AI & Code Maintenance:

This white paper provides a realistic assessment of AI in 2025 for development teams: its actual capabilities, its limitations, and how to leverage it effectively for analyzing, documenting, and maintaining complex codebases.

Written for developers, architects, and Ops professionals, this document focuses on practical use cases, technical constraints, and best practices for integrating AI into your existing code tools and workflows.

1. Executive Summary

2. Context: Why AI is becoming essential for development teams

3. The state of AI in 2025

3.1. ANI, AGI, ASI: What we actually have today

The literature identifies three theoretical stages of AI evolution:

For practical purposes, we will remain in the ANI domain for years to come, with powerful but specialized models that depend heavily on their training data and context.

3.2. From conversation to agents: conversational, generative, reasoning, and agentic AI

Since 2022, ANI models have evolved through several stages:

4. AI capabilities and limitations for development tasks

4.1. Raw performance: IQ tests and benchmarks

Recent benchmarks show that some models achieve human-level performance on standardized IQ tests. On paper, AI appears 'comparable' to humans in certain areas...

...but these scores primarily measure raw capability on well-defined tasks (formal problem-solving, logic, text comprehension). They do not reflect robustness, adaptability, or the ability to manage a complex, evolving system like an enterprise codebase.

4.2. Adaptability tests: Where AI still falls short

To better assess AI limitations, organizations like the ARCPrize Foundation have developed tests that measure adaptability to novel tasks with implicit rules and shifting context.

Key findings:

• Most models struggle to exceed a 60% success rate on puzzles that humans solve with ease;
• On tests combining adaptability and efficiency, most models score below 10%, with top performers barely reaching 20%.
• These scores are improving far more slowly than the raw performance measured by traditional benchmarks.

In other words, AI is highly capable when given well-defined tasks in formats it has seen before, but remains fragile when required to:

• Generalize from unfamiliar contexts;
• Balance conflicting constraints;
• Take a step back and reason about a complex system, such as a multi-million-line enterprise application.

4.3. Reliability, hallucinations, and the case for co-intelligence

These limitations directly affect developers' daily work:

• AI cannot gauge its own reliability. It may produce incorrect answers with complete confidence, sometimes backed by seemingly coherent reasoning.
• Models are highly context-sensitive: without precise instructions and relevant data, they tend to give vague responses or simply tell users what they want to hear.

For instance, the Stack Overflow Developer Survey 2025 reports that 84% of developers already use AI tools, yet 46% do not trust the outputs, and nearly half say these tools struggle with complex tasks.

This underscores the value of co-intelligence:

• Either AI produces a result that humans verify;
• Or humans produce a result that AI helps verify and refine.

In software maintenance, the second approach is often more effective: AI serves as a review and analysis assistant, while the developer retains ownership of the final decision.

5. Security, compliance, and governance: Prerequisites for AI adoption

The primary barriers to enterprise AI adoption today relate to security and compliance: leaked secrets, personal data exposure, industry regulations, liability concerns, and more.

Key 2024–2025 trends

• Companies are establishing AI governance policies:
  • 59% have already implemented AI usage and control policies (source: Pacific);
  • 98% plan to increase their AI governance budget (source: OneTrust).

• Europe has enacted the AI Act to regulate AI use and strengthen trust in critical AI systems.
• All major paid model providers (OpenAI, Claude, Gemini, etc.) now offer contractual options that guarantee data confidentiality:
  • No use of customer data for model retraining;
  • Isolated environments;
  • Limited and controlled logging.

For DevOps teams, this means AI can be integrated into development workflows without exposing code or business data to additional risk—provided you:

• Select providers that meet regulatory requirements (GDPR, AI Act, etc.);
• Clearly define what data is shared with AI;
• Incorporate these considerations into security policies and architecture reviews.

6. AI adoption in 2025: Individuals and businesses

6.1. Individual adoption

Individual AI use has grown dramatically:

• An estimated 1 billion regular users worldwide, including 850 million using OpenAI products.
• In the United States, 62% of the population uses AI at least weekly (source: Pew Research).
• Between 19% and 27% of users report using it for work (source: Gallup).

Many developers now rely on AI to:

• Write unit tests;
• Get quick algorithm explanations;
• Scaffold functions or infrastructure scripts;
• Rewrite tickets or documentation.

6.2. Enterprise adoption

On the enterprise side, the picture is paradoxical:

• 78% of companies already use AI in at least one function,
  
• including 36% in IT (source: McKinsey).
  
• Yet only 1% say they are 'fully' leveraging it (source: Pew Research).
• Only 12% of U.S. employees have received formal AI training (source: Pew Research).
• Still, 17% of companies report an EBIT improvement of at least 5% attributable to AI (source: McKinsey).

7. Applying AI to code analysis and maintenance

7.1. The limitations of 'raw' AI for developers

Asking a ChatGPT-style model to 'Explain this function' or 'Tell me the impact of this change' based on a simple copy-paste is inadequate.

Limitations become apparent immediately:

• Incomplete context: AI sees a snippet but has no visibility into dependencies, data models, or configuration constraints.
• Partial view: impossible to reason holistically from isolated fragments.
• Silent errors: AI may miss edge cases, tables, or triggers without realizing it—and without flagging what's missing.

For complex, mission-critical systems, the risk is operational: refactoring guided solely by AI may break functionality, cause outages, or degrade performance.

7.2. Why a tooled framework matters

For AI to be productive in application maintenance, it needs:

Complete, structured context

• Inventory of objects (procedures, triggers, tables, UI components, services, etc.);
• Dependency and call-path analysis (cross-references, call graphs...);
• Metadata on the application, data volumes, and performance.

Structured, controlled input

• Clear objective: explain, document, estimate impact, identify risk;
• Expected output format (list, table, summary with details, etc.);
• Explicit guardrails (do not fabricate, flag missing information, verify).

Integration with existing tools

• Invoked from the code analysis tool or IDE;
• With logging, traceability, and peer review support.

To meet these requirements, Visual Expert leverages static code analysis to supply structured application data to AI, enabling contextualized, verifiable responses.

7.3. Example: Integrating reasoning AI into a tool like Visual Expert

For each feature:

• Visual Expert automatically assembles the context (code, dependencies, metadata);
• Formats a structured prompt for the AI model;
• Returns a response that the developer can review, supplement, or revise.

8. Toward the first software maintenance agents

Today's features foreshadow the first software maintenance agents:

• Capable of suggesting fixes or preparing comprehensive refactoring;
• Validating certain assumptions (e.g., whether a change meets maintainability and security standards);
• Eventually, automating multiple steps in the build pipeline, especially continuous code checks in a CI/CD environment.

Near-term safeguards

In the near term, AI limitations dictate several safeguards:

• No autonomous execution in production;
• Mandatory review by a developer or architect;
• Validation through manual or automated tests (unit, integration, performance).

9. Practical recommendations for developers, architects, and ops

10. Conclusion: Co-intelligence as a competitive advantage

AI will not replace developers, architects, or Ops professionals. However, teams that integrate it thoughtfully into their tools and processes will gain a decisive edge: faster maintenance cycles, deeper insight into existing applications, reduced risk in certain areas, and greater agility when facing major evolutions.

The question today is no longer whether AI is 'good enough,' but how to use it effectively:

• Within a solid governance framework;
• Embedded in existing tools and processes;
• Supported by solutions that deliver rich context;
• With humans remaining at the center of technical decisions.

Solutions that combine advanced static analysis with reasoning AI—like Visual Expert—point the way forward: they transform AI into a credible, productive software maintenance assistant that meets the security and compliance demands of modern IT environments.