Introduction
In 2026, AI-assisted development is nothing new. According to a report from Anthropic, developers have integrated AI tools into an average of 60% of their daily work. But paradoxically, while code output has accelerated, the stability of software delivery hasn't improved alongside it. Companies have poured resources into adopting AI, yet the efficiency gains often fall short of expectations.
So, where does the problem lie?
This post outlines three phenomena I've recently observed, hoping to offer some different perspectives.
1. Writing Code by Hand Is Becoming a Traditional Craft
After this year, writing code by hand might gradually become an ancient art.
AI has dramatically compressed the costs of development and testing, forcing a reinvention of software workflows. In most scenarios, humans no longer need to write every single line of code. However, clarifying requirements and defining boundaries are still tasks that require human intervention. Breaking down vague requirements into finer details and providing enough context for an AI to execute correctly are things AI still can't do.
I've had a very real feeling lately: AI develops so fast that ideas sometimes can't keep up with the speed of implementation.
But this doesn't mean we should just hand everything over to AI. AI is fundamentally a capability amplifier. If your ability is 0, multiplying it by 100 still results in 0. The real advantage belongs to those who understand the entire software development lifecycle: people with a problem-solving mindset, a grasp of underlying principles, attention to detail, and the flexibility to apply technology. It is their productivity that will be magnified several times over by AI.
What the Data Says
According to Anthropic's 2026 Agentic Coding Trends Report, 78% of Claude Code sessions now involve multi-file editing (up from just 34% in Q1 2025), and the average session duration has increased from 4 minutes to 23 minutes. AI is evolving from "assisting with code snippets" to "global collaboration based on an understanding of the overall workflow."
However, the 2025 DORA Report presents a noteworthy set of data:
Metric | Change in High-AI-Adoption Teams |
|---|---|
Task Completion Volume | +21% |
PR Merge Volume | +98% |
PR Review Time | +91% (New Bottleneck) |
Change Failure Rate | No improvement, even negatively correlated |
MTTR (Mean Time to Repair) | No improvement |
Code output has indeed accelerated, but review speed can't keep up, and software delivery stability hasn't improved with AI. Simply put: AI can help you write faster, but if your understanding isn't deep enough, you're just accelerating the creation of bugs. The time required for human tasks may, ironically, become the new bottleneck after AI acceleration.
Compared to rushing to have AI write code, understanding business processes and mastering system architecture have become more important than ever.
2. Why AI Adoption Isn't Accelerating Traditional Enterprises
Many people assume that adopting AI will lead to a massive boost in efficiency, but reality is often different. The reason lies not in the technology itself, but in organizational structure and collaboration models.
The Hidden Costs of Siloed Workflows
A common development process in the past looked something like this: PM → UI/UX Design → Frontend Development → Backend Development → System Integration → Deployment. Each stage has a clear owner, and the division of labor is well-defined, but every handoff creates the potential for information gaps.
Common real-world scenarios include: requirements needing repeated back-and-forth clarification during handoffs; waiting and coordination time approaching or even exceeding actual development time; the frontend team finishing their work only to find the API spec doesn't match, and after the backend fixes it, the UI flow has changed again.
A more practical problem is that some teams want to implement automated processes (CI/CD, automated testing, Design-to-Code), but the resistance is often not technical. Instead, their development and collaboration models are still stuck in a mindset of manual handoffs, making it impossible to connect the workflows.
AI Needs Full Context, But Organizational Structures Can't Provide It
For AI tools to produce consistently high-quality results, there's a key prerequisite: sufficient and complete context. But if information is fragmented across different roles, the context the AI receives is incomplete, and the quality of its output will naturally suffer.
I've observed this myself. When I only receive the requirements for my specific part of the project, lacking the surrounding context, the code AI generates is technically correct but doesn't fit into the overall workflow. This isn't an AI problem; it's an information silo problem.
Five Common Barriers to Adoption
The biggest obstacles to truly implementing AI in an organization are usually not technical:
-
Multiple organizational layers make it difficult to centralize information, preventing AI from getting the full context.
-
Learning curve: Adopting new tools requires an adjustment period, and output may temporarily decrease in the short term.
-
Inertia of existing processes: Workflows and divisions of labor have been in place for years, and change takes time.
-
Vague requirements: No matter how powerful, AI can't clarify "what exactly needs to be done" for you.
-
Insufficient business understanding: Writing prompts without understanding the business logic produces output that looks correct but is practically unusable.
This is even more pronounced with legacy systems. On average, enterprises lose about $370 million annually due to technical debt, and poorly categorized data can increase AI adoption costs by as much as 40%. The prerequisite for AI to provide acceleration is that the system itself already has sufficient documentation, clear processes, and an understandable architecture.
According to the Anthropic report, developers have integrated AI into about 60% of their daily work, but only 0–20% of tasks can be fully delegated to AI. This figure is very telling: AI is an excellent collaborator, but "human understanding" remains the core bottleneck.
3. The AI-Driven Trend Towards Full-Stack
Since AI requires complete context, several recent trends have aimed to shorten the handoff chain:
Handoff Point | Method | Effect |
|---|---|---|
Design → Frontend | Figma MCP standards, converting designs directly to code | Reduces back-and-forth communication costs |
Frontend → Backend | Monorepo, OpenAPI (OAS), tRPC | Synchronizes interface documentation with code |
Backend → External Systems | Standardized API documentation, OAS | Lowers integration costs |
As tools increasingly bridge these "handoff" gaps, teams need people who understand the overall workflow and architecture more than they need specialists in a single technical domain.
Role Boundaries Are Blurring
In practice, we can observe more and more cross-disciplinary work: backend engineers starting to handle parts of the frontend or process design, PMs needing to get involved in technical specification details, engineers participating in requirements clarification and project planning, and developers needing to master prompt engineering and model management.
CWS Technology even predicts that by 2026, full-stack developers will evolve into "AI-Stack Developers," requiring cross-disciplinary skills spanning prompt engineering, DevOps integration, and AI governance.
This doesn't mean roles are being replaced; it's more that tools are lowering the barrier to entry, causing the work content between roles to overlap. Research from Faros AI also indicates that about 27% of AI-assisted work falls into the category of "things we didn't do before," such as interactive dashboards and exploratory tools. AI isn't just replacing the existing division of labor; it's creating new scopes of work.
Conclusion
AI has certainly made code production faster, but moving fast doesn't mean you're moving in the right direction.
Looking back at these three phenomena, I think they all point to one thing: understanding "what to do" and "why to do it" has become more critical than the speed of writing code. Only those who can feed complete context to AI can truly unlock its value. As role boundaries blur, the ability to communicate with people from different domains has also become more necessary than ever.
Perhaps the focus in the future will no longer be on a finer division of labor, but on striking a balance between specialized depth and holistic understanding, allowing teams to reduce communication costs while maintaining expertise. This might be the most important direction for teams to consider in the age of AI.
Comments