ADM AI Expert Series #2: Our Journey with AI-Driven Development
As CTO and AI Lead at ADM Interactive, I am in charge of implementing AI-agenti software development across the company. It’s been a journey full of surprises, some frustrations, and quite a few “aha” moments that I want to share with you today.
Currently, there are two camps in the Software Engineering field: enthusiasts and
skeptics about AI-driven development. Most enthusiasts are representatives of
solo developers or small development teams. Individuals and tiny teams
achieving $1-100 million in annual revenue just by using AI agents instead of
human software developers. Companies like Lovable reaching $100 million ARR
with just 45 employees. Skeptics are mostly those who develop large
enterprise-grade systems working in a bigger development teams.
Well, applying AI-driven development in a 6-8 member development team
working on enterprise-grade projects? That’s a completely different game. At
ADM we decided to pilot it! Spoiler: it’s challenging, yet we did it!
Our Setup with Claude Code
As of today, the best AI coding agent implementation I have tried has been
done by Anthropic, with their Claude Code. Thats why we decided to use
Claude Code for our development process.
The first thing we did was establish clear rules. We created a CLAUDE.md file in
project that contains project description, coding style, which libraries and
components to use, how to deploy and test, etc. – basically everything that is
needed for development. This same file was shared across all team members,
so every instance of Claude Code would follow the same patterns.
We also created comprehensive documentation for our existing codebase that
we kept updating during development. Without this, Claude Code would
suggest solutions that didn’t fit with what we already had.
To make Claude Code more effective, we connected Figma through MCP so it
could retrieve UI mockups directly. We also set up a Puppeteer server through
MCP so Claude Code could test functionality itself: then Claude Code can
actually open a browser, go to a specific URL, and test end-to-end
functionality, as a human QA Testing Engineer would do.
The Documentation Problem
Here’s where we hit our first real bottleneck. We had requirements
documentation – lots of it. But it wasn’t working for Claude Code. There was no
single document that would be single-point-of-truth for AI, and it became a
problem – AI needed clarifications and more detailed instructions from the
developer, and multiple iterations to achieve desired functionality.
So, we needed to create shorter, more specific version of requirements
documentation specifically for AI. This wasn’t just about reformatting – it
required going back to stakeholders to clarify ambiguities that human
developers probably would just figure out but AI wouldn’t.
It’s interesting that this aligns with what other companies are experiencing.
According to industry research, European fintech unicorns achieved 30-40%
cost reductions through AI adoption but faced quality issues. Klarna, for
example, cut 40% of their workforce using AI, only to reverse course and hire
humans again when their AI-first approach degraded customer service quality.
The pattern is clear: AI amplifies existing problems, including unclear
requirements.
The “It’s Done” Problem
Another surprise: when Claude Code says all required functionality is complete,
you need to check. In our experience, we quite often need one or even two
more iterations before all required functionality is actually implemented. A
human eye needs to supervise and review everything.
This matches the broader industry paradox. Research shows that despite AIgenerated
code introducing 10 times more security vulnerabilities and frequent
reliability issues, developers are spending $200-400 monthly on these services
because they’ve become indispensable for modern development workflows.
We’re all accepting these trade-offs because the productivity gains are real.
That’s why we implemented mandatory code review for all AI-generated code
before committing. It adds time, but it’s saved us from numerous issues.
How We Changed Our Development Process
We’ve essentially reduced our Agile approach – no sprints, no grooming
sessions, no planning meetings that take half a day. We kept only daily
standups for status updates and information exchange. Everything else shifted
to continuous AI-driven development.
This might sound extreme, but it makes sense when AI can generate in hours
what used to take days. Traditional estimation and planning become less
relevant. And we’re not alone in this – Spotify achieved an 86% adoption rate of
AI-powered development tools, and they have also moved away from traditional
sprint structures.
Our Key Discovery: It’s Not Just About Code
At ADM Interactive, we believe the solution isn’t just using AI for coding. The
real breakthrough came when we started applying AI to the entire end-to-end
flow: requirements identification and analysis, solution brainstorming and
design, and documentation creation. These were actually our main bottlenecks.
Now, before we code, we use AI to help translate what stakeholders want into
clear technical specifications. We use it to generate architectural proposals. We
also create documentation before coding starts – AI generates it, humans
review it, stakeholders approve it, and only then does implementation begin.
This shift has been more impactful than just faster coding. It’s helped us
identify unclear requirements early, consider alternatives we might have
missed, and maintain documentation that’s actually up-to-date, always.
What We’ve Learned
Working with AI coding agents in a team setting requires different thinking than
solo development.
You need:
- Standardization: That CLAUDE.md file isn’t optional. Without it, you’llget eight different coding styles from eight team members using AI.
- Better and clear requirements: If your requirements are unclear to human engineers, they’re useless to AI. We had to get much better at writing clear, specific requirements.
- Human oversight: Always. Every piece of AI code needs human review, at least very brief. The AI doesn’t catch everything, especially integration issues between different parts of the system.
- Process changes: Traditional Agile doesn’t fit anymore. You need something more fluid that matches the speed AI enables.
- Broader AI application: Don’t just use AI for coding. Apply it to requirements, design, documentation – the entire pipeline.
Moving Forward
The transition to AI-driven development isn’t just about adopting new tools. It’s
about rethinking how we build software. Despite all the challenges we’ve faced,
AI-driven development has genuinely accelerated our delivery. We’re building
features faster, maintaining better documentation, and our developers are
focusing on what matters most – solving complex problems!
If you’re considering adopting AI-driven development for your team and want to
avoid the pitfalls we’ve encountered, feel free to reach out.