By now most of us have heard the Klarna story. The company deployed a hugely successful AI customer-service agent that handled millions of conversations in its first month, across 23 markets and 35 languages, doing the work of roughly 700 full-time agents. Klarna stopped hiring and let its service workforce shrink around it. Then the company realized that while the agent was winning on efficiency and cost, it was damaging customer relationships because it had been optimized for the wrong goal. Klarna started hiring humans again.
The comfortable reading is that AI can’t handle nuance. The more useful reading is that the agent succeeded at exactly what it was given. The agent didn’t fail. The objective did.
And an objective can fail without being wrong. Klarna genuinely wanted lower costs and faster resolutions, and the agent delivered both. What the objective missed was the rest of the picture: the customer relationships that make a fintech business durable never made it into the agent’s goal. The objective wasn’t so much wrong as incomplete, and the agent pursued the incomplete version at a scale no human team could have matched.
An agent’s objective is more than a prompt. A prompt gets written by one person in an afternoon. An objective sits at the end of a chain: an executive team sets a strategic direction, mid-level leaders translate it into projects and workflows, and somewhere in that translation a goal gets encoded for the agent to optimize. Every link in the chain is a judgment call about what the business actually wants. When the direction is stale, or the translation captures only part of the problem, agents don’t surface that. They just execute.
I think this kind of objective failure is becoming a defining risk of the agentic era, for two reasons that are usually discussed separately. Viewed together, they look like a pair of scissors opening: one blade rising, the other falling, and the gap between them widening.
The first blade: demand for collective judgment is rising
Three things are driving demand up at once.
Autonomy windows are lengthening. Current agents run for hours, sometimes days, without human intervention, and the checkpoints have stretched to match. The longer the run, the more work each review has to do: by the time a human looks, the agent has already committed hundreds of small decisions in one direction. The quality of the judgment at that checkpoint, and of the objective behind it, becomes the whole ballgame.
Implementation decisions are multiplying below the executive level. Most decisions about where agents belong get made by mid-level leaders selecting projects and redesigning workflows against the strategic direction the executive team has set; Gartner’s agent-adoption guidance treats that workflow-redesign judgment as preceding every deployment. That division of labor is right. It also raises the stakes on the direction itself, because a hundred mid-level judgment calls inherit whatever clarity, or ambiguity, the executive team produced. Executives don’t need to be hands-on in deployment. They need to set a direction settled enough to delegate against, approve the workstreams, and own the outcomes.
Intent has to be settled before it can be encoded. Intent engineering, the emerging discipline of making organizational purpose machine-readable so agents optimize for what the business actually needs, presupposes a settled purpose to encode. Most organizations don’t have one at the resolution that agents require. You can’t make machine-readable what was never human-agreed.
Of course, judgment isn’t a new skill. What changes in agentic workflows is judgment’s importance: it becomes the control gate, the checkpoint where human intent either enters the system or doesn’t. Agrawal, Gans, and Goldfarb laid out the underlying economics years ago in Prediction Machines (2018) and Power and Prediction (2022): when prediction gets cheap, judgment becomes the valuable complement. Most of the current discourse applies that at the individual level: your judgment, your taste, your ability to direct a model. The organizational layer is nearly silent. And that’s the layer that matters, because one person’s judgment can direct an agent. Only collective judgment can direct an enterprise.
Naming the gap: alignment latency
Here’s a working definition. Alignment latency is the elapsed time between the moment your AI deployment surfaces a strategic question and the moment your executive team produces an answer it actually owns.
Agents operate in minutes and hours. Executive alignment, in most established organizations, operates in weeks and quarters: a calendar of steering committees, offsites, and pre-reads. During the gap, the agents don’t wait, and the mid-level teams selecting projects don’t either. Both keep executing against yesterday’s direction at full speed. Klarna’s alignment latency ran somewhere over a year, and its agent handled millions of conversations during that gap.
The instinctive response is to get faster at aligning. Compress the calendar, tighten the governance loop, stand up an AI council. That’s the right instinct, and it runs straight into the second blade.
The second blade: the supply of shared understanding is degrading
Collective judgment doesn’t come from nowhere. It runs on shared context: a reasonably accurate mutual picture of what colleagues know, intend, and worry about. The evidence is accumulating that AI-mediated work is reducing collaboration and shared context because AI makes it easier to work alone.
The clearest field data comes from an eight-month study of a roughly 200-person U.S. technology company, published by Harvard Business Review in February 2026 (Aruna Ranganathan and Xingqi Maggie Ye). They expected AI to lighten workloads. Instead it intensified them: people worked at a faster pace, absorbed tasks that used to belong to colleagues (product managers writing code, researchers doing engineering), and let work spread into moments that used to be breaks. The tools reduced dependence on others; where the old move was to ask a colleague, the new move was to ask the model. Their role didn’t require these changes. AI made doing more alone feel possible, so people did more, alone.
Read as a productivity story, that’s encouraging. Read as a collaboration story, it’s alarming. The handoffs, cross-checks, and informal corrections that used to happen between people are exactly what got squeezed out, and the organization’s mutual picture of itself got thinner while its output got larger.
There’s a mechanism underneath this, and it now has a research literature and a name: deskilling. The sharpest evidence is clinical. A multicentre study published in The Lancet Gastroenterology & Hepatology (August 2025) found that after a few months of routine AI-assisted colonoscopy, endoscopists’ detection rates when working without the AI fell from 28.4% to 22.4%. Experienced specialists became measurably worse at their core skill after months of delegating it. What we hand to a tool, we stop exercising, and what we stop exercising atrophies. That applies to collaboration as much as to individual expertise. In research on human-AI collective intelligence, a dashboard built to track what a team knows ended up undermining the team’s collective cognition; members offloaded the work of knowing who knows what, and the connective tissue researchers call transactive memory thinned. Collaboration is a set of skills — asking, challenging, cross-checking, knowing who to pull in — and they obey the same rule.
