AI-Assisted Coding: From Individual Experimentation to Measurable Software Delivery Improvement

AI-assisted coding has quickly become part of modern software development.

Developers use AI to generate code, explain unfamiliar codebases, create tests, troubleshoot errors, refactor existing solutions, and increasingly complete larger units of development work. The potential is significant, but simply giving developers access to AI coding tools does not guarantee better software delivery.

One developer may use AI extensively while another barely touches it. Teams may develop their own prompting approaches. AI-generated code may enter the codebase without consistent review expectations. And organizations may track licenses and adoption while having little evidence that AI is actually improving cycle time, quality, rework, or delivery performance.

The gap between individual AI productivity and system-level software delivery performance matters.

DORA's 2025 State of AI-Assisted Software Development research, based on nearly 5,000 technology professionals, found that AI's impact is best understood as an amplifier of the surrounding software delivery system. Strong practices amplify the benefits of AI, while weaknesses in the system become magnified.

GitHub's enterprise research with Accenture also illustrates the opportunity. Developers using GitHub Copilot showed increases in pull requests, pull request merge rates, and successful builds. GitHub also emphasizes measuring AI's impact using telemetry and metrics that reflect an organization's own goals and development workflows.

Organizations don't simply need developers using AI.

They need an intentional capability for AI-Assisted Coding.

 

What Is AI-Assisted Coding?

AI-assisted coding is the use of AI to better enable the creation and modification of source code while maintaining or improving the quality and performance of software delivery.

It goes well beyond asking an AI assistant to autocomplete the next few lines of code.

AI can support developers with activities such as:

• Generating initial code
• Modifying or refactoring existing code
• Creating tests
• Understanding unfamiliar code
• Troubleshooting errors
• Generating boilerplate
• Exploring implementation approaches
• Reviewing code
• Working through larger development tasks

The objective is not simply to generate more code. The real opportunity emerges when these activities become part of a deliberate AI workflow rather than isolated uses of an AI coding assistant.

A strong AI-assisted coding capability helps teams produce software more efficiently while maintaining or improving quality, reliability, maintainability, and delivery performance.

That distinction becomes increasingly important as AI coding tools become more powerful. GitHub's guidance recommends treating AI-generated code as code that still requires human review, testing, secure coding practices, and standard code-review controls.

The question therefore evolves from:

"Are our developers using AI?"

to:

"Have we developed repeatable AI-assisted coding practices that improve how software is delivered?"

 

Figure Out Where You Are

Before deciding how to improve AI-assisted coding, identify where your teams are today.

LAI's AI-Assisted Coding Maturity Model progresses through five stages. Each represents an increase in how consistently AI is applied, integrated into the AI workflow, measured, and improved.

The objective is not to move every team to Optimizing as quickly as possible.

The objective is to understand where AI-assisted coding is today, what meaningful progress looks like, and what teams should improve next.

 

The AI-Assisted Coding Maturity Model

Lean Agile Intelligence's AI-Assisted Coding Maturity Model provides a five-stage progression:

Starting → Emerging → Enabling → Operationalizing → Optimizing

The maturity model progresses from little or no AI support in coding, through independent developer experimentation and shared practices, to AI-assisted coding becoming an integrated and continuously improving part of the development workflow.

Each stage helps teams understand what their current state looks like, what evidence demonstrates progress, and what practical actions move the capability forward.

 

Starting: Coding Remains Primarily Manual

At the Starting stage, AI is either unavailable to developers or rarely used to create and modify source code.

Developers continue performing most coding tasks without AI assistance. Teams may have heard about AI coding tools, but there has been little meaningful experimentation or discussion about where they could improve development work.

This is not necessarily a technology problem. In some organizations, tools are technically available but developers have not been given the clarity, encouragement, enablement, or opportunity needed to incorporate them into their work.

What This Looks Like

Common signals include:

• No AI coding tools installed
• Pull requests contain no AI-generated code
• Boilerplate code is repeatedly recreated
• AI usage is absent from code-review discussions

The important signal is not simply whether the organization owns AI licenses. It is whether AI is meaningfully participating in the coding workflow.

