Every engineering leader I talk to says the same thing: their organization is "leveraging AI." When I ask what that actually means, the answer usually lives somewhere between "we bought Claude Code licenses" and "we have a working group looking at agentic patterns." The honest version is that most organizations are stuck in phase one and don't know it.
I've spent the last two years working through what AI adoption actually looks like inside engineering organizations, both as an operator and an advisor to PE-backed fintech and SaaS companies. The pattern is consistent. There are four phases. Each phase requires different things. Most companies announce phase three while operating in phase one, and the gap between announcement and reality is where the productivity claims fall apart.
Phase One: Individual Experimentation:
This is where everyone starts. Copilot or ClaudeCode licenses are purchased for Engineering. Some engineers use them aggressively, some never log in. The fast adopters quietly become more productive. The skeptics keep doing things the old way. There's no shared standard, no measurement, no governance.
Phase one feels productive because individuals are getting faster. But the organizational gain is invisible. You cannot point to a sprint that delivered more, a release that shipped sooner, or a feature that landed cheaper. The productivity is there but it is trapped inside individual workflows.
Most engineering organizations I look at are here. They have tools. They do not have an AI strategy.
Phase Two: Team Standardization:
Phase two is when the organization stops treating AI as a personal tool and starts treating it as a shared one. Teams pick a primary platform and stick to it. Code review processes adapt to the reality that some of the code being reviewed was drafted by an AI. Security scanning gets tuned for AI output patterns, which look different from human output patterns. Prompt libraries start to emerge, often informally at first.
The harder work in phase two is cultural. Senior engineers who built careers on craftsmanship sometimes resist. Junior engineers who skipped the craftsmanship step sometimes over-trust the output. The engineering manager's job becomes calibrating both groups against a shared standard.
Most organizations that have moved past phase one stop here. They get a real productivity bump, maybe 20 to 30 percent on certain workflows, and they declare victory. That is fine if your goal is incremental improvement. It is not enough if your goal is structural change.
Phase Three: Workflow Integration:
Phase three is where AI stops being a tool engineers use and starts being part of the platform itself. Code generation gets tied directly to design systems. Test creation runs as part of the CI pipeline. Documentation drafts itself from the code. Refactoring becomes a multi-step agentic workflow rather than a manual project.
This is where the productivity numbers start to look meaningful. When I implemented an AI development program inside an engineering organization, I realized 3x gains in productivity. Industry research backs the pattern. GitHub's own studies show developers using Copilot complete tasks roughly 55 percent faster on isolated work, and the multiplier compounds when AI is integrated into the platform rather than bolted onto it. Engineer sentiment shifts as well. Skeptics convert quickly once the work becomes more interesting rather than less. That is what phase three looks like when it works.
It is also where most efforts fail. Phase three requires real platform investment. It requires engineering leaders to make decisions about which workflows to automate first, which to leave alone, and how to handle the parts of the codebase where AI output is not yet trustworthy. It requires governance. Most organizations skip the governance work and end up with a productivity gain that looks great in the first month and starts producing technical debt by the third.
The organizations that get phase three right have a few things in common. They have a clear picture of what their codebase looks like. They have an honest assessment of where AI is reliable and where it is not. They have an executive who understands that the goal is structural leverage, not just faster typing.
Phase Four: AI-Native Engineering:
Phase four is the rarest and the one most leaders talk about without understanding what it actually means. In phase four, the engineering organization is designed from the ground up assuming AI is a co-developer. Architecture decisions get made with AI involvement in mind. Documentation gets written for both humans and machines because both will be reading it. Code review evolves into something different. The role of the engineer shifts toward judgment, design, integration, and verification, while AI handles drafting, transformation, and large-scale refactoring.
I have not seen a fully phase-four organization yet. I have seen pieces of it. The companies closest to it tend to be smaller, newer, and led by founders who built their assumptions about software development inside the AI era rather than retrofitting them.
The interesting question for established organizations is not whether to aspire to phase four. It is whether the journey from phase three to phase four requires rebuilding the engineering culture you spent ten years building. My current view is that for most companies, the answer is yes. That is not a comfortable answer.
Why This Matters The reason I find this framework useful is that it forces honesty about where an organization actually sits. Most CTOs I talk to want to talk about phase three and four. The engineering reality is usually phase one with aspirations. Closing that gap is where the work is.
If you are an engineering leader trying to figure out where to invest your next dollar of AI effort, the framework is straightforward. Diagnose your real phase. Resist the urge to skip phases. Phase two before phase three. Phase three before phase four. The companies that try to leapfrog tend to spend a lot of money and produce demos rather than durable productivity.
If you are a PE investor or board member trying to evaluate whether a portfolio company is genuinely getting AI leverage or just running a workshop, ask one question: what is the measured productivity delta and where in the workflow does it come from. The answer will tell you whether the organization is in phase one with a marketing budget or doing the harder work of building real leverage.
The companies that get this right over the next two years will run materially leaner organizations than their competitors. The ones that do not will keep announcing AI strategies while the gap widens.