You’ve got a legacy system problem and everyone’s telling you AI will solve it. When you start looking at vendors you’ll find dozens of tools all claiming they can help.
Here’s what matters: code comprehension tools and code generation tools solve different problems. Comprehension tools like CodeConcise from Thoughtworks and Claude Code analyse your existing codebase to generate documentation and extract business logic. Generation tools like GitHub Copilot write new code.
If you’re dealing with legacy modernisation, comprehension tools deliver more ROI. If you’re building new features on a modern stack, generation tools accelerate development. Most modernisation projects need both, but in the wrong order they waste your time and money.
This guide fits into the broader AI modernisation landscape, comparing the major tools, explaining when to use consulting firms versus platform services, and providing a framework for evaluating vendors without getting locked in.
What’s the Difference Between Code Comprehension and Code Generation Tools?
Code comprehension tools analyse existing codebases to generate documentation and extract business logic. Code generation tools write new code through AI-powered completion.
The distinction matters because legacy modernisation needs understanding first. Greenfield development needs speed.
Engineers spend more time reading code than writing it. Code comprehension tools tackle this directly. They build knowledge graphs using databases like Neo4j, parse code into Abstract Syntax Trees, and apply RAG for queries. The output is functional specs from undocumented code and business rules you didn’t know existed in your legacy systems.
CodeConcise uses abstract syntax trees in graph databases with edges showing control flow. The ingestion pipeline extracts structure without using LLMs. AI handles summarisation afterwards.
Generation tools work differently. They focus on snippet-level prediction, IDE integration for real-time suggestions, and boilerplate for CRUD operations. GitHub Copilot excels at this but struggles with multi-file reasoning.
So here’s the simple test: are you reverse engineering COBOL before migration? Use comprehension. Are you writing unit tests for refactored microservices? Use generation.
The ROI follows the budget. Legacy modernisation eats up 60-70% of effort understanding existing systems. Greenfield development puts 90% into writing new code.
When Does Code Comprehension Have More ROI Than Code Generation?
Code comprehension delivers higher ROI when you’re reverse engineering undocumented legacy systems, extracting business rules, or planning tech stack migrations. Generation tools are better for greenfield development and test creation.
Organisations spend 60-70% of modernisation budgets on understanding existing systems. That’s where comprehension creates value.
For undocumented COBOL applications, comprehension tools do the heavy lifting. Thoughtworks research shows a potential two-thirds reduction in reverse engineering time—from 6 weeks to 2 weeks per 10,000-line module. For large programmes that works out to a potential saving of 240 FTE years.
Tech stack migrations follow the same pattern. Moving from Angular to React requires deep understanding before you can rewrite anything. Cognizant research demonstrates 30-50% reduction in discovery time.
Generation tools flip the equation for greenfield scenarios. New microservices needing REST boilerplate? Test coverage expansion? New features in modern codebases? That’s generation territory. GitHub Copilot studies show 20-35% productivity gains.
The decision framework is straightforward. If you lack functional specifications, start with comprehension. If you know what to build but need speed, start with generation. If you’re modernising legacy, plan for comprehension taking 60-70% of your effort.
Without comprehension tools, manual reverse engineering takes 6-8 weeks. With them, 2-3 weeks. That makes the business case pretty clear.
How Do CodeConcise, Claude Code, and Cursor Compare for Legacy Modernisation?
CodeConcise is Thoughtworks’ proprietary tool for reverse engineering mainframe systems, available only through consulting. Claude Code offers 200k token context windows for multi-file reasoning, accessible via API. Cursor bridges comprehension and generation with IDE integration. GitHub Copilot focuses on snippet-level generation with limited modernisation utility.
CodeConcise uses code-as-data where parsers extract structure into Abstract Syntax Trees stored in a graph database. The comprehension pipeline traverses the graph using depth-first search. AI handles summarisation. For technical details on how these knowledge graph architectures work, see the deep dive on AST parsing and Neo4j implementation.
CodeConcise excels at mainframe expertise and business rule extraction. It’s the best for large-scale mainframe modernisation and undocumented COBOL. The catch? It’s an internal Thoughtworks tool. You can’t buy it as SaaS. You need to engage Thoughtworks for a consulting programme at premium rates.
A Thoughtworks experiment tested CodeConcise with Claude Code for adding language support. Normally takes two to four weeks. Claude Code identified all changes for Python support in half a day—a 97% time saving. But it failed completely on JavaScript, producing code referencing non-existent packages. So results are inconsistent.