How to Progress to Emerging

Start with structured experimentation, not a mandate that every developer immediately use AI.

Identify several common coding activities where the potential benefit is easy to understand and relatively easy to evaluate.

For example:

• Creating repetitive boilerplate
• Generating unit tests
• Explaining unfamiliar code
• Refactoring a small function
• Creating documentation
• Troubleshooting an error

Give a small group of developers access to an approved AI coding tool and ask them to deliberately test it against these tasks.

The goal at this stage is learning:

Where does AI actually help our developers?

Practical Example: Run a Two-Week AI Coding Experiment

Select 5–10 developers and identify three common development tasks.

For each task, capture:

Task → Normal Approach → AI-Assisted Approach → Time/effort → Quality → Developer Learning

For example:

At the end of two weeks, hold a 60-minute retrospective:

Where did AI help? Where did it create additional work? What would we use again? What should we avoid?

Do not attempt to create enterprise-wide standards yet.

The purpose is to create enough real experience to move from assumption to evidence

 

At the Emerging stage, AI-assisted coding is happening primarily through individual developer initiative.

Multiple developers are experimenting with AI and incorporating generated code into their normal development work. Different developers use different prompts, techniques, tools, and approaches.

This experimentation is valuable because developers are discovering where AI fits naturally into their work.

But the capability remains largely individual rather than shared.

What This Looks Like

Observable signals include:

• AI references appearing in pull requests
• AI-generated code appearing in the codebase
• Developers sharing useful prompts informally
• Developers maintaining personal AI coding approaches
• Different AI techniques emerging across the same team

The organization now has useful experience to learn from.

The challenge is preventing that learning from remaining trapped with individual developers.

How to Progress to Enabling

Start harvesting what developers are learning.

Ask:

• Which coding tasks consistently benefit from AI?
• Which prompts or approaches work particularly well?
• Where does AI-generated code require significant rework?
• What context improves the results?
• What review concerns appear repeatedly?
• Which developers have discovered practices worth sharing?

Use those findings to identify common patterns.

DORA's research emphasizes that successful AI adoption is a systems problem rather than simply a tools problem. Moving beyond isolated individual usage requires strengthening the practices and environment surrounding the technology.

Practical Example: Run an AI Coding Practice Harvest

Ask developers to submit one AI-assisted coding example from the previous two weeks.

For each example, capture:

• Coding Task: What was the developer trying to accomplish?
• AI Approach: How was AI used?
• Prompt / Context: What information was provided to the AI?
• Result: Did it save time, improve quality, or create rework?
• Learning: What should another developer know before trying it?

Review the submissions as a team and identify the five practices worth making repeatable first.

For example:

Generate unit tests from existing code, but require developers to review coverage and validate edge cases before acceptance.

This begins converting personal AI techniques into shared team knowledge.

 

Enabling: Shared AI Coding Practices Are Defined

At the Enabling stage, teams begin establishing shared and consistent AI-assisted coding practices.

Common tools, prompts, review expectations, and usage patterns become visible. Developers have greater clarity around how and where AI should be used, while teams continue learning what works in their specific context.

This is an important transition.

The goal is not to eliminate developer autonomy or prescribe exactly how every developer interacts with AI. The goal is to create enough shared guidance that AI-assisted coding becomes repeatable and reliable rather than completely dependent on individual technique.

What This Looks Like

Evidence includes:

• Published AI coding guidelines
• Shared IDE AI tooling
• Reusable coding prompt templates
• Consistent AI patterns appearing in code reviews
• Agreed expectations for reviewing AI-generated code
• Shared practices beginning to appear across teams

The organization has begun answering:

"What does effective AI-assisted coding look like here?"

How to Progress to Operationalizing

Create a lightweight AI Coding Playbook based on what teams have already learned.

The playbook should answer practical questions such as:

• Which AI coding tools are supported?
• What coding tasks are good candidates for AI assistance?
• When is AI-generated code appropriate?
• What information should never be included in prompts?
• What human review is required?
• What testing expectations apply?
• How should developers provide context to AI?
• Which prompts or workflows should teams reuse?

