The software industry has sold the same thing for twenty years: finished interfaces for humans. Features were fixed, ready to use, and users clicked through pre-designed paths. Value flowed through the GUI. A different deliverable is taking shape now. It isn’t for people to click, but for AI to orchestrate. Its users aren’t employees at screens, but AI agents performing tasks for them. This shift isn’t just a product manager’s dream. Real corporate environments are already seeing it happen. Klarna’s internal restructuring offers the clearest look at this new reality.
Klarna built an internal knowledge graph with tools like Neo4j over the last two years. They reorganized data scattered across 1,200 SaaS systems into a unified semantic layer. CEO Sebastian Siemiatkowski says data equals knowledge. The goal wasn’t just saving on license fees. It was turning fragmented data into knowledge AI can consume directly.
On top of this layer, Klarna deployed Kiki, an internal AI assistant. Employees use natural language to query and operate on company knowledge. Kiki is already in production. Sebastian also demonstrated, in a 20VC interview, prototypes using Claude built on open source accounting and CRM systems. These show how AI can handle tasks that once required multiple specialized tools. While these remain prototypes, they signal a clear direction.
Engineering facts include the unified knowledge layer and Kiki. Claude-based cross-system agents show the future, not yet a verified production capability.
Klarna is worth analyzing because it’s doing something fundamentally different. The team isn’t building a smarter interface. It’s rebuilding the operational foundation. Systems once only humans could use are being transformed into systems AI can orchestrate.
This new deliverable has three layers.
The first is the hard foundation. This includes payment clearing, risk models, merchant networks, and core functions like records, permissions, transactions, and audits. These can’t be hallucinated. They are the foundation.
The second is the knowledge layer. Klarna’s Neo4j graph encodes business semantics, organizational structures, and rules into a format AI can use. Without this, AI faces fragmented data and can’t do much.
The third is the AI operation layer. Kiki and the Claude prototypes turn complex operations that once required switching between GUIs into high-level actions AI can trigger in one step.
I call this a Generative Kernel. Instead of delivering finished furniture, you deliver core parts, assembly instructions, and tools. AI then assembles and operates them as needed. This pattern is already appearing in AI programming. If Stripe redesigned its developer experience for the AI era, it would do something similar. It would provide the hard base of payment APIs, a knowledge system for AI like best practices and edge cases, and tools to turn complex tasks into reliable calls. Klarna targets internal agents while Stripe targets external ones, but the pattern is the same. This model is spreading from dev tools to enterprise software.
A reassignment of roles is a better description for the fate of the GUI. In traditional software, the GUI handles three things: entry for operations, information display, and process constraints. When the operator shifts from human to agent, these functions change.
Entry points in the GUI weaken significantly. Agents use APIs, tool calls, and code. Buttons and forms are slow and imprecise for them. Information display remains but changes purpose. It’s no longer the operator’s main view. It becomes an observation window for supervisors. Humans use it to review results, find anomalies, and handle exceptions. Process constraints move to natural language. Instead of interface design preventing errors, we use skill boundaries, permission sets, and agent contracts.
The GUI goes from being the only interface to a window for watching agents interact with software. It still has value, but it no longer owns the control plane. That plane is moving toward natural language, agents, and tool calls.
This matches our experience in AI development. When using Claude Code or Cursor, you don’t click IDE menus often. You describe intent in natural language. AI calls the compiler, test framework, and version control. The IDE’s GUI is still there, but it’s for previewing results and auditing errors. Klarna is bringing this same shift to enterprise software.
When agents take over, the type of reliability we need changes.
Traditional SaaS provides process determinism. Users click buttons and fill forms in a precise order. This works because humans need guidance and predictable feedback.
Agents care about result determinism. Given a goal and constraints, can the system stably deliver the right outcome? Specific paths or the number of pages visited don’t matter to the agent. It needs clear input-output contracts, verifiable standards, and feedback on errors. This doesn’t mean everything becomes result-based. The hard base still needs highly controlled processes. Agents handle the querying, orchestration, and cross-system operations on top of that base.
This was the core design choice in Klarna’s rebuild. When Kiki executes a cross-departmental process, the specific clicks don’t matter. What matters is whether it met the goal within constraints and if it can self-correct or ask for help when it drifts.
Combining Klarna’s practice with AI development experience reveals several insights for builders.
First, a unified knowledge layer is a requirement, not an option. Klarna spent massive effort on data unification because AI can’t act on fragmented data. Any team wanting AI agents to be effective must first ensure their data and knowledge are in a state agents can consume.
Second, design standards are changing. If you’re building a product, ask if your core capabilities can be called by an agent directly. Is your business knowledge encoded for AI, or does it only exist in manuals and UI design? Do you provide tools that let AI reach deterministic results?
Third, GUIs won’t vanish, but relying only on them is dangerous. If a product’s value is tied entirely to the interface, that value will be bypassed when agents become the primary operators. Lasting value lies in the core capabilities and the knowledge layer. Interfaces are just one channel, and they are becoming less central.
Fourth, this isn’t vibe coding. Vibe coding says anyone can write code with AI. Klarna’s work is deeper. It rebuilt how enterprise data is organized so AI can perform complex operations. One lowers the barrier to generating code. The other changes the architecture of enterprise computing. Both move in the same direction: operators shift from humans to agents, and deliverables shift from finished GUIs to generative kernels.
This story has gathered some noise. A 20VC interview with Klarna’s CEO was titled “SaaS is Dead.” This led to claims that AI is replacing SaaS or that it’s just a CEO telling stories. Sebastian later clarified his stance. In TechCrunch, he said they did not replace SaaS with an LLM. He even suggested it might not be the end of Salesforce, but the opposite. In a TIME interview, he admitted AI works for simple tasks while human support remains a premium service.
These clarifications matter. “AI replacing Salesforce” is a better headline than “rebuilding knowledge graphs with Neo4j,” but it hides the real work. Klarna’s core move wasn’t replacing CRM functions with an LLM. It was unifying fragmented data into a knowledge layer for AI to use. Without that first step, the second doesn’t work.
Sebastian’s point about switching costs is also vital. He makes versions of this argument in both the 20VC interview and the later TechCrunch clarification: the real blow comes when AI solves the switching cost of data. When agents can pull data from old systems and plug it into a new knowledge layer, the moats built on data lock-in weaken. This is a powerful idea, but it depends on building that knowledge layer and achieving result determinism in complex workflows. These conditions aren’t yet proven at scale.
The transition is still in its early stages. We don’t know how many of the Claude-based prototypes will make it to production. The cost of maintaining a unified knowledge layer and the reliability of agents in complex tasks are still open questions.
A clear direction has emerged. As AI agents become the primary operators of software, design will shift from being human-friendly to being agent-orchestratable. For builders, the most important thing to track isn’t headlines about AI replacing SaaS. The deliverable itself is being redefined. We are moving from finished interfaces for humans to generative kernels for agents: capabilities, knowledge, and verifiable feedback loops.