The ones that keep coming back to the same meeting were never yours to solve. In the age of AI agents, tensions move to the core of the job.
A company handed two-thirds of its customer service to AI, shed the equivalent of 700 support agents, and called it solved. A year later, the CEO said the words every leader dreads: We went too far. It was not an AI mistake. It was a category mistake, the same one that breaks reorgs, rewrites, and roadmaps. Many of your hardest calls were never problems with answers. These were tensions to be managed, and AI agents just turned spotting the difference into a survival skill.
This is the first article in a short series on three cognitive lenses for thinking under pressure: tension, connection, and reduction.
In early 2024, Klarna launched an AI assistant built with OpenAI and, within its first month, said the bot was handling two-thirds of all support conversations and doing the work of 700 agents. From 2022 to 2024, the support headcount fell by roughly that many due to a hiring freeze and attrition. Efficiency was solved, and the company said so loudly. Then the quality complaints arrived, customer trust slipped, and by May 2025, CEO Sebastian Siemiatkowski admitted the company had gone too far. Klarna started rehiring humans into a blended model in which AI handles volume and people handle judgment.
Read quickly, and this looks like a story about AI falling short. It is not. The company did not pick the wrong tool; it misclassified the question. “How much of this work should AI carry, and how much should humans?” has no permanent answer. It moves with the model, the customer, and the quarter. Treat a moving question as settled, and the bill always comes due.
A problem has a solution. You diagnose it, fix it, and it stays fixed. A late invoice run, a broken deployment pipeline, a gap in the sales-coverage map: real problems, and the satisfying kind, because once handled, they leave. A tension is different in kind, not a different degree. It is a pair of values, each with a real claim, that never stop needing each other. Speed and quality. Centralization and decentralization. Purpose and profit. Human judgment and machine scale. You do not get to pick one and be done, because the moment you starve either side, it comes back wearing new symptoms.
Here is the loop almost every leadership team runs without naming it. One value feels under-supplied this quarter, so the team swings hard toward it and gets applause for two or three quarters. Then the other value, the one quietly starved by the swing, starts surfacing as a fresh batch of complaints. The instinct is to swing back, and the cycle restarts from the opposite pole. Smart teams can run this for years, each swing dressed up as a bold new initiative, before anyone notices it is one tension oscillating rather than a sequence of problems being solved. Rehiring the support team is simply the backswing. It would have been far cheaper to name the tension before the first one.
What is new is the speed. Every team is now dividing labor between humans and agents, and that division is a tension, not a setting you configure once. Agent autonomy against human oversight. Velocity against control. These used to be annual offsite questions; they now sit within daily operations, and the bill for getting them wrong arrives faster than the year it took to surface above. The right setting moves with each new model, and a moving optimum is the signature of a tension. Anyone selling a permanent answer to “how much should we let AI run” is selling the next overshoot.
None of this is new thinking, only newly urgent. A consultant named Barry Johnson founded the discipline in 1975 and published the canonical book, Polarity Management: Identifying and Managing Unsolvable Problems, in 1992. (Thanks, Aram, for pointing me in that direction.) Unsolvable is the load-bearing word. Jim Collins gives the same idea a more memorable name. In Beyond Entrepreneurship 2.0, he contrasts the “tyranny of the or,” which forces every choice into A or B, with the “genius of the and.” Disciplined thinkers, in his research, are comfortable holding both.
The strongest reason to trust this is that serious people keep arriving at the same place through different doors and naming it differently. Johnson calls it a polarity. Collins calls it the genius of the and. On The Curiosity Shop podcast this year, Brené Brown and Adam Grant call it a paradox, and Grant’s plain definition is the one worth keeping: “two opposites coexisting,” which are “often the source of our best ideas, our most important decisions, but they can drive us crazy too.” That last clause is the part the tidier names leave out. The discipline is the willingness to sit with something that drives you a little crazy rather than resolve it for relief.
