Many individuals have tried AI instruments and walked away unimpressed. I get it — many demos promise magic, however in follow, the outcomes can really feel underwhelming.
That’s why I need to write this not as a futurist prediction, however from lived expertise. Over the previous six months, I turned my engineering group AI-first. I’ve shared earlier than concerning the system behind that transformation — how we constructed the workflows, the metrics, and the guardrails. As we speak, I need to zoom out from the mechanics and speak about what I’ve discovered from that have — about the place our occupation is heading when software program growth itself turns inside out.
Earlier than I do, a few numbers as an example the dimensions of change. Subjectively, it feels that we’re transferring twice as quick. Objectively, right here’s how the throughput developed. Our complete engineering group headcount floated from 36 in the beginning of the 12 months to 30. So that you get ~170% throughput on ~80% headcount, which matches the subjective ~2x.
Zooming in, I picked a few our senior engineers who began the 12 months in a extra conventional software program engineering course of and ended it within the AI-first manner. [The dips correspond to vacations and off-sites]:
Observe that our PRs are tied to JIRA tickets, and the typical scope of these tickets didn’t change a lot by the 12 months, so it’s nearly as good a proxy as the info can provide us.
Qualitatively, trying on the enterprise worth, I really see even greater uplift. One purpose is that, as we began final 12 months, our high quality assurance (QA) group couldn’t sustain with our engineers' velocity. As the corporate chief, I wasn’t pleased with the standard of a few of our early releases. As we progressed by the 12 months, and tooled our AI workflows to incorporate writing unit and end-to-end assessments, our protection improved, the variety of bugs dropped, customers turned followers, and the enterprise worth of engineering work multiplied.
From massive design to fast experimentation
Earlier than AI, we spent weeks perfecting consumer flows earlier than writing code. It made sense when change was costly. Agile helped, however even then, testing a number of product concepts was too expensive.
As soon as we went AI-first, that trade-off disappeared. The price of experimentation collapsed. An concept may go from whiteboard to a working prototype in a day: From concept to AI-generated product necessities doc (PRD), to AI-generated tech spec, to AI-assisted implementation.
It manifested itself in some superb transformations. Our web site—central to our acquisition and inbound demand—is now a product-scale system with tons of of customized elements, all designed, developed, and maintained immediately in code by our inventive director.
Now, as a substitute of validating with slides or static prototypes, we validate with working merchandise. We check concepts reside, study quicker, and launch main updates each different month, a tempo I couldn’t think about three years in the past.
For instance, Zen CLI was first written in Kotlin, however then we modified our thoughts and moved it to TypeScript with no launch velocity misplaced.
Instead of mocking the options, our UX designers and mission managers vibe code them. And when the release-time crunch hit everybody, they jumped into motion and stuck dozens of small particulars with production-ready PRs to assist us ship an excellent product. This included an in a single day UI format change.
From coding to validation
The following shift got here the place I least anticipated it: Validation.
In a conventional org, most individuals write code and a smaller group assessments it. However when AI generates a lot of the implementation, the leverage level strikes. The true worth lies in defining what “good” seems like — in making correctness express.
We assist 70-plus programming languages and numerous integrations. Our QA engineers have developed into system architects. They construct AI brokers that generate and keep acceptance assessments immediately from necessities. And people brokers are embedded into the codified AI workflows that permit us to attain predictable engineering outcomes through the use of a system.
That is what “shift left” actually means. Validation isn’t a stand-alone perform, it’s an integral a part of the manufacturing course of. If the agent can’t validate it’s work, it could possibly’t be trusted to generate manufacturing code. For QA professionals, this can be a second of reinvention, the place, with the suitable upskilling, their work turns into a important enabler and accelerator of the AI adoption.
Product managers, tech leads, and knowledge engineers now share this accountability as effectively, as a result of defining correctness has grow to be a cross-functional talent, not a task confined to QA.
From diamond to double funnel
For many years, software program growth adopted a “diamond” form: A small product group handed off to a big engineering group, then narrowed once more by QA.
As we speak, that geometry is flipping. People interact extra deeply in the beginning — defining intent, exploring choices — and once more on the finish, validating outcomes. The center, the place AI executes, is quicker and narrower.
It’s not only a new workflow; it’s a structural inversion.
The mannequin seems much less like an meeting line and extra like a management tower. People set path and constraints, AI handles execution at velocity, and folks step again in to validate outcomes earlier than selections land in manufacturing.
Engineering at the next degree of abstraction
Each main leap in software program raised our degree of abstraction — from punch playing cards to high-level programming languages, from {hardware} to cloud. AI is the subsequent step. Our engineers now work at a meta-layer: Orchestrating AI workflows, tuning agentic directions and abilities, and defining guardrails. The machines construct; the people resolve what and why.
Groups now routinely resolve when AI output is protected to merge with out evaluate, how tightly to sure agent autonomy in manufacturing programs, and what indicators really point out correctness at scale, selections that merely didn’t exist earlier than.
And that’s the paradox of AI-first engineering — it feels much less like coding, and extra like considering. Welcome to the brand new period of human intelligence, powered by AI.
Andrew Filev is founder and CEO of Zencoder
