Try to Automate Your Entire Job. What's Left Is Why You're Valuable.

Klarna was the poster child. The Swedish fintech replaced roughly 700 customer service agents with AI chatbots, cut its workforce by 24%, and CEO Sebastian Siemiatkowski told anyone who would listen that "AI can already do all of the jobs that we as humans can do." Resolution times improved, labor costs dropped, and the boardroom loved every slide in the quarterly review.

Then the customers started leaving, because the bots handled straightforward questions but couldn't handle nuance, refunds that required judgment, or the loyalty dynamics that keep a fintech's margins alive. By early 2025, Klarna reversed course and started rehiring humans. Siemiatkowski conceded: "It's so critical that you are clear to your customer that there will always be a human if you want."

Klarna's mistake was not using AI; it was that the people who decided what to automate had never worked a customer service shift, had never navigated the emotional terrain of a frustrated customer threatening to close their account, and had no way to know what they were losing until it was gone.

The 74% Problem

BCG's October 2024 survey found that 74% of companies struggle to extract any scaled value from their AI investments: not marginal value, not disappointing-but-measurable value, but nothing at all beyond isolated pilots that never graduate to production. McKinsey, the firm most companies hire to fix this, has its own confession: 70% of large-scale transformations fail.

The failure mode is always the same, predictable enough to script in advance. Consultants interview employees for a week, build a slide deck mapping job functions to AI capabilities, and leave. The slide deck gathers dust because the people who built it never actually did the work they were proposing to automate, and the employees who do that work every day were never asked which parts require judgment and institutional memory that no process document captures.

Michael Polanyi named the underlying epistemological barrier in 1966 when he wrote: "We can know more than we can tell." A machinist knows when a lathe sounds wrong before any sensor detects a deviation. A nurse knows a patient is declining from micro-expressions that no clinical decision rule captures. David Autor formalized it as Polanyi's Paradox in 2014: if you cannot articulate what you know, you cannot program a machine to replicate it. And the person least equipped to articulate your tacit knowledge is someone who has never done your job.

The Jagged Frontier

Harvard Business School's 2023 study of 758 BCG consultants using GPT-4 revealed what they called a "jagged technological frontier." On tasks inside the AI's competence boundary, consultants improved by 40%, but on tasks outside it, performance dropped 23% because the AI didn't just fail to help with hard problems but actively degraded the work of experienced professionals by anchoring them to confident, wrong outputs that felt authoritative enough to override their own judgment.

The frontier is jagged, which means it cannot be drawn on a whiteboard by anyone, including the people who built the models. No one can predict in advance which of your specific tasks fall on which side. The METR study makes this visceral: experienced developers were 19% slower with AI assistance and believed they were 20% faster, producing a 39-percentage-point perception gap that should terrify anyone relying on self-reported AI productivity gains. If even the people using the tools misjudge this severely, imagine how far off an outside consultant's assessment will be.

The only way to map the frontier for your job is to walk along it yourself, task by task, and see where the ground holds.

The Math

Consider a mid-size company with 1,000 employees evaluating AI augmentation. Two paths.

Path A: Hire consultants. A typical MBB engagement runs 10 to 14 weeks with 4 to 6 consultants at published day rates of $6,500 to $12,000. Total: $3 million to $5 million for a PowerPoint deck identifying automation candidates across 15 to 30 job functions, with a timeline to first actionable change of 4 to 6 months and a success rate for scaled value of just 26%.

Path B: Give every employee AI tools and the mandate to experiment. A Copilot, ChatGPT Team, or Claude Pro license costs $50/month per employee, totaling $600,000 per year for 1,000 people, and the timeline to a first experiment is measured in hours, not months. Each employee runs dozens of micro-experiments against their own tasks, discovering what AI handles well and what it butchers, and no slide deck is required because the knowledge stays in the organization when the experimenter is the organization.

Path B costs 5x to 8x less per year than a single Path A engagement, but the real gap is informational, not financial. A consultant interviews you for 45 minutes and decides that 30% of your job is "LLM-addressable." You, after two weeks of actually trying, know that the 30% is really 12%, and a different 18% you never would have guessed turns out to be trivially automatable. The consultant's map diverges from the territory the moment it's drawn.

Why This Is Offense, Not Defense

The instinct is to frame self-automation as survival strategy, but that framing is wrong, and if you accept it, you limit yourself to playing a game where the best possible outcome is keeping what you already have.

The person who automates 40% of their routine work doesn't lose 40% of their role; they gain 40% of their capacity back, and that changes everything. The hard problems that have been sitting in the backlog for months, the analysis nobody had bandwidth for, the cross-functional project that needed someone with deep context and available hours: that's where the reclaimed time goes. Self-automation is not about protecting your current job description. It is about expanding into work that was previously impossible given your bandwidth constraints.

Shopify CEO Tobi Lütke understood this when he declared in an April 2025 memo: "Reflexive AI usage is now a baseline expectation." Teams must demonstrate that a job cannot be done by AI before requesting new headcount. note what he did not do: he did not hire McKinsey, he did not create a top-down AI Transformation Office, and he did not ask a committee to study the question for six months. He pushed the automation mandate to the people who actually do the work.

When thousands of employees simultaneously run micro-experiments against their own tasks, the organization collectively maps the AI frontier faster than any centralized analysis could. Each person contributes one data point about their own domain, and together they build a map that no consulting firm, regardless of price, has the domain expertise to construct. This is the open-source model applied to organizational knowledge: distributed contributors, each with unique context, producing a collective artifact no single author could match.

The Organizational Design Problem

Individual initiative alone is insufficient. An MIT Technology Review and Infosys study found that 83% of business leaders say psychological safety directly impacts the success of AI initiatives. If employees fear that discovering automatable tasks means discovering their own redundancy, the rational move is to hide the results. Game theory wins over good intentions every time.

Organizations that want self-automation to work need three structural elements. First, reward discovery, not just productivity. The employee who maps the frontier and shares what they learned is producing organizational intelligence worth more than the efficiency gain itself. Second, redefine roles rather than eliminating them. When someone automates report generation, their job becomes the analysis that the reports were supposed to inform, the work that was always backlogged because everyone was too busy formatting spreadsheets. Third, make knowledge-sharing mandatory, not optional. An employee who experiments in isolation produces individual optimization. An employee who documents and publishes their findings produces organizational learning. The difference between these two outcomes is the difference between 1,000 disconnected experiments and a continuously updated map of what AI can and cannot