And most companies don't know they're in it.
In January 2025, Klarna announced its AI agent was doing the work of 853 full-time employees and had saved the company $60 million.
By mid-2025, Klarna was frantically rehiring the human agents it had fired.
This is not a story about AI failing. It's a story about AI succeeding at exactly the wrong thing — and the distinction between those two outcomes is the most important unsolved problem in enterprise AI right now.
The Intelligence Race Is Over
For the past three years, the AI conversation has been framed as an intelligence race. Who has the best model? Who tops the benchmarks? Who has the biggest context window?
That framing made sense when models were the bottleneck. Models are not the bottleneck anymore.
Claude Opus, Gemini, GPT — these are all extraordinarily capable systems. The differences between them matter far less than most people think. What matters enormously is what you give them to want.
The race has shifted. It is no longer about who has the smartest AI. It is about who has built the organizational infrastructure that lets AI operate with the fullest, most accurate, most strategically correct understanding of what the organization is actually trying to accomplish.
This is the intent race. And most companies don't know they're in it.
What Klarna Actually Got Wrong
Let's be precise about the Klarna story, because the popular reading is wrong.
Klarna's AI agent handled 2.3 million conversations in its first month. Across 23 markets, 35 languages. Resolution times dropped from 11 minutes to 2. The CEO projected $40 million in savings.
The popular reading: AI can't handle nuance. The AI failed.
The accurate reading: The AI was extraordinarily good at exactly what it was asked to do. That was the problem.
Klarna gave its agent the goal: resolve tickets fast.
Klarna's actual organizational goal was: build lasting customer relationships that drive lifetime value in a competitive fintech market.
Those are profoundly different goals. They require profoundly different decisions at every point of customer interaction.
A human agent with five years at Klarna knew the difference intuitively. She knew when to bend a policy. She knew when to spend three extra minutes because the customer's tone said they were about to churn. She knew when efficiency was the right move and when generosity was the right move.
She knew this because she absorbed Klarna's real values — not the ones on the website, but the ones encoded in daily decisions, in the stories veterans tell new hires, in the unwritten rules about which metrics leadership actually cares about when push comes to shove.
The AI agent knew none of it. It had a prompt. It had context. It did not have intent.
When Klarna fired 700 human agents, those agents took with them the institutional knowledge that had never been documented. The knowledge that everyone just knew. The $60 million in savings wasn't nearly enough to cover the reputational damage from becoming the public face of AI gone wrong. Klarna spent the rest of 2025 trying to buy back what it had discarded.
The Microsoft Copilot Parallel
Klarna isn't an isolated case. Look at what happened with Microsoft Copilot.
85% of Fortune 500 companies adopted it. Only 5% moved from pilot to larger-scale deployment. Only about 3% of the total Microsoft 365 user base became paid users. Bloomberg reported Microsoft slashing internal sales targets after the majority of salespeople missed their goals.
Inside companies with six-figure Copilot deals, employees resisted. They preferred ChatGPT. They preferred Claude. They preferred anything that felt like it understood what they were actually trying to do.
The standard explanation centers on UX problems and model quality. Those are real. But they're not the fundamental issue.
The fundamental issue is this: deploying an AI tool across an organization without organizational intent alignment is like hiring 40,000 new employees and never telling them what the company does, what it values, or how to make decisions. You get lots of activity. You get AI usage metrics on a dashboard. You get almost no measurable impact on what the organization is actually trying to accomplish.
That's not a tools problem. That's an intent gap.
The Numbers That Should Terrify You
The investment in enterprise AI is real and accelerating. 57% of companies are putting between 21 and 50% of their digital transformation budgets into AI automation. 20% are investing over half.
And yet: 74% of companies globally report they have yet to see tangible value from AI. McKinsey found 30% of AI pilots failed to achieve scaled impact. Deloitte's 2026 State of AI report found that 84% of companies have not redesigned jobs around AI capabilities, and only 21% have a mature model for agent governance.
Massive investment. Widespread deployment. Mixed results.
These numbers aren't a contradiction once you understand the underlying problem. Organizations have solved "can AI do this task?" They have completely failed to solve "can AI do this task in a way that serves our organizational goals at scale with appropriate judgment?"
That second question is an intent engineering question.
Three Disciplines, Three Eras
To understand where we are, you have to understand how we got here.
Prompt Engineering was the first discipline of the AI age. Individual, synchronous, session-based. You sit at a chat window, craft an instruction, iterate the output. Personal skill, personal value. This era produced a thousand "how to write the perfect prompt" blog posts. Most of them were terrible.