Claude Code runs on a terminal-based interface for easy integration. You need an API key with pay-as-you-use pricing. It’s best for mid-market organisations wanting tool control and API-first approaches. But you’ll need custom integration work.
Cursor reached $1 billion in revenue in November 2025. They cherry-pick the best models, wrap them in the best interface for AI coding. They automated 80% of their own support tickets. Monthly subscription, per-seat pricing.
Cursor delivers strong IDE integration with hybrid comprehension-generation. Best for development teams wanting a unified tool. Weaknesses include context management complexity and less focus on reverse engineering.
GitHub Copilot brings maturity and Microsoft integration with the lowest learning curve. But limited multi-file reasoning makes it poor for reverse engineering. Best for greenfield development and entry-level AI adoption.
Choose CodeConcise for mainframe modernisation with consulting budget. Choose Claude Code for mid-market with internal AI expertise. Choose Cursor for development-led organisations needing both comprehension and generation. Choose GitHub Copilot for greenfield focus and Microsoft shops.
Multi-tool strategies are common. Use CodeConcise or Claude Code for comprehension plus Cursor or Copilot for generation. Avoid multiple comprehension tools though—you’ll create duplicate knowledge graphs and confuse yourself.
What Role Do Consulting Firms Play in AI-Driven Modernisation?
Consulting firms provide advisory services, methodology frameworks, proprietary tools, and implementation expertise that platform tools alone can’t deliver.
Thoughtworks differentiates through Martin Fowler thought leadership, evolutionary architecture, and CodeConcise integration. They run advisory through implementation for full-service modernisation. Works best when architecture quality matters and you have consulting budget. Boutique approach with 10-50 consultants per engagement.
Cognizant brings Skygrade platform and global delivery scale. Programme management for multi-year transformations with large engineering teams. Suits enterprise-scale organisations with complex portfolios. Their research shows 30-50% OpEx reduction.
Cognizant research found 85% of executives concerned their tech estate will block AI integration. Three-quarters plan to complete modernisation within two years. But 63% cite complexity as an obstacle, 50% cite talent, and 93% have retired only 25% of tech debt.
98% plan to use systems integrators. That validates the consulting model.
Engage consulting when your strategy is unclear, you have expertise gaps, or you need change management. Use platform tools when strategy is defined, you have internal capability, or you’re doing tactical migrations.
The hybrid model works well—consulting for strategy in weeks 1-6 with platform tools for execution in weeks 7-12.
Value beyond tools includes change management, methodology transfer, and risk reduction through battle-tested approaches.
Should You Build Custom AI Tools or Buy Vendor Solutions?
Build custom tools when you have unique requirements, need strategic control, have internal AI expertise, and face multi-year modernisation. Buy vendor solutions when speed matters, you want proven approaches, have limited resources, or use standard tech stacks.
33% of organisations cite vendor lock-in concerns as driving build decisions. But the calculus is shifting. 2020-2023 builds made sense due to immature tools. 2024-2026 buys are increasingly attractive as tools mature and costs drop.
Build for proprietary domain-specific languages, niche tech stacks, or industry compliance needs. Build when AI capability is a competitive differentiator. Build when you have ML engineering teams and infrastructure already in place.
Example: Financial services firm with proprietary trading languages, regulatory requirements, and internal AI research.
Buy when you need immediate capability to meet board deadlines. Buy for proven approaches reducing risk. Buy when you have small teams or standard languages like Java and .NET.
Example: Mid-market SaaS company modernising a Java monolith with no AI team and an 18-month deadline.
The hybrid approach works too—buy the foundation then extend with internal knowledge. Wrap vendor tools in internal APIs for future switching.
Cost analysis matters. Build costs include ML salaries at £150k-250k per engineer plus infrastructure. Buy costs include subscriptions at £50-150/month or consulting at £200k-500k. Build breaks even at 50-100+ developers or 20-30 person-years of work.
Run a 3-month pilot before committing. Negotiate data residency, export capabilities, and pricing caps.
How Do You Mitigate Vendor Lock-In Risk in AI Modernisation?
Use API abstraction layers that decouple tools from core systems. Use open standards like Kubernetes and Docker. Deploy across multiple clouds.