When a team starts treating a tension as a tension, three things change, and none of them is soft. The conversation stops being a debate to win, because permanently winning either side is the failure mode. The job becomes describing both poles fairly, including the exact cost each one imposes when it dominates. That sounds like a group hug. It is closer to the opposite, because it denies everyone the relief of being declared right.
The team also starts watching for warning signs that hide in plain sight as recurring complaints. Falling satisfaction scores are a leading indicator that the human pole has been starved, but a team that never sets the tension up to be watched reads the signal as a CEO mea culpa rather than a dashboard. Name a tension, and its two failure signatures, and the next overshoot announces itself while a cheap correction is still on the table. With agents in the loop, that early warning matters more because the swing comes faster.
And the work changes shape. The job is not to resolve the tension; it is to notice when one pole has overshot and pull deliberately back. Small, recurring corrections, not heroics. It looks less like a turnaround and more like a thermostat, because a dramatic announcement is itself an overshoot.
For anyone who operates or invests in software, this is not abstract. The companies that compound through a full ownership cycle are rarely the ones that found the perfect operating model; they are the ones that kept adjusting as conditions moved. Growth against efficiency. Founder velocity against institutional discipline. Now, human craft against AI agent scale. Each breaks the same way when a team decides it has finally been solved.
AI sharpens that risk to a point. The human-and-agent balance you set at the start of a hold is not the balance that should exist at exit, because the technology underneath it compounds across the very three-to-five-year window in which you are building the business to sell. Set it once in year one and defend it; by year three, you are optimizing for a capability frontier that no longer exists. The buyer is underwriting the business as it will run then, with the agents that exist then. The teams that keep retuning the line as the models improve reach the table with a company that is structurally more valuable than the teams that solved it once. The classic stall was defending last cycle’s answer. The new one is defending an answer that AI’s own rate of change has already obsoleted underneath you.
So here is the move, and it works for a founder with 7 people and a chief product officer with 700. The artifacts compress as the scale shrinks; the discipline does not. Pick the loudest tension on your team, the one dragging the same argument into the same meeting. A good 2026 candidate: how much to let the agents run versus how much a human signs off. On a whiteboard, write the genuine upside of each pole in one sentence. Under each, write the specific cost that pole inflicts when pushed too far. Then ask the room one question: which of those two costs are we paying right now? The disagreement stops being about who is right and becomes about where on the map you currently sit, which is a question you can actually act on.
Twenty honest minutes tend to produce more clarity than the last three meetings on the same topic combined. Run it as a recurring move, and it changes something deeper than any single decision: it changes how the team handles disagreement itself. Once you can hold a tension rather than solve it, the next question is where the leverage sits to govern the swing. That is a question about systems, and it is where this series goes next.
Pick a side to lean today. Then keep the other one alive on the team, because in the age of AI agents, you will need it back faster than you think.
Originally published at orion.beehiiv.com.
Sources
Barry Johnson, Polarity Management: Identifying and Managing Unsolvable Problems, 2nd ed., HRD Press, 1992. ISBN 9780874251760.
Jim Collins and William Lazier, Beyond Entrepreneurship 2.0, Portfolio, 2020. ISBN 9780399564239.
Fortune, “Klarna plans to hire humans again, as new landmark survey reveals most AI projects fail to deliver,” 2025-05-09. https://fortune.com/2025/05/09/klarna-ai-humans-return-on-investment/ Accessed 2026-06-10.
Entrepreneur, “Klarna Is Hiring Customer Service Agents After AI Couldn’t Cut It on Calls, According to the Company’s CEO.” https://www.entrepreneur.com/business-news/klarna-ceo-reverses-course-by-hiring-more-humans-not-ai/491396 Accessed 2026-06-10.
The Curiosity Shop with Brené Brown and Adam Grant, “Exploring the Paradoxes of Human Nature,” 2026-05-28. https://www.youtube.com/watch?v=5tVqjcbs3io Accessed 2026-06-10.