Context Engineering is the discipline the industry is currently grappling with. Anthropic defined it in September 2025 as the shift from crafting isolated instructions to crafting the entire information state that an AI system operates within. LangChain's Harrison Chase put it bluntly: "Everything's context engineering." Building RAG pipelines, wiring MCP servers, structuring organizational knowledge so agents can access it. This is where the action is right now.
Context engineering is necessary. It is not sufficient.
Intent Engineering is the third discipline, and almost nobody is building for it yet.
Context tells agents what to know. Intent tells agents what to want.
Intent engineering is the practice of encoding organizational purpose into infrastructure — not as prose in a system prompt, but as structured, actionable parameters that shape how agents make decisions autonomously. It is the layer that would have told Klarna's AI agent: yes, you can resolve this ticket in 90 seconds, but this customer has been with us for three years and their tone indicates frustration. Spend the extra time. Offer them a specialist. The goal is retention.
Without intent engineering, you get what Klarna got: a technically brilliant agent optimizing for exactly the wrong objective.
The Three Layers of the Intent Gap
The intent gap operates across three distinct layers. Getting any one of them right is helpful. Getting all three right is the difference between having AI tools and having an AI-native organization.
Layer 1: Unified Context Infrastructure. Right now, every team building agents rolls their own context stack. One team pipes Slack data through a custom RAG pipeline. Another manually exports Google Docs into a vector store. A third built an MCP server connecting to Salesforce but not Jira. A fourth doesn't know the other three exist.
This is the shadow agents problem, and it mirrors the shadow IT crisis of the early cloud era — except the stakes are higher because agents don't just access data, they act on it. The Model Context Protocol, which Anthropic introduced in late 2024 and donated to the Linux Foundation in December 2025, is the most promising standardization attempt. But protocol adoption and organizational implementation are very different things. Having a standard doesn't help if you haven't decided which ports to install, who maintains them, or what gets plugged in.
Layer 2: Coherent AI Worker Toolkit. One person uses Claude for research and ChatGPT for drafting. Another uses Cursor for code and Perplexity for fact-checking. A third built a custom agent chain. A fourth is copy-pasting into a chat window. None of these workflows are transferable, measurable, or improvable by anyone else.
The difference between individual AI use and organizational AI leverage is enormous. It's the difference between one good hire and a system that makes everybody better. Organizations are giving people tools without giving their agents the organizational context and data that let those tools deliver real value. Tools deployed without organizational infrastructure become very expensive toys.
Layer 3: Intent Engineering Proper. This is the layer that almost certainly doesn't exist in your business. And it requires something genuinely new.
OKRs were designed for people. They encode human-readable goals. They assume a manager can look a direct report in the eye and say "here's what matters this quarter" and trust that the report will interpret that guidance through a mesh of institutional context, professional norms, and personal judgment developed over months and years.
Agents have none of that. An agent does not absorb your company culture through osmosis at all-hands meetings and hallway conversations and happy hours. When a human employee joins a company, alignment happens through a hundred informal mechanisms. None of that works for agents. Agents need explicit alignment, and they need it before they start working, not six months after.
This means organizations need to develop something that mostly doesn't exist: machine-readable expressions of organizational intent.
That is not "put the OKRs in the prompt." It is a cascade of specificity that most organizations have never had to produce because humans could fill in the gaps. It requires goal structures that are agent-actionable, not human-aspirational. It requires delegation frameworks that decompose principles into decision boundaries. It requires feedback mechanisms that measure alignment drift, not just task completion.
Why This Hasn't Been Built
Three reasons, and they're all real.
First, it's genuinely new. Before agents could run autonomously over long time horizons, the human was the intent layer. Long-running agents break that model. We have agents that run for weeks now. We will soon have agents running for months.
Second, it's a two-cultures problem. The people who understand organizational strategy — executives, department heads — are not the people who build agents. The people who build agents are not the people who understand organizational strategy. MIT found that AI investment is still viewed primarily as a tech challenge for the CIO rather than a business issue requiring leadership across the organization. CIOs can build infrastructure. Intent comes from the entire leadership team.
Third, it's genuinely hard. Making organizational intent explicit and structured requires something most organizations have never done. Goals live in slide decks and OKR documents that get half-read and referenced at performance reviews once a year. They live in leadership principles that get cited but never operationalized. They live in the tacit knowledge of experienced employees who know what to do in ambiguous situations even though they've never been told. Nobody has strong muscles here because most organizations have never had to exercise them.
What a Solution Looks Like
The solution has three components and they build on each other.