API abstraction puts an internal layer between AI tools and application code. Switch from Claude Code to CodeConcise by changing the API implementation without touching application code. Costs 10-15% development overhead but saves months of migration work later.
Open standards use Kubernetes for orchestration. Avoid cloud-specific services like ECS or Cloud Run. Use EKS, AKS, or GKE for portability.
Multi-cloud avoids single-provider dependency. Don’t commit to Azure Mainframe Modernisation if it locks you into Azure for a decade. Run primary workloads on AWS, disaster recovery on Azure, burst capacity on GCP. Multi-cloud adds 15-20% complexity but provides negotiating leverage.
Run annual benchmarking testing 2-3 alternative tools. Every 18-24 months, run a 2-week PoC with emerging tools. Share results with your current vendor during renewals.
Negotiate data export rights in standard formats. Get portability clauses requiring vendor migration assistance. Cap price increases at CPI plus 5%.
Prefer API-accessible tools like Claude Code over consulting-only tools like CodeConcise. Choose tools with open integration patterns like Cursor with MCP servers.
High lock-in risk: single-vendor platforms like Azure Mainframe Modernisation, proprietary formats, consulting-delivered tools. Low lock-in risk: open-source tools, multi-cloud deployments, containerised approaches.
Lock-in is acceptable when you’re committed to the vendor’s infrastructure anyway, when they’re the only option handling your tech stack, or when exit costs are manageable.
What Platform Services Are Available for Mainframe Modernisation?
Azure Mainframe Modernisation provides automated COBOL-to-cloud migration, best for Microsoft-committed organisations. AWS Migration Hub offers multi-cloud support and migration tracking. Google Anthos focuses on multi-cloud Kubernetes for hybrid deployments.
Azure delivers automated COBOL-to-C# conversion, JCL to Azure Batch, mainframe data to Azure SQL. Strengths include deep Microsoft integration and automated conversion. Weaknesses are Azure lock-in and limited PL/I support. Best for Azure enterprise agreements and .NET shops.
AWS Transform is an agentic AI service that transforms monolithic COBOL into microservices. Automatically analyses source code for technical documentation. Extracts business logic and provides data lineage analysis.
AWS’s AI assistant Kiro handles specification-driven development. Generates microservice specifications, databases, and infrastructure as code. Enables architects to design with formal specifications before implementation.
Strangler fig pattern enables progressive modernisation. Replace monolithic applications gradually while keeping the original running. Both systems operate simultaneously during coexistence.
Google Anthos delivers multi-cloud Kubernetes management and hybrid cloud deployments. Provides a Kubernetes-based abstraction layer running on-premises and across clouds. Strengths include cloud portability and avoiding single-cloud lock-in. Weaknesses require Kubernetes expertise and add complexity.
Use platform services for standard patterns and tactical migrations. Use consulting firms for complex architecture and portfolio-wide transformation. The hybrid model uses consulting for strategy then platform services for execution.
Platform services complement comprehension tools. Use Claude Code or CodeConcise to understand your mainframe, then platform services to migrate it.
How Do You Evaluate AI Modernisation Tools for Your Tech Stack?
Use five weighted criteria. Language support weighs 30%. Integration weighs 25%. Pricing weighs 20%. Vendor track record weighs 15%. Support quality weighs 10%.
Run proof-of-concept pilots on representative code before committing. Involve architecture leadership and development teams.
Language support at 30% asks whether the tool handles your legacy languages—COBOL, PL/I, assembly. Modern coverage for Java, .NET, Python. Framework support for Spring and React. Database dialects for DB2 and Oracle.
Integration at 25% covers CI/CD pipelines, IDE compatibility, Git workflows, and API availability. Collaboration features for knowledge sharing.
Pricing at 20% evaluates transparency, alignment with usage patterns, and total cost of ownership. Does pricing scale linearly or have economies of scale? Monthly, annual, or multi-year contracts?
Vendor track record at 15% looks at case studies in your industry, reference customers of similar size, and technology maturity. Company stability and roadmap transparency.
Support quality at 10% examines service level agreements, support channels, training, community resources, and professional services.
PoC framework runs 2-4 weeks. Select 2-3 representative code samples. Define success criteria before starting. Include 2-3 developers, 1 architect, 1 product owner. Deliver a PoC report with recommendation.
Common mistakes include feature checklist bias, pilot bias with cherry-picked examples, sunk cost fallacy, hype cycle trap, and analysis paralysis. Set a 3-month decision deadline.
