AI Migration Landscape

This research surveys the AI-assisted code migration landscape as of February 2026, covering enterprise platforms, open-source tools, documented success stories, and the capability boundaries of current LLMs for migration tasks.

Enterprise Platforms

Tool	Vendor	Maturity	Key Metric
Amazon Q Transform	AWS	Most mature	1.1B lines analyzed, 810K hours saved
IBM watsonx Code Assistant for Z	IBM	Production	COBOL-to-Java with Granite models
GitHub Copilot Workspace	Microsoft	GA	Agentic multi-step migration
Google Gemini Code Assist + codmod	Google	GA	Large-scale automated codemod
Microsoft CAMF	Microsoft Research	Research	Code-Aware Model Framework for migrations

AWS Transform is the most production-proven platform (1.1 billion lines across Amazon’s internal Java migration). Microsoft and IBM are investing in agentic multi-step approaches.

Open-Source and Community Tools

Tool	Focus	Approach
OpenRewrite / Moderne	Java/JVM framework upgrades	Deterministic AST-based recipes
OpenHands	General-purpose AI coding agent	Agentic loop with sandbox execution
vFunction	Monolith decomposition	Automated service boundary detection

Documented Success Stories

Enzyme to React Testing Library

Metric	Value
Scale	3,500 files
Timeline	6 weeks (estimated 18 months manually)
Team	6 engineers
Compression ratio	~12x faster

Approach: State-machine pipeline (analyze, transform, validate, iterate). Retry loops with validation errors fed back to the LLM. Rich context: 40K-100K token prompts pulling up to 50 related files. Simple retry outperformed sophisticated prompt engineering.

Java 8 to Java 17

Metric	Value
Scale	30,000+ production applications
Savings	4,500 developer-years of effort
Tool	Amazon Q Transform
Cost savings	$260M estimated

The largest documented AI-assisted migration. Same-language version upgrade with well-defined transformation rules.

.NET Framework to .NET 8

Metric	Value
Timeline	Hours instead of months
Approach	Copilot-assisted with CAMF

Applied to Xbox platform components. Same-ecosystem upgrade with strong tooling support.

Organization	Migration	Result
Commonwealth Bank of Australia	COBOL to Java	40M lines scoped
US Air Force	COBOL to Java	Proof of concept
Various (IBM)	COBOL to Java	watsonx Code Assistant

No publicly documented case of an AI-assisted full ERP system migration exists. The closest examples are module-level extractions and framework upgrades within the same language.

AI Capabilities and Limitations

What LLMs Can Reliably Do

Capability	Confidence	Notes
Type mapping (Python to Go/Kotlin)	High	Mechanical, well-understood
Simple function translation	High	Pure functions, no side effects
Data model extraction from JSON	High	ERPNext DocType JSONs are ideal input
Test generation from specifications	Medium-High	Table-driven tests, edge cases
Business rule extraction	Medium	Depends on code clarity
Controller method translation	Medium	Requires rich context

What LLMs Cannot Reliably Do

Limitation	Impact	Notes
Cross-file dependency resolution	Critical	Hooks-driven execution paths are invisible
Framework behavior replication	Critical	ORM, permissions, workflow — emergent behavior
Dynamic typing inference	High	`frappe.get_value()` returns untyped data
Multi-tenancy architecture	High	Bench/site model is framework-level
Regional compliance logic	High	Scattered across hooks and overrides
Performance optimization	Medium	AI produces correct but unoptimized code

Academic Benchmarks

Metric	Range
Correct translations (general)	2.1% - 47.3%
Correct translations (specialized/fine-tuned)	90%+
Scalability cliff	~100 lines per translation unit
Asymmetric proficiency	Better translating TO Python than FROM Python

The “Jobol” Problem

Naive AI translation produces syntactically valid but unidiomatic code — dubbed “Jobol” (Java that reads like COBOL). For cross-language ERP migration, this manifests as target language code that mirrors Python patterns instead of using idiomatic constructs.

Code Intelligence Tools

Tool	Max Context	Key Feature
Sourcegraph Cody	1M tokens	100K+ repo support, enterprise-grade
Augment Code	200K tokens	500K files, 70% win rate over Copilot
Cursor	272K tokens	50K files, IDE-integrated
CodeCompass	Qdrant-based	Open-source, MCP-integrated
GitHub Copilot	8K-64K effective	Ubiquitous but limited context

For complex legacy codebases, standard IDE-level tools see individual files. The complexity requires seeing the full dependency graph — structural indexing, semantic search, and cross-file dependency resolution.

DDD Extraction from Legacy Code

Approach	Method
Quantized Code Llama + LoRA	Automated metamodel generation from code
Prompt-to-model pipelines (Qlerify)	Natural language to domain model
Mono2Micro + Context Mapper	Trace analysis to service boundaries

Eric Evans’s key insight: AI works for generic subdomains (CRUD, reporting, basic validation) but core domains need human involvement. For ERP systems, the accounting engine, tax calculation, and stock valuation are core domains requiring human judgment.

Parity Testing Approaches

Approach	Tool	Use Case
Shadow testing / traffic mirroring	Diffy, GoReplay, Speedscale	Run both systems in parallel, diff responses
Feature flags (4 modes)	LaunchDarkly, custom	NEW, OLD, MIRROR, PARITY modes for gradual cutover
Cross-database data diffing	Datafold	Compare data outputs between systems
Dual Run	Google Cloud	Run legacy and modern simultaneously
Consumer-driven contracts	Pact	Define expected behavior, verify both comply
AI-generated test suites	Custom (3-agent approach)	Extract logic, generate tests, generate passing code

AI Compression Ratios

Migration Type	Manual Timeline	AI-Assisted Timeline	Compression
Test framework migration	18 months	6 weeks (Airbnb)	~12x
Java version upgrade (large)	Years	Months (Amazon)	~5-10x
.NET framework upgrade	Months	Hours (Microsoft)	~100x
Legacy codebase migration	12-26 months	4-12 months	~2-3x
ERP system migration	12-24+ months	Unknown	Unknown

AI compression ratios are highest for homogeneous, repetitive transformations (Java 8 to 17, Enzyme to RTL) and lowest for heterogeneous, business-logic-heavy transformations (ERP migration).

Critical Caveats

Statistic	Source
57% of migrations take longer than anticipated	Industry surveys
83% fail or exceed budgets	Gartner
Vendor accuracy claims require verification on your codebase	Microsoft (early 2026)

Eight Key Patterns

1. Multi-Agent Pipelines

Analyze, translate, validate, iterate — the dominant architecture for AI-assisted migration.

2. Retry Over Prompting

Simple retry with error feedback outperforms sophisticated prompt engineering.

3. Rich Context Wins

Choosing the right related files matters more than prompt engineering.

4. Specialized Models

Specialized models outperform general-purpose LLMs for targeted language pairs.

5. Strangler Fig + AI

The lowest-risk combination for incremental legacy replacement.

6. Parity Is Non-Negotiable

Shadow testing and parity validation before production cutover.

7. Smaller Teams, Higher Skills

5-10x fewer people, but they need architectural judgment.

8. Verify Vendor Claims

Gap between demo results and production results is significant.

ERP Migration Market

Metric	Value
AI-in-ERP market (2023)	$4.5B
AI-in-ERP market (projected 2033)	$46.5B
SAP S/4HANA migration rate	37% of 35,000 customers
Typical ERP migration timeline	12-24+ months

No documented case of AI-assisted full ERP system migration exists publicly. PearlThoughts’ ERPNext experiment is among the earliest documented attempts at AI-assisted ERP cross-language migration.