Avoid creating a 40-page policy document.

Developers need guidance they can use inside the development workflow.

GitHub, for example, supports repository-specific instruction files that give AI coding tools information about project conventions, build commands, testing expectations, coding styles, and workflows. Its guidance recommends keeping those instructions concise and actionable.

Practical Example: Create Repository-Level AI Coding Instructions

Create a simple repository instruction file or equivalent team guidance covering:

Build

• Standard build commands

Testing

• Required test commands

• Expectations for new tests

Architecture

• Important design constraints

Coding Standards

• Language and framework conventions

Workflow

• Branching practices

• Required validation before committing

AI Review

• AI-generated code requires human review

• Tests must pass

• Security concerns require human validation

Then create 5–10 reusable prompts for recurring activities such as:

• Generate tests
• Refactor code
• Explain a component
• Investigate a defect
• Review an implementation
• Create initial documentation

This gives developers a reliable starting point while preserving room for experimentation.

 

Operationalizing: AI Becomes Part of the Standard Development Workflow

At the Operationalizing stage, AI-assisted coding is embedded into established development workflows and teams begin measuring its effect on software delivery.

Teams no longer rely on individual developers remembering when or how to use AI.

AI-assisted coding practices become part of the standard AI workflow, and the organization evaluates whether those practices reliably produce the intended outcomes.

What This Looks Like

Observable evidence includes:

• Pull request templates incorporating relevant AI information
• Definition of Done including AI code-review expectations
• Development practices accounting for AI-assisted work
• AI-generated code consistently reviewed and tested
• Delivery measures used to evaluate the effect of AI-assisted coding
• Rework and quality trends becoming visible

At this stage, AI usage becomes visible not only in individual code artifacts but in the system through which software moves from idea to production.

How to Progress to Optimizing

Choose specific points in the software development workflow where AI should be intentionally incorporated.

For example:

Story → Plan → Code → Test → Review → Merge

An AI-assisted workflow could evolve toward:

Story → AI-Assisted Planning → AI-Assisted Coding → AI-Generated Test Suggestions → Human + AI Code Review → Merge

The objective is not to insert AI into every step.

It is to identify where AI consistently improves the flow of work and make those practices part of normal delivery.

Then measure the result.

Useful measures include:

• Coding cycle time
• Pull request lead time
• Rework
• Build success
• Defect rate
• Code-review time
• Change failure percentage
• Developer experience

GitHub's Accenture study provides an example of looking beyond adoption metrics. The study examined pull requests, merge rates, successful builds, usage telemetry, and developer experience rather than measuring AI success solely through licenses or prompt volume.

Practical Example: Create an AI-Assisted Pull Request Workflow

Update your pull request template to include:

AI Assistance

• Was AI used in creating or modifying this code?
• Where was it used?
• What AI-generated code required significant modification?

Validation

• Were AI-generated changes reviewed by the developer?
• Were appropriate automated tests run?
• Were security and quality considerations reviewed?

Then add relevant AI practices to the team's Definition of Done:

AI-generated code has been reviewed, tested, and held to the same engineering standards as human-generated code.

Run the approach with one or two teams for 6–8 weeks.

Compare:

Before AI Workflow → After AI Workflow

using measures such as:

• PR cycle time
• Review time
• Rework
• Build failures
• Escaped defects

The question is no longer simply:

"Did developers use AI?"

It becomes:

"Did our AI-assisted development workflow improve software delivery?"

 

Optimizing: AI Coding Practices Continuously Improve

At the Optimizing stage, AI-assisted coding is an established part of software development and teams continuously improve how AI participates in the coding workflow.

AI capabilities evolve rapidly.

Models improve. Coding agents take on larger tasks. Development teams discover better prompting and context strategies. New risks emerge. Workflows that worked six months ago do not automatically remain the best use of AI.

Organizations at this stage deliberately inspect and improve how AI participates in software development.

What This Looks Like

Observable signals include:

• Regular reviews of AI's impact on coding workflows
• AI handling increasingly larger units of development work
• Improved context-enrichment and prompting approaches
• Positive cycle-time, quality, or rework trends
• Successful practices spreading across teams
• New AI capabilities being evaluated through controlled experiments

The defining characteristic is not more AI.