Document decisions for institutional memory. Maintain vendor comparison matrices. Capture lessons learnt for next evaluation.
Review primary tools annually. Benchmark 1-2 alternatives on the same criteria. Use results for contract renegotiation.
Once you’ve selected your tools, you need an execution plan. The 90-day AI modernisation implementation playbook covers tool selection within a broader modernisation roadmap, including vendor evaluation criteria, team structure decisions, and phased rollout strategies. Understanding the code comprehension vs generation thesis provides strategic context for these technical decisions.
FAQ Section
Is CodeConcise Available as a Commercial SaaS Product?
CodeConcise is a proprietary Thoughtworks tool, available only through consulting engagements. It’s not sold as SaaS. To use it, you need to engage Thoughtworks for a modernisation programme. Alternatives include Claude Code via Anthropic API or building custom tools with Neo4j and open-source parsers.
Can I Use Multiple AI Tools Together?
Yes, it’s common. Many organisations use comprehension tools like CodeConcise or Claude Code for reverse engineering alongside generation tools like Cursor or Copilot for new development. Avoid multiple comprehension tools—you’ll duplicate knowledge graphs and make extra work for yourself. But combining comprehension with generation works well. Use API abstraction for future vendor switching.
How Do I Avoid Vendor Lock-In?
Use API abstraction layers. Use open standards like Kubernetes and Docker. Deploy across AWS, Azure, and GCP. Negotiate data export rights and portability clauses. Run annual benchmarks. Design for portability from the start.
Should I Build Custom AI Tools or Buy Vendor Solutions?
Buy if you need speed, have standard tech stacks, lack AI expertise, or want proven approaches. Build if you have unique requirements, need strategic control, have ML expertise, or operate at scale where vendor costs exceed build costs. The hybrid approach buys foundation then extends with custom logic.
Which Consulting Firms Specialise in AI-Driven Legacy Modernisation?
Thoughtworks brings evolutionary architecture with CodeConcise, ideal for transformation and architecture redesign. Cognizant offers large-scale integration with Skygrade platform, suited for enterprise programmes. Kyndryl provides mainframe expertise via IBM heritage, best for mainframe-to-cloud infrastructure. Choose based on scope, scale, and technology focus.
What’s the Difference Between Rehosting, Replatforming, and Refactoring?
Rehosting moves applications to cloud without code changes. Fastest at 3-6 months with lowest transformation. Replatforming makes minimal changes to leverage cloud capabilities. Balanced at 6-12 months. Refactoring rebuilds with modern architectures. Slowest at 9-18 months but highest transformation. Choose based on risk tolerance, timeline, and architecture goals.
How Long Does an AI Modernisation Tool Proof-of-Concept Take?
2-4 weeks is typical. That gives you time to test representative code samples, hands-on evaluation by developers, and assessment across your criteria. Shorter lacks depth. Longer delays decisions. Define success criteria before starting.
What Programming Languages Do AI Modernisation Tools Support?
CodeConcise excels at mainframe languages—COBOL, PL/I, assembly, Natural. Claude Code supports most languages—Java, .NET, Python, JavaScript, C++. Cursor focuses on popular modern languages with IDE integration. Copilot covers mainstream languages but has limited legacy support. Language support is your highest-weighted criterion at 30%.
How Much Do AI Modernisation Tools Cost?
Generation tools use per-seat subscriptions—Copilot around £10/month, Cursor around £20/month. Comprehension tools use usage-based API pricing like Claude Code, or consulting fees like CodeConcise at £200k-500k for programmes. Platform services use consumption-based cloud costs. Build vs buy breaks even at 50-100+ developers or 20-30 person-years. Get total cost of ownership including licensing, infrastructure, training, and support.
What Are MCP Servers and Why Does Cursor Use Them?
MCP servers are Cursor’s context engineering approach, providing structured context beyond code snippets. They enable integration with documentation, databases, and APIs. AGENTS.md files define project-specific context rules. This differentiates Cursor from snippet-level tools by enabling multi-file reasoning and domain knowledge injection.
How Do Knowledge Graphs Improve Code Comprehension?
Knowledge graphs represent code as a semantic network of entities and relationships rather than text. This enables queries like “find all functions modifying customer credit limits” that are impossible with text search. CodeConcise uses Neo4j to store relationships, enabling reverse engineering and business rule extraction. It provides structured context for AI models beyond token sequences.
