The Rebuild
What to redesign. What to unlock. Who to put in charge.
This article was co-authored with Oliver Dlouhy, CEO of Kiwi.com
In an earlier piece, we argued that AI is killing the century-old organizational pyramid. The companies that have rebuilt around small, autonomous teams directing networks of agents are operating at productivity levels that have no precedent in the history of business.
Most leaders now accept the diagnosis. The question they are actually asking is harder: how do I get there from here?
The answer is almost never about the technology. The tools are accessible. What separates the 6% of companies that McKinsey’s 2025 State of AI identifies as genuine high performers from the 88% that are using AI but not pulling ahead is organizational: how processes are redesigned, how institutional knowledge is made available to agents, and who is leading the change. That gap is what this article is about.
Redesign, not automation
The default move for most organizations is to take an existing process and hand it to AI. Feed the model the documentation. Ask it to do what the team used to do. It seems logical. It almost never works.
At Kiwi.com, a mid-market European travel technology platform, this was the first mistake. Eighteen months later, revenue is up 35% year on year and a market segment that did not exist when the transformation began now contributes meaningfully to the business. Here is how it happened.
The biggest early mistake was expecting the technology to perform on human terms.
Booking verification is the clearest example. When a booking attempt fails, the resolution process involves confirmation emails, e-ticket numbers, payment records, provider portals, reservation status, airline contacts, and decisions about refunds and cancellations. In practice it is a messy sequence of interdependent checks, each dependent on the one before it.
The first approach was to give the model the existing process documentation and ask it to reach the right final decision. The documentation was thorough, but it had been written for experienced professionals who understood the full flow, remembered the exceptions, and knew instinctively when to move from one step to the next. For an LLM, too much judgment was required at once. The success rate was around 70%. In booking operations, 30% wrong decisions is not automation. It is a different and more expensive problem.
The breakthrough came from redesigning the booking process as a sequence of narrow AI decisions. Instead of asking the model to determine the final outcome, the workflow now guides it step by step: find the confirmation email, check whether an e-ticket number exists, check whether the booking is under 22 hours old, verify all segments, decide whether the airline needs to be contacted. Each step became more reliable because the model only saw the context relevant to that specific decision.
Less context overall, but much more relevant context at each moment. The success rate went from around 70% to around 95% in one to two weeks.
The model did not suddenly become smarter. The work was redesigned for AI. That is the most important practical lesson from everything done at Kiwi. Start with a question about structure, not capability. Not ‘can AI do this?’ but ‘what is the right sequence of narrow decisions for AI to make?’ Getting that redesign right is the actual job.
When people inside Kiwi saw what was possible, they stopped waiting to be told what to automate. They brought their own use cases. Transformation stopped being something that happened to the organization and started being something the organization did to itself.
Nathan, Kiwi’s customer support agent, now consistently outperforms prior performance benchmarks on customer satisfaction by 15 to 20 percent. The work that remains is entirely judgment work: the complicated itinerary, the distressed traveler, the situation no process can anticipate.
Your biggest AI constraint isn’t the technology
The next major transformation at Kiwi was data analytics. Around twenty analysts with full backlogs, writing SQL, building dashboards, supporting decision-makers across the company.
The data existed but getting to it took weeks. By the time analysis arrived, the decision had often already been made.
Through an agent named Stephen, accessible by tagging it on Slack, anyone in the company could get the analysis they needed in real time, 24 hours a day. Decisions that had previously been made on instinct were now made on data. The whole organization got smarter, faster.
Narrow context, better decisions. The right context at the right moment in a narrow sequence outperforms all the context at once.
What made this possible was not the model. It was the work done before the model was involved. The single biggest constraint on how fast any organization can move with AI is context: the institutional knowledge of what the organization is trying to build, how the industry works, who the customers are, and why things are done the way they are. In most organizations, that knowledge lives in human minds rather than in systems. Until you change that, your agents are operating on incomplete information, and their output reflects it.
Making what your organization knows machine-readable is not a technology project. It is a leadership priority that belongs on the CEO’s agenda. The future belongs to companies that treat models as components, and treat orchestration, context, and proprietary knowledge as their true differentiators.
Budget airlines are the clearest example of what this unlocks. For many smaller carriers, maintaining integrations previously made no economic sense. With automation, the system now detects when an airline changes its website, proposes new code, tests it, and creates a merge request for human review. Those engineers are now designing the systems that maintain themselves. Kiwi has a record number of active airline partnerships and a market segment that did not exist eighteen months ago now contributing meaningfully to revenue. The pyramid had not just been expensive. It had been limiting what the business could attempt.
The one thing AI cannot do
The leaders creating disproportionate value in this environment share a specific capability. In the forthcoming book, A-Players: How to Hire the People Who Matter Most in the New World, written by Anish Batlaw, Jessica Neal and Ram Charan, we call this Frame Intelligence: the ability to reframe a problem before anyone else has recognized it as one. To get upstream of the premise. To expose what the rest of the organization is taking for granted and open up possibilities the existing frame made impossible.
