In recent times, many individuals have experimented with AI tools but often found themselves disappointed. This skepticism is understandable; numerous demonstrations might promise impressive capabilities, but the actual outcomes can sometimes fall short of expectations.
Rather than offering a speculative vision of the future, I aim to share insights based on firsthand experience. Over the past six months, I transitioned my engineering team to an AI-first approach. Previously, I discussed the foundational elements of this transformation, including our methodologies, metrics, and safety measures. Now, I want to step back from the technical details and reflect on what I’ve learned regarding the evolving landscape of software development as we shift toward an AI-centric model.
To illustrate the magnitude of this change, consider some key statistics. Subjectively, it feels as though our pace has doubled. Objectively, our throughput has evolved significantly: our engineering team decreased from 36 members at the start of the year to 30, resulting in approximately 170% productivity with about 80% of our original headcount, reinforcing that subjective perception of a ~2x speedup.
Focusing on our senior engineers, I can depict changes from the start to the end of the year, marking their transition from traditional software practices to the AI-first methodology. [The dips in productivity align with vacations and off-site events]:
It’s important to note that our pull requests are associated with JIRA tickets, and the average scope of these tickets has remained relatively stable throughout the year, making this a reliable proxy for measuring our progress.
From a qualitative perspective, assessing business value reveals an even greater increase. Early on, our quality assurance (QA) team struggled to keep pace with the speed of our engineers. As the leader of the company, I was concerned about the quality of some of our initial releases. However, as we developed our AI workflows to incorporate unit and end-to-end testing, we saw an improvement in our coverage, a reduction in bugs, and an increase in user satisfaction, ultimately boosting the business value of our engineering efforts.
From Big Design to Rapid Experimentation
Previously, we dedicated weeks to refining user flows before commencing coding, which was necessary when change was costly. Agile practices helped, but even then, testing multiple product concepts remained expensive.
Transitioning to an AI-first strategy has eliminated that trade-off. Experimentation costs have drastically declined. Ideas can migrate from conceptualization to a functional prototype within a day: from brainstorming to AI-generated product requirements, to AI-crafted technical specifications, and finally, to AI-assisted implementation.
This shift has led to remarkable transformations. Our website, crucial for acquisition and inbound demand, has evolved into a product-scale system with countless custom components, all designed, developed, and maintained directly in code by our creative director.
Now, rather than validating concepts with slides or static mock-ups, we test with live products. We quickly validate ideas, learn at a faster pace, and roll out significant updates every two months—a rhythm I couldn’t have envisioned three years ago.
For instance, Zen CLI was initially created in Kotlin but was later transitioned to TypeScript without losing any release velocity.
Our UX designers and project managers actively code features instead of merely mocking them up. When the crunch time for releases arrives, they swiftly jump in to fine-tune numerous minor details, submitting production-ready pull requests to facilitate a successful launch. This included implementing a UI layout change overnight.
From Coding to Validation
A surprising shift has occurred in the realm of validation.
In traditional organizations, most team members focus on writing code while a smaller group is tasked with testing it. However, with AI taking on a significant amount of the implementation work, this dynamic has shifted. The emphasis now lies in clearly defining what “good” looks like and making correctness explicit.
Our platform supports over 70 programming languages and countless integrations. Our QA engineers have transformed into system architects, creating AI agents to generate and manage acceptance tests directly based on requirements. These agents are incorporated into our codified AI workflows, leading to consistent and predictable engineering outcomes.
This exemplifies what “shift left” truly means. Validation is no longer a standalone function but an integral part of the production process. If an agent cannot validate its output, it cannot be entrusted to generate production code. For QA professionals, this is a transformative moment: through proper upskilling, their roles evolve into vital catalysts for AI adoption.
Now, product managers, tech leads, and data engineers also share in this responsibility, as defining correctness has become a cross-functional competency rather than a role limited to QA.
From Diamond to Double Funnel
For decades, the software development process has taken on a “diamond” shape: a small product team hands off to a larger engineering team, which then narrows through QA.
Today, this model is undergoing a significant transformation. Humans are more involved at the beginning—defining intent and exploring options—and again at the end, validating outcomes. The middle stage, where AI operates, has become faster and more streamlined.
This is not merely a new workflow; it represents a structural inversion.
The new model resembles a control tower rather than an assembly line. Humans set the direction and parameters, AI manages execution at speed, and people return to validate outcomes before they are finalized for production.
Engineering at a Higher Level of Abstraction
Each significant advancement in software has elevated our level of abstraction—from punch cards to high-level programming languages, from hardware to the cloud. AI represents the next horizon. Our engineers now operate at a meta-layer: orchestrating AI workflows, adjusting instructions and capabilities, and defining guardrails. Machines handle construction; humans determine what and why.
Teams routinely decide on the conditions under which AI output can be merged without review, how closely to constrain agent autonomy in production, and what indicators truly signify correctness at scale—decisions that were simply nonexistent before.
This is the irony of AI-first engineering: the focus shifts from coding to critical thinking. Welcome to a new era of human ingenuity, empowered by AI.
Andrew Filev is the founder and CEO of Zencoder
Welcome to the VentureBeat community!
Our guest posting program allows experts to share their insights, providing neutral, in-depth explorations of AI, data infrastructure, cybersecurity, and other innovative technologies reshaping the future of business.
Read more from our guest post program — and check out our guidelines if you’re interested in contributing an article of your own!
