In my previous publication I tried to explain why I have a quarrel with a phenomenon I called "Agentic Evangelization" - by which I mean uncritical optimism about Agentic Coding being a magic wand that fixes all software delivery challenges with a single spell.

Today I'd like to expand on that thought by addressing an aspect of the Software Development Lifecycle that rarely gets mentioned in the ever-optimistic Agentic narrative.

What Agentic Coding does well (most of the time) is increasing the productivity of software implementation. That's an important distinction. It is primarily about creating code - and, at least today, only partially about everything around it.

Let's assume for a moment that this problem truly has been solved. We now have a mature engineering workflow that works like a Swiss watch: Specification → Context retrieval → Task planning → AI implementation → Automated testing & validation → Human review → Merge → Deployment → Monitoring.

We have a robust harness, great tooling and engineering playbooks. Code generation itself is no longer a challenge. What is crucial to realize, however, is that implementation is only one stage of a much larger delivery system.

IT projects come in immensely different shapes and sizes. We may be building a greenfield application for a SMB, modernizing a legacy enterprise platform, migrating an existing system where the target architecture is already well understood, or creating an entirely new digital product for a heavily regulated industry with hundreds of stakeholders and months of discovery work ahead.

The SDLC by principle is similar in all of those scenarios, but the relative importance of its individual phases changes dramatically.

For a straightforward migration or greenfield project, implementation may indeed be the heaviest lift. But for large enterprise programs, the main weight often lies elsewhere - it’s product discovery, stakeholder alignment, business analysis, architecture, governance, validation. This is why I believe many discussions around Agentic Coding focus on the wrong question.

The question is not whether Agentic Coding makes developers faster. It (usually) does.

The real question is whether implementation is actually the main constraint that limits the success of your delivery.

In an enterprise program we participated in, the client introduced Agentic Coding into the development workflow. The rollout focused almost exclusively on the implementation phase of SDLC. Initially, the results looked promising. Development throughput increased noticeably.

However, it quickly became apparent that the organization as a whole wasn't prepared for that increase in capacity. Product teams struggled to provide enough refined requirements. QA suddenly had significantly more output to validate. The same happened with compliance and security screenings. Planning became more difficult because implementation stopped being the pacing function of the program. The delivery team started producing work faster than downstream functions could absorb it. The system became faster locally, but not globally, the organization became less balanced and even more prone to project failure than before.

For decades, research from Standish Group, PMI, McKinsey and other respected institutions has consistently shown that software projects rarely fail because developers write code too slowly.

The dominant causes remain remarkably consistent:

unclear or constantly changing requirements
weak business sponsorship
poor stakeholder alignment
unrealistic expectations
ineffective governance
uncontrolled scope growth
organizational complexity

Implementation speed has simply never been identified as one of the dominant predictors of project success or failure. It’s a significant cost-driver, of course. But a success determinator? Available data shows not usually.

So why are we so uncritically optimistic?

Perhaps because implementation is the most visible part of software development. It is measurable. It is easy to demonstrate. Watching an AI generate thousands of lines of working code in minutes is genuinely impressive. But software delivery is a system problem rather than a coding problem. Optimizing one component does not necessarily optimize the entire thing.

So then… Can AI improve the system as a whole? I absolutely believe that it can, but it’s a far deeper and more difficult transition than buying a few enterprise licenses.

AI can significantly improve product discovery by synthesizing user research, market feedback and stakeholder interviews.

It can accelerate UI and product design by generating interface concepts, interactive prototypes and multiple design alternatives long before implementation begins.

It can improve requirements definition by identifying inconsistencies, ambiguities and defining acceptance criteria.

It can strengthen project and program management by increasing observability across delivery pipelines (Jira/Confluence/Slack/Calendar/Github + Claude via MCP works like a charm).

Those are only examples, but the point is that successful AI adoption is much broader than Agentic Coding and in many cases, it seems to now be the main (or even only) point of focus for the industry.

So then let’s put Agents everywhere and they will solve all our problems, right? Not necessarily. Broadening AI adoption beyond software implementation is a step in the right direction, but software delivery remains a socio-technical discipline. As in the examples invoked above, AI can improve many activities across the SDLC, yet organizations still need clear priorities, effective governance and aligned stakeholders to turn faster execution into successful outcomes. AI may accelerate it, but it cannot replace leadership, ownership, or organizational alignment.

In conclusion, I really believe AI has the potential to become one of the most important strategic enablers our industry has ever seen. It may very well improve the success rate of software projects. But as of now, there is no credible evidence that it does. Not because it doesn't work, but simply because we don't know yet.

I remain optimistic about it, but also careful.

Thank you for your attention to that matter.

MS

AI makes software implementation faster. But will it make your project successful?

Related articles