Two years into the generative AI boom, a strange pattern has emerged across the tech industry: most companies have access to roughly the same AI tools, yet the gap between the ones who benefit from them and the ones who don’t keeps widening. According to Sergii Kravtsov, CEO and co-founder of ConnectiveOne and co-founder of Evergreen IT Development, that gap has almost nothing to do with which model a company uses — and everything to do with whether it rebuilt its processes around AI, or just handed employees a new tool and hoped for the best.
“Most companies I talk to bought subscriptions, gave everyone access, and waited for a transformation to happen on its own,” Kravtsov says. “What they got instead was maybe a 10 to 20 percent productivity bump. That’s roughly where the majority of teams — including plenty of well-funded ones — are still stuck today.”
Kravtsov has spent over 18 years in IT, first building software for other companies through Evergreen, which has delivered more than 400 projects and products, some of which became leaders in their own market niches, before founding ConnectiveOne, an AI-powered communications platform that now serves millions of end users. That dual vantage point — running a services company that implements AI transformation for clients, and a product company that has to make the same approach work on its own systems — has shaped a framework he now describes to almost every founder and executive he talks to.
The gap between having AI and running it
Kravtsov draws a sharp line between what he calls AI adoption and AI management.
“AI adoption is when your team has access to the tools. Licenses are bought, some people are excited, some are skeptical, and that’s basically where it stops,” he explains. “AI management is when those same tools operate as a controlled, observable system inside your actual workflow — where you can see what was done, at what cost, and against what quality standard, before anything reaches a human for final review.”
The distinction sounds subtle, but it’s the difference between a plateau and compounding returns. Teams stuck in adoption mode tend to report the same failure pattern: prompts get copy-pasted into chat channels instead of built into shared documentation, context resets with every new session, and nobody can say what a specific feature actually cost to build — in time or in AI usage — let alone whether the output meets the standard the team would hold a human to.
“The mistake isn’t usually technical,” Kravtsov says. “Teams run multiple AI tools at once without clear rules or quality checks, get inconsistent results, and conclude the technology isn’t ready. In most cases the technology was fine. The process around it wasn’t.”
Rebuilding, not bolting on
The turning point for Kravtsov’s own teams came when they stopped asking “where can we add AI to what we already do” and started asking “what would this process look like if AI were part of the system from the start.”
Applied to software development, that meant restructuring the entire delivery pipeline — from requirements gathering to design, coding, testing, and documentation — around specialized AI agents, each with a narrowly defined role, rather than one general-purpose assistant fielding every kind of request. “An AI agent behaves a lot like a new hire,” Kravtsov says. “Give it a vague, unlimited scope and the output gets unpredictable fast. Give it a specific role with clear boundaries — this is your job, this is where it starts, this is where it ends — and the output quality stabilizes dramatically.”
The result, at ConnectiveOne, was a roughly 300 percent increase in development velocity year over year, measured against the company’s own prior baseline. But Kravtsov is quick to caveat the number. “Velocity by itself isn’t the metric I’d trust,” he says. “What actually tells you the system is sound is whether output quality holds steady as you scale up how much work you’re running through it. Speed is a byproduct of that being true — not something worth celebrating on its own.”
It’s not only an engineering story
Kravtsov is careful to note that this pattern isn’t confined to software teams. Drawing on organizational turnarounds he’s led — including work with an insurance company and a SaaS business — he’s increasingly focused on applying the same logic to non-technical functions: customer support, sales, HR, finance, and internal operations.
“The hidden problem in most companies right now is that they’re still running on the old model — everything held together by people, manual handoffs, and whoever happens to have the most context in their head,” he says. “AI doesn’t just make up for a shortage of developers. It makes up for a shortage of analysts, support staff, and operations people too. But only if you’re willing to redesign the process, not just hand someone a tool and hope it helps.”
That shows up concretely at ConnectiveOne itself, where the platform’s own customer-facing AI agents needed the same kind of structural rethink as the company’s internal dev pipeline. Early on, the team noticed that support and sales conversations reset every time a session ended — the AI effectively forgot who it was talking to. The fix wasn’t a bigger or smarter model; it was building a dedicated knowledge layer that lets the platform’s agents retain context about a customer across sessions, rather than starting cold every time. “That single piece of infrastructure mattered more than any model upgrade we made,” Kravtsov says. “It’s the difference between AI that answers a question and AI that remembers who it’s talking to — and for a support product, that difference is close to everything.”
Where to actually start
For founders and executives who feel like they’ve fallen behind — or worse, tried an “AI transformation” that fizzled — Kravtsov’s advice runs counter to the instinct to move fast and broad.
“Don’t try to roll AI into everything at once,” he says. “Pick one process your team already does constantly, one where success or failure is easy to measure, and where a mistake doesn’t carry serious consequences. Build real structure around automating just that. Prove it works, measure the actual before-and-after, and only then move to the next process.”
It’s a slower, less dramatic story than the “we transformed the whole company” narratives that dominate conference stages — and, in Kravtsov’s experience, a far more reliable one. “I’ve seen too many companies try to do this in one big leap and end up changing nothing, because the process was too chaotic to produce a real result,” he says. “The companies that actually pull ahead aren’t the ones with access to a better model. Right now, access to the technology is roughly the same for everyone. They’re the ones who started rebuilding the system around it earlier than everyone else.”
