I've been watching the developers struggle with the same problem for months now. They're all excited about AI development tools - and rightfully so - but their projects keep veering off course in ways that remind me of the early days of agile adoption.
Last week, I sat in a project review where we implemented a feature for one of our products. The development team had used GitHub Copilot extensively, the code was technically sound, but nobody in the room recognized the final product. It wasn't anyone's fault, really. The specifications had been verbal, the requirements evolved through Teams messages, and the AI had faithfully built exactly what it was asked to build in each isolated conversation.
This isn't unique to AI projects, but AI amplifies the problem in ways we're still learning to handle.
The Reality of AI-Powered Development
Most of our clients are now using some form of AI assistance - ChatGPT for brainstorming, Copilot for code generation, or Claude for documentation. The productivity gains are real, but so are the unexpected detours.
Here's what I keep seeing: developers ask an AI to "build a shopping cart," get something functional, then later ask for "payment integration" and receive a completely different cart implementation. Ask for user accounts the next day, and you might get a third version that breaks the payment flow you just added.
Each conversation exists in a vacuum. Unlike human developers who remember yesterday's decisions and maintain some consistency, AI starts fresh every time. That's both its strength and its biggest weakness.
I've started calling this "conversational drift" - each interaction nudges the project in a slightly different direction until you end up somewhere completely unexpected.
Think of it like building a house where each contractor shows up with no knowledge of what the last one did. The foundation might be solid, but the walls, wiring, and roof end up mismatched.
The Tools That Make It Possible
A new category of tools has emerged specifically to solve this problem. Rather than just writing better documentation, these tools create structured workflows that AI assistants can follow consistently.
- OpenSpec is particularly good for teams working with existing systems. If you're adding features to current applications, it excels at tracking what's changing versus what stays the same. It's free and works with any AI tool your team already uses.
- GitHub Spec Kit offers the most comprehensive approach if you're starting fresh projects. It includes built-in quality gates that prevent AI from over-engineering solutions.
- Kiro provides an enterprise-grade solution with advanced automation, though it comes with higher costs and requires more setup.
These tools work by creating what I call "executable specifications" - structured blueprints that any AI assistant can follow, whether your team uses GitHub Copilot, Claude, or ChatGPT.
How This Transforms Your Projects
Rather than saying "build user authentication" you specify: "Users register with email and password. Passwords must be at least 8 characters. Registration sends a confirmation email. Users can't access the system until confirmed. Failed login attempts are tracked and accounts lock after 5 failures."
The difference is profound. AI gets clear guidance, stakeholders can review expectations before development starts, and everyone knows what "done" looks like.
If you're managing projects: Specifications become your project timeline. Instead of guessing what developers are building, you can track progress against documented features.
If you're writing requirements: Your user stories now need to include edge cases and error conditions. AI doesn't make assumptions about "obvious" behaviour.
If you're testing software: The specifications double as your test cases. Every scenario described becomes something you can verify.
Making the Transition
You don't need to revolutionize your entire development process overnight. Pick your next AI-assisted project and try this approach:
- Install one of the recommended tools (most take under 5 minutes).
- Create a change proposal describing what you want to build.
- Let the tool generate structured specifications from your requirements.
- Review and refine the specs with stakeholders.
- Use your existing AI tools to implement the solution based on the specs.
You'll probably find that the extra planning time upfront saves significantly more time during development, testing, and stakeholder reviews.
The Competitive Advantage
Organizations that master spec-driven development will deliver AI projects faster and more reliably than competitors still treating AI like magic. The companies figuring this out now will be building more software, faster, with higher quality while their competitors struggle with inconsistent AI outputs and project overruns.
At Advania, we're helping clients navigate this transition because the alternative - continuing to treat AI like human developers who read between the lines - leads to projects that technically succeed while failing to solve real business problems.