At the infrastructure level: A composable, vendor-agnostic architecture that enables agents to operate across systems, tools, and models securely and at scale. MCP provides the protocol layer. But the organizational implementation requires decisions about data governance, access controls, freshness guarantees, and semantic consistency that no protocol makes for you. The companies that get this right will treat it as a core strategic investment, not an IT project.
At the workflow level: A shared, living map of which workflows are agent-ready, which are agent-augmented with human in the loop, and which remain human-only. Not a static document filed in Confluence and forgotten. An operating system that evolves as agent capabilities improve and organizational context infrastructure matures. Organizations that do this well are likely creating a new role: something like an AI Workflow Architect, sitting between engineering, operations, and strategy.
At the alignment level: The genuinely new thing. Goal translation infrastructure that converts human-readable organizational objectives into agent-actionable parameters. This includes decision boundaries, escalation logic, value hierarchies — how the agent resolves tradeoffs — and feedback loops that measure and correct alignment drift over time.
The Architecture That Makes This Real
Here's what this looks like when you actually build it.
Intent cascades down through three layers. Conflicts escalate up. No child layer can contradict its parent — it can only specialize, refine, or constrain further.
GLOBAL INTENT LAYER
Corporate values, non-negotiables, top-level tradeoffs
No override. This is the constitution.
└── DEPARTMENT INTENT LAYER
Team goals, departmental tradeoffs, domain priorities
Can specialize global values. Cannot contradict them.
└── AGENT INTENT LAYER
Role-specific behavior, authority, escalation logic
Can specialize department intent. Cannot contradict it.
The rule is simple and it matters: child layers can make constraints stricter. They cannot make them looser. A department can say "escalate sooner than global requires." An agent can say "escalate sooner than the department requires." Neither can say "escalate later."
Every agent boots up with an Intent Configuration — a typed, versioned, auditable object that tells it not just what to do but what to want. Not prose. A structured schema with explicit tradeoff hierarchies, escalation triggers, authority boundaries, success definitions, and feedback requirements.
Before any intent configuration touches a real user, it passes through a synthetic validation sandbox. Think of it as TDD for agent behavior — test-driven intent engineering. You run the configuration against a battery of injected scenarios: a frustrated longtime customer, an enterprise client requesting a policy exception, a customer whose sentiment tanks mid-conversation. If the config handles them correctly according to its declared tradeoff hierarchy, it ships. If it fails any scenario, the system identifies exactly which rule produced the wrong outcome. The sandbox is a hard deployment gate, not just a report.
And every decision an agent makes carries a traceable audit chain back to the specific version of the specific rule that produced it. Not an alignment score. Component scores: objective adherence, boundary compliance, escalation correctness, outcome quality. Every decision is traceable to config lineage. This is compliance-ready by default — something regulators will demand by 2027 at the latest.
The Management Innovation of This Era
If OKRs were the management innovation that let Intel align thousands of humans to shared objectives in the 1970s, intent engineering is the management innovation that lets organizations align hundreds — or thousands, or tens of thousands — of agents to those same objectives in 2026.
While those agents operate at speeds and scales no human manager can supervise.
We do not have twenty years for this to become standard practice. The urgency is different.
The most important AI investment in 2026 is not a model subscription. It is not another Copilot license. It is organizational intent architecture — making your company's goals, values, decision frameworks, and tradeoff hierarchies discoverable, structured, and agent-actionable.
The company with a mediocre model and extraordinary organizational intent infrastructure will outperform the company with a frontier model and fragmented, unaligned organizational knowledge every single time.
The Lesson Klarna Learned the Hard Way
Klarna's story was not "AI doesn't work." The AI worked brilliantly. That was the problem. It was so good at optimizing for the measurable objective that nobody noticed it was destroying the ones that really mattered — trust, relationship quality, brand integrity.
The 700 human agents who were let go took with them the institutional knowledge that had never been documented. The knowledge that everyone just knew. The age of humans just know is ending.
Intent engineering is the discipline of making what humans know explicit, structured, and machine-actionable. Not because the humans are leaving — although some of them will — but because the agents arriving to work alongside them cannot function without it.
Context without intent is like a loaded weapon with no target. We have spent years building AI systems.
2026 is the year we learn to aim them.
Jason Brashear is a senior software developer and AI systems architect with 30 years of experience building production systems. He is the creator of ArgentOS, an intent-native multi-agent operating system, and a partner at Titanium Computing. He writes about the intersection of AI architecture, organizational design, and the future of agentic systems.
Follow him on GitHub: webdevtodayjason
Top comments (0)