The gap between individual AI productivity and team performance

As a product manager at Upsun with a computer engineering background, Kateryna Dvornichenko had spent months researching competing tools in the agentic development space, running tests, comparing features, and building a picture of where the market was heading. She realized the tools were impressive, but something kept standing out.

"Collaboration was not the strong point of any of them," she says. "Everyone stays on their own machine with their own setup."

That observation became the foundation of her thinking about what Upsun Dispatch™ needed to be. More importantly, it pointed to a problem that most conversations about AI productivity in software engineering tend to overlook: making one developer faster doesn't necessarily make the team faster.

The productivity that stays local

The numbers on individual AI productivity show that developers are moving faster, writing more code, and handling more tasks in a day than they could a year ago. The tools have genuinely changed what one person can do in a given hour.

But productivity that stays on one person's machine does not automatically translate to team productivity. The workflows an engineer configures, the prompts they refine, and the context they build up over weeks of working with an agent- all of it lives locally. So when they leave, it goes with them. Meaning that if someone needs to build on their work, they start from scratch.

"People have kind of workflows already on their machines," Dvornichenko explains, "but it's all individual." Basically, the problem isn't that the tools aren't working; it's that the gains aren't accumulating anywhere the team can access and build on.

The bottleneck that follows the speed

As individual productivity increases, the bottleneck can shift from execution to coordination. This gap between individual and team performance shows up first in the review queue: one engineer produces code faster than the team can absorb it, and senior engineers spend their days on work they didn't write and may not have the context to evaluate efficiently.

Then, it impacts the approval process, where a single person becomes the bottleneck for every decision that needs sign-off. And, finally, in cost visibility, where managers are paying for AI usage across the team with no clear picture of what is being spent or whether it is producing results.

"If you're paying a subscription, for example, and you're paying like 200 euros per person," Dvornichenko says, "but you don't know if they're actually using all the tokens that are part of the subscription." The spend is real, while the insight is not.

These are different symptoms of the same problem: AI tools were largely designed around individual users, while the infrastructure needed to turn individual gains into team-level productivity wasn't part of the design.

What team-level AI actually requires

Dvornichenko identifies three things that a genuinely team-level approach to AI needs to provide: trustability, collaboration, and cost transparency. They sound straightforward, but in practice, many AI development tools weren't designed to provide them at the team level.

Trustability means a full, traceable record of what happened. "You do an action in one tool and then in another," she says, "and then the agent works, and you need to see how we ended up with this chunk of code in production, and you don't have much of an idea." Without traceability, you can't build trust in the system. And without trust, teams stay in full supervision mode indefinitely, which defeats the purpose of having agents at all.

Collaboration means anyone on the team can participate in the process, not just the person who set up the workflow on their laptop. "Anyone from the team can approve things," Dvornichenko says. "It's not only you who can approve, and then you're kind of a bottleneck on some actions potentially." 

When workflows live on shared infrastructure rather than personal machines, the whole team can see what is running, contribute to it, and build on what others have done.

Cost transparency means knowing what you are spending before and after a workflow runs. It means management has a single place to understand AI expenditure across the team, not a patchwork of individual subscription receipts and token bills that arrived after the fact.

Together, these capabilities turn AI from something individual engineers use in isolation into something a team can actually operate, oversee, and improve together.

The mindset shift the tools cannot make alone

Another dimension of this problem is one Dvornichenko is careful not to understate. Closing the gap between individual and team performance is not just a tooling problem. It is also a mindset problem, and tools cannot solve these kinds of problems on their own.

Teams need to move from thinking about AI as something individual engineers use in isolation to thinking about it as something the whole organization participates in. As Dvornichenko puts it: "We can accompany this mind shift, but we cannot do this shift alone." 

The tools that helped individual engineers move faster have not yet caught up with the challenge of making teams move together. The next step in AI-assisted development isn't simply making individual engineers faster; it's making that productivity compound at the team level.