It is better AI-assisted software delivery over time.

How to Sustain and Continuously Improve

Create a continuous learning loop around AI-assisted development:

Measure → Learn → Experiment → Adapt → Measure Again

Look beyond individual prompt techniques.

Ask broader workflow questions:

• Which coding activities should AI now handle?
• Which activities should remain primarily human?
• Can AI take on larger units of work safely?
• Where does AI still produce excessive rework?
• What repository context improves results?
• Which AI-generated changes fail review most often?
• What has changed in cycle time or quality?
• Are developer-level improvements creating bottlenecks elsewhere?
  • This becomes especially important when AI improves one part of delivery while constraints remain elsewhere across teams and workflows.

DORA's 2025 research reinforces the importance of examining AI within the broader delivery system rather than optimizing isolated developer activity. AI amplifies strengths and weaknesses, making system-level measurement especially important.

Practical Example: Run a Quarterly AI Coding Workflow Review

Every quarter, bring engineering leaders and representative developers together and review four areas:

1. Usage

Where is AI being used?

• Code generation
• Testing
• Refactoring
• Debugging
• Code review
• Larger agentic tasks

2. Outcomes

What has changed?

• Cycle time
• Rework
• Quality
• Build success
• Developer experience

3. Friction

Where is AI creating problems?

• Poor context
• Incorrect code
• Excessive review
• Security concerns
• Prompt inconsistency
• Workflow bottlenecks

4. Experiments

What should we test next?

For example:

Hypothesis: Providing repository architecture and testing instructions to the AI will reduce rework on AI-generated pull requests.

Or:

Hypothesis: Allowing an AI coding agent to complete bounded defects end-to-end will reduce cycle time without increasing escaped defects.

Add selected experiments to an AI Coding Improvement Backlog, assign an owner, identify the expected measure, and review results during the next session.

GitHub's guidance increasingly emphasizes providing AI coding systems with repository-specific context, planning before coding, verification, and structured workflows such as explore → plan → code → verify → commit.

Optimization means continually discovering how much of the coding workflow AI can effectively support while improving—not degrading—the performance of the overall engineering system.

 

Key Takeaway

AI-assisted coding maturity isn't measured by how much code AI generates. It is demonstrated when teams consistently use AI within their development workflows to improve delivery outcomes—and continuously get better at doing so.

 

From AI Adoption to AI Operationalization

Moving from AI coding adoption to AI operationalization means progressing from tool access and individual experimentation to shared practices, integrated workflows, measurement, and continuous improvement.

Giving developers an AI coding assistant creates access.

Individual experimentation creates learning.

Shared practices create repeatability.

Embedding those practices into the development workflow creates operationalization. That is the shift from AI coding adoption to AI coding operationalization. For a broader look at this transition beyond coding, see AI Adoption vs. AI Operationalization: Why Deploying Tools Isn't Enough.

Measuring and improving those workflows creates sustained capability.

The AI-Assisted Coding Maturity Model therefore progresses from:

Manual Coding → Individual AI Use → Shared Practices → Integrated AI Workflow → Continuous Improvement

This matters because local developer productivity does not automatically translate into better software delivery.

If AI enables developers to generate code faster but creates additional review, rework, testing, security, or deployment problems downstream, the organization has optimized one activity rather than the overall system.

When teams intentionally integrate AI into the coding workflow, preserve strong engineering practices, measure the effect, and adjust based on evidence, AI-assisted coding becomes part of a stronger software delivery capability.

That is the shift from AI coding adoption to AI coding operationalization.

 

Understand Where Your Organization Stands

Lean Agile Intelligence's AI assessments help organizations establish a baseline across critical AI capabilities, identify gaps, prioritize improvement opportunities, and measure progress over time.

For AI-Assisted Coding, that means looking beyond whether developers have access to AI coding tools and understanding whether AI usage has progressed from isolated experimentation to shared practices, integrated development workflows, measurable outcomes, and continuous improvement.

Assess your AI capabilities and identify what your organization needs to operationalize next.