AI tools unreliable; documents structured approach to mainframe modernization with semantic preservation.
— Market data compilation shows 75%+ of organizations now use AI for modernization; legacy modernization market valued $24.98B in 2025, projected $56.87B by 2030; 95% of ATM swipes use COBOL code.
— IBM releases watsonx v2.8.0 with agentic chat interface, enabling AI agents to orchestrate multi-step legacy analysis and transformation tasks with architectural reasoning across mainframe codebases.
— IBM announces Project Bob, unifying watsonx COBOL (System Z) and RPG (IBM i) code assistants into single platform, extending legacy modernization tooling across IBM's broader platform ecosystem.
— Thoughtworks case study: agriscience company assessed 15-year-old, 30-component, 2.2M-LOC legacy system in one month using CAST Imaging, delivering technical risk quantification and foundation for modernization planning.
— IBM releases watsonx v2.7 with AI-powered business rule discovery and natural language COBOL code generation in VS Code, expanding language support and addressing SME knowledge loss challenges.
— Mainframe consulting firm evaluates watsonx v2.6+ based on hands-on experience, noting COBOL generation as 'compelling' for legacy transformation and Assembler explanation as 'time-saver,' citing Omdia analyst recognition.
— IBM research paper presents automated evaluation system for COBOL-to-Java translation in watsonx Code Assistant for Z, combining analytic checkers with LLM-as-judge techniques for scalable quality assessment.
— Stack Overflow survey (49,000+ developers) reveals 80% AI adoption but only 29% trust in accuracy; 66% spending extra time fixing 'almost-right' AI code—signal of validation overhead in migration workflows.
— OpenLegacy whitepaper outlines phased coexistence strategy for safe legacy modernization; cites Forrester research showing 353% ROI and describes platform capabilities for rapid discovery and service transformation.
— IBM announces watsonx Code Assistant for i public preview for RPG legacy code explanation, trained on 10M+ lines of RPG including deprecated versions (RPG II, III, OPM).
— IBM researchers present automated functional validation framework for COBOL-to-Java translation in watsonx, using symbolic execution and semantic equivalence testing to address LLM reliability in production.
— IBM releases watsonx Code Assistant for Z 2.6 on June 27, 2025, adding AI agents for autonomous COBOL code generation, expanded language support (Dutch, Korean), and assembler code explanation.
— IBM and CAST announce partnership leveraging CAST's industry-leading dependency analysis technology to enhance application discovery phase, enabling faster and more accurate modernization assessment.
— Technical analysis examines generative models' capability to replace manual reverse engineering of legacy COBOL, Fortran, and Assembly systems, addressing time and consistency challenges in code understanding.
— Forrester commissioned TEI study validates OpenLegacy's Digital Driven Integration platform achieves 353% ROI for enterprise customers modernizing monolithic legacy systems.
— OpenLegacy Hub SaaS product listed on AWS Marketplace, enabling cloud-native legacy modernization with automated dependency analysis and structured integration planning.
— Survey of 200+ mainframe IT leaders shows only 13% deployed AI in production for legacy systems, with 32% planning and 22% testing—adoption remains nascent despite vendor advances.
— European automotive manufacturer accelerated legacy modernization by using generative AI to cut reverse engineering time by two-thirds on 15M-line codebase, reducing SME overhead.
— Comparative study of ChatGPT, Copilot, and other AI tools shows all produce non-replicable, semantically incorrect COBOL-to-Java translations, highlighting persistent technical limitations.
— IBM watsonx Code Assistant for Z v2.x GA documentation detailing COBOL-to-Java transformation, semantic equivalence validation, and deployment options on z/OS.
— Leading global insurance company documented 4x acceleration in modernization planning using CAST Imaging to analyze 110M lines of COBOL in four weeks without disrupting operations.
— Software AG ends support for legacy Unix platforms (HP-UX, AIX, Solaris) by December 31, 2024, creating market pressure for rehosting legacy applications to Linux x86.
— NTT DATA technical analysis documents critical challenges in generative AI-assisted COBOL migration (hallucinations, rule-based limitations) and proposes Java redesign guidelines for accuracy.
— AWS Prescriptive Guidance officially lists CAST Highlight as partner tool for legacy application discovery and analysis across 50+ languages including COBOL, supporting cloud migration planning.
— IBM GA releases watsonx Code Assistant with COBOL-to-Java translation and mainframe application modernization, supporting SaaS and on-premises deployment for enterprise modernization.
— IBM IBV report confirms 71% of Fortune 500 companies rely on mainframes, establishing sustained market demand for legacy system modernization across banking, insurance, telecom, retail, and airlines.
— CAST partnership with Google Cloud demonstrates modernization of .NET legacy applications for hundreds of clients worldwide, showing multi-language portfolio migration at scale.
— IBM expands watsonx on-premises availability to include COBOL Code Explanation feature, enabling organizations to deploy AI-assisted legacy code understanding without cloud dependency.
— US property and casualty insurance company achieved 80% reduction in developer time to understand legacy applications using IBM Watsonx Code Assistant for Z.
— Survey of 500 IT leaders shows 86% deploying or planning GenAI for mainframe modernization with 114-225% one-year ROI; 43% report lacking skills to leverage AI capabilities.
— Italian national insurance agency (INAIL) deployed CAST Imaging for analyzing and migrating 700 legacy applications to cloud, expanding from 100-app pilot to full portfolio.
— IBM Research tutorial (ISEC 2024) surveys legacy modernization challenges, architectural paradigms, and generative AI opportunities including automated testing and code generation.
— CAST Imaging winter update adds advanced filtering, impact analysis for UI/database changes, and new DB2-to-MySQL migration Advisor, expanding tool capabilities for legacy analysis.
— Industry opinion highlights persistent adoption barriers: shortage of programmers skilled in legacy languages (COBOL, C/C++), system complexity, and 2038 problem risks.
— Federal CTO discusses how federal agencies can use AI to explain legacy code, accelerate greenfield development, and expedite vulnerability remediation in memory-unsafe languages.
— ICSE 2024 workshop paper presents method for translating COBOL to Java using LLMs with high-resource refinement, addressing data scarcity and improving translation correctness.
— Gartner Distinguished VP notes IBM watsonx Code Assistant 'has no case studies at this time to validate its claims,' highlighting validation gap despite product GA.
— Practitioners emphasize management and cultural barriers ('if they haven't done it by now, it's a management problem') and technical challenges ('challenge is getting that last bit of correctness').
— CAST releases AI-driven Advisors feature in CAST Imaging to provide personalized recommendations for mainframe refactoring and cloud migration.
— IBM launches watsonx Code Assistant for Z as GA product for mainframe application modernization, signaling vendor investment in COBOL-to-Java translation.
— Developer community discussion highlights technical challenges: LLM mistakes, legacy system dependencies, and need for semantic understanding beyond 1:1 translation.
— SoftRoad deployed AI-assisted COBOL-to-Java migration for Mazda, NEC, Fujitsu and others, completing 950+ migrations with automated testing and cost reduction.
— CAST announces automated 'cloud blocker' detection in CAST Imaging to identify and visualize inhibitors preventing legacy app migration to cloud.
— Case studies from Google, Manulife, and Accenture demonstrate software intelligence tools cutting modernization time by 50%+ across multiple industries.
— Gartner predicts GenAI tools will reduce legacy modernization costs by 70% by 2027; Slalom experiments show 73% faster language migration.
— Gartner research via Kyndryl identifies critical adoption barriers: lack of clear IT strategy positioning and retirement of mainframe expertise.