At Netflix, they did not try to optimize the DVD business. Before streaming was obvious to anyone, they questioned whether physical media was the right business at all. That is Frame Intelligence. In an AI transformation, it is the difference between asking how to automate what you already do and asking what you would build if you were starting today.
AI cannot do this. It can optimize within a frame with extraordinary speed and accuracy. What it cannot do is decide that the frame is wrong or construct a genuinely new one. That upstream capacity is now the scarcest and most valuable human contribution in any organization.
Most leaders never question the frame they inherit. They optimize within it, measure their performance against it, and describe their success in its terms. The leaders who create disproportionate value ask whether the frame is right in the first place.
Frame Intelligence works alongside two related capabilities that A-Players identifies as the full cognitive engine of leaders navigating this transition well. Adaptive judgment: the discipline to determine what actually matters and act on it at the right moment. And learning velocity: the rate at which a leader moves from new signal to updated mental model to changed behavior. Not whether they learn, but how fast the loop closes. In a world where conditions can invalidate an inherited operating model within months, that speed is now a primary source of competitive advantage.
What the rebuild looks like
Kiwi today looks fundamentally different from the company it was eighteen months ago. Oliver initiated the transformation and continues to shape its direction. The design and implementation work that he has done alongside his team, and what they have built does not have an obvious precedent.
The organizing unit is what Oliver calls a Mission-Aligned Team: a small group of professionals working alongside a network of agents, brought together around a specific outcome and dissolved when the work is done. The agents are not durable entities. They are spun up for a task and killed when it is complete. For a single complex task, the team might spin up dozens of agents with specific instructions, each handling a narrow slice of the problem, and terminate them all when the work is done.
The number of agents deployed does not reflect the amount of work. It reflects the complexity of the context. When context is complex, it is split across more agents to prevent hallucination. When context is simple, a single agent can do the work of dozens of people, with compute as the only constraint. The system is designed so that each agent sees exactly what it needs to see and nothing more.
Some agents run in loops, triggered by specific events, with professionals adjusting instructions only when something changes. Some amplify what an individual human can do. Some simply automate tasks end to end. The relationship between professionals and agents is not a standard management hierarchy. Professionals set direction, define constraints, and intervene at the moments that require genuine judgment. Everything else the system handles itself.
In cases where agents are about to make a potentially costly or irreversible decision, the team deploys a devil’s advocate: an agent whose sole job is to challenge the conclusions of every other agent before the decision is executed. They are not just designing for speed. They are designing for judgment.
The productivity numbers reflect all of this. Merge requests per engineer have nearly doubled year on year. Interestingly, the total number of merge requests across the company fully recovered within four weeks of the restructure. The organization did not slow down during the transition. It accelerated through it.
For a century, the constraint on organizational scale was human attention: you could only coordinate as many people as your management structure could hold together. What the team at Kiwi has built removes that constraint. The ceiling is now compute, and not headcount. The 35% revenue growth and the near doubling of engineering output are early signals of what becomes possible when that ceiling disappears.
There is a deeper implication worth sitting with. Organizational structure, the reporting lines, the functions, the business units, the periodic restructuring programs that consumed years of leadership attention, was never an end in itself. It was a solution to a coordination problem: how do you get large numbers of people working toward a common output without chaos? That solution carried a cost. Structure that reduced coordination cost also reduced adaptability. You could not move one piece without disturbing everything connected to it.
What Oliver and his team have built suggests that when agents handle coordination, structure becomes less load-bearing. The question shifts from who reports to whom to “what is the mission” and “do we have the context” to pursue it.
About the authors
Anish Batlaw is a Managing Director at General Atlantic, where he leads talent strategy across the firm’s global portfolio of growth companies. His forthcoming book, A-Players: How to Hire the People Who Matter Most in the New World, co-authored with Jessica Neal and Ram Charan, will be published in Q4 2026.
Oliver Dlouhy is the founder and CEO of Kiwi.com, a global travel technology platform. Over the past eighteen months he has led one of the most substantive AI transformations of a mid-market company, rebuilding Kiwi around Mission-Aligned Teams and a network of specialized AI agents.
Sources
McKinsey & Company. (2025). The State of AI in 2025: Agents, innovation, and transformation. 1,900+ respondents, 105 countries. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
Nadella, S. (2026, January). Remarks on firm sovereignty and AI. World Economic Forum, Davos.
Batlaw, A., Neal, J., & Charan, R. (2026). A-Players: How to Hire the People Who Matter Most in the New World. Forthcoming Q4 2026.
The views expressed in this piece are solely those of the authors and are provided for informational purposes only. They should not be construed as investment, financial, legal, or other professional advice. Kiwi.com is a portfolio company of General Atlantic. Kiwi is not a representative sample of GA’s current and former investments.