According to Gartner’s Hype Cycle, AI-assisted coding has reached the “Peak of Inflated Expectations” and it’s easy to see why given its potential.  However, adoption in some organizations is limited and productivity gains are variable. Part of the challenge is that productivity itself is surprisingly difficult to measure.

Our goal is to separate genuine value from hype and understand where AI-assisted coding can make a meaningful difference today.

The pace of change

One thing that’s hard to argue with is the pace of change. The length and complexity of tasks that AI can complete without human intervention continues to increase and, in some areas, that rate of improvement appears to be accelerating.

Industry veteran Andrej Karpathy described late 2025 as an “agentic inflection point” for coding. While earlier tools often required frequent correction and intervention, newer generations are producing larger, more coherent and more reliable chunks of code.

Importantly, the progress is not only coming from improvements in foundation models. Much of the value is being driven by specialized tooling built around those models.

Today’s development landscape includes several levels of AI assistance:

• AI-powered search for fast answers
• Conversational assistants that generate code from prompts
• IDE-integrated agents that understand an entire codebase and make targeted changes
• Agentic engineering tools that can perform complex development tasks with minimal supervision
• Emerging autonomous systems that proactively identify issues and suggest improvements

However, increased capability does not eliminate the need for human oversight. In many cases, it makes that oversight even more important.

Use cases delivering value today

While some claims about AI coding remain speculative, there are already several areas where organizations are seeing consistent benefits.

Data ingestion and transformation

One area where we’re already seeing practical value is in data engineering and migration work.

Teams working with large numbers of data processing pipelines can use AI tools to translate code between languages and frameworks, helping accelerate modernization projects. For example, converting legacy data processing logic to distributed computing environments like Spark can be significantly faster when AI generates an initial version of the translated code.

In our experience, this can be particularly useful when working through large numbers of scripts. The AI isn’t removing the need for human review, but it can help accelerate that first pass and provide a useful starting point.

Developers still need to validate business logic, confirm coding standards and account for platform-specific requirements. But reducing the effort required to perform the first pass across hundreds of scripts can create substantial efficiencies.

Code quality reviews

AI is also proving useful as a code reviewer.

Modern coding agents can analyze pull requests, summarize changes and identify potential issues that traditional automated testing tools might miss. Examples include spotting inconsistent naming conventions, identifying outdated comments, highlighting unused variables or detecting areas where functionality and documentation have drifted apart.

These capabilities give faster feedback to the developers to catch and fix issues in their own code. And they help reviewers focus their attention on the more complex and critical aspects of the application while reducing the chances of simple mistakes reaching production.

Boilerplate and repetitive coding

Perhaps the clearest area of value lies in routine development work.

Scaffolding new components, applying repetitive changes across large codebases and generating standardized tests are all tasks that follow predictable patterns. AI tools are particularly effective in these scenarios because they can recognize and replicate those patterns quickly and consistently.

When the work is repetitive and easy to validate, AI assistance can remove significant amounts of manual effort without introducing significant risk.

The jagged frontier

Despite the progress, AI-assisted coding still has limitations.

Researchers studying AI performance often describe a “jagged frontier”, a concept that captures the uneven nature of AI capability. Some tasks benefit enormously from AI assistance, while others see little improvement or even experience declines in quality. The difficulty is that these boundaries are not always obvious in advance.

Tasks that tend to sit inside the frontier include:

• Boilerplate code generation
• Repetitive development activities
• Code translation and migration
• Pattern recognition
• Test generation
• Documentation support

Outside the frontier tend to be tasks that require broader context, judgement or organizational understanding. These include:

• Architecture and design decisions
• Project planning and prioritization
• Complex system integration
• Resolving ambiguous requirements
• Understanding organizational context
• Strategic trade-off decisions

A coding agent can see the code in front of it. It can’t necessarily see the systems around it, understand how a particular company defines a customer, or know when an assumption needs to be challenged rather than accepted.

As Karpathy argues, the areas where humans continue to add the greatest value are aesthetics, judgement, testing and oversight. Developers may not need to write every line of code, but they’re still essential for guiding, reviewing and directing AI agents and their outputs.

The practical response

We’ve seen how AI-assisted coding is developing rapidly, that it can be extremely helpful, but that it still has limitations and requires expert oversight. These operating principles can be a helpful way to think practically about the use of AI-assisted coding:

Tailor the solution
AI systems naturally default towards common patterns. Left alone, they will produce solutions based on what they have seen before.

Companies tend to compete on small nuances: how they define a customer, how they build products, how they operate and the decisions they make. These nuances need to be specified and actively managed by project teams.

The goal is not to accept the default answer. The goal is to ensure the solution reflects the unique requirements of the organization.

Create a clear vision and direction
As development becomes faster, alignment becomes more important.

When developers get stuck, they can increasingly ask an agent for help rather than another member of the team. That’s useful, but it also means teams need to be more intentional about documenting and sharing decisions, assumptions and project direction.

That helps keep people aligned, but it also helps keep the agents aligned.

Be intentional
Different projects benefit from AI in different ways. Some may gain more value during development, others during testing, review or requirements analysis.

The key is being intentional about where you’re using it and how you’re validating the outputs.

Embrace what helps
Some developers are already achieving significantly better results than others. Understanding why, sharing successful approaches and standardizing effective practices can help entire teams benefit from the technology.

That includes sharing prompts, aligning on tools, establishing review processes and building consistent approaches to quality assurance. As the technology evolves, the opportunity isn’t simply to automate more work, but to make better use of both AI capabilities and human expertise.

Human expertise still matters

From our experience, the biggest gains come when teams combine AI assistance with strong human oversight. Across many real-world projects, we’re using AI to speed up aspects of development and deliver better solutions. But maintaining trust in those applications requires expert human oversight, particularly when solutions must be tailored, enterprise standards upheld and reliability assured.