The organizational version has a name too: researchers at Aalto University call it the vicious circle of skill erosion. In their case study of an accounting firm, reliance on automation bred complacency at both the individual and the organizational level. People gradually stopped monitoring the work, maintaining the competence, and assessing the outputs, and nobody noticed until the system was removed and the firm discovered its staff could no longer perform core accounting tasks. That’s the challenge-and-verify habit that collective judgment runs on, dying quietly. Analyses of enterprise AI failure modes flag the same trust miscalibration as a behavioral risk: when a culture treats AI output as authoritative, people stop challenging it. Add shadow AI, with individuals quietly routing work through unapproved tools, and organizational reasoning starts migrating out of shared space entirely.
Demand rising. Supply degrading. That’s the scissors.
What this argument is, and what it isn’t
Two honest caveats.
First, the scissors is a synthesis. No single study demonstrates both blades. I’m connecting an economics literature and a management discourse on one side with field studies and collective-intelligence research on the other. I find the connection compelling, but you should weight it as an argument rather than a finding.
Second, the field evidence is early; the work intensification study is in-progress research. And the erosion mechanisms are behavioral rather than inherent to the technology, which is the good news. Behavioral mechanisms respond to deliberate practice: teams that keep cross-checking, keep challenging output, and keep deciding together keep the capability. Nothing here is inevitable.
The meeting paradox
There’s a twist that makes this harder. We all know the pain of a calendar so full of meetings that it feels like little gets done, and the particular pain of the meeting that could have been an email or a Slack message. Agents and async tooling have destroyed whatever excuse those meetings had, so teams are cutting them.
The danger is cutting judgment-forming coordination with the same knife. Some gatherings just move information, and machines now do that better. Others convert disagreement into alignment: they surface the conflicts between competing priorities across corporate silos and resolve them into something an agent can safely be pointed at. The test is simple. If a meeting ends with information distributed, it was a broadcast. If it ends with a decision somebody owns, it was judgment infrastructure. Cut broadcasts ruthlessly. Strengthen judgment infrastructure deliberately.
The mid-market read
A scope note: this argument is about established organizations — the 300-person SaaS company, the 2,000-person insurer with regulators, legacy systems, and a board — not the small startup or the one-person practice.
For a mid-market leadership team, three moves follow.
Measure your alignment latency. Take the last strategic question your AI activity surfaced: a pilot that had unintended outcomes, a workflow a vendor wants to expand, a quality complaint that implicates an automated step. How long did it take to reach an answer the executive team owns? That number is your exposure window, because while you’re deciding, agents and implementation teams continue to execute.
Audit what work intensification is deleting. Ranganathan and Ye’s prescription is deliberate norms (an “AI practice”) rather than hoping individuals self-regulate. Start with an inventory: where did cross-checking, handoffs, and informal correction actually live in your organization, and do they still happen?
Check the objective for completeness before you scale the agent. An agent’s objective encodes some of the organization’s intents and silently drops the rest. Before scaling, name everything the business actually wants from the workflow, then check the objective against that list. Klarna wanted cost and speed, and got them. Cost and speed just weren’t the whole intent.
I realize I sell alignment work, which is why I encourage you to look at the data rather than take my word for it. Demand for collective judgment is rising no matter what you do. Whether its supply erodes is a choice, which means the width of the gap is a choice too. This is one pair of scissors an organization can close.
Even in an age of AI, human collaboration is still the ultimate competitive advantage.
Sources
In order of appearance.
- Klarna (February 2024) — “Klarna AI assistant handles two-thirds of customer service chats in its first month”. Rollout figures: conversations, markets, languages, 700-agent workload equivalence.
- Bloomberg (May 8, 2025) — “Klarna Turns From AI to Real Person Customer Service” (subscription). CEO Sebastian Siemiatkowski on lower quality and the return to human hiring.
- Nate B. Jones (February 2026) — analysis of the Klarna case as goal misspecification rather than AI failure, and of intent engineering. Video essay; no stable URL.
- Gartner — “Redesign Your Workflows for Strategic Adoption of AI Agents” (licensed research; paraphrased).
- Ajay Agrawal, Joshua Gans, and Avi Goldfarb — Prediction Machines (2018) and Power and Prediction (2022), Harvard Business Review Press.
- Aruna Ranganathan and Xingqi Maggie Ye (February 9, 2026) — “AI Doesn’t Reduce Work—It Intensifies It”, Harvard Business Review. Eight-month field study; described by the authors as in-progress research.
- Krzysztof Budzyń et al. (August 2025) — “Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy”, The Lancet Gastroenterology & Hepatology. The 28.4% → 22.4% detection-rate decline.
- Pranav Gupta, Thuy Ngoc Nguyen, Cleotilde Gonzalez, and Anita Williams Woolley (2025) — “Fostering Collective Intelligence in Human–AI Collaboration: Laying the Groundwork for COHUMAIN”, Topics in Cognitive Science. The team-knowledge-dashboard finding and transactive systems model; on offloading, see also Sparrow et al., Science (2011).
- Tapani Rinta-Kahila, Esko Penttinen, Antti Salovaara, Wael Soliman, and Joona Ruissalo (2023) — “The Vicious Circles of Skill Erosion: A Case Study of Cognitive Automation”, Journal of the Association for Information Systems.
- Dr. Philip Takyi (March 2, 2026) — “The human failure modes of enterprise AI in 2026”, The Business & Financial Times (FinSec Series). Trust-calibration failures and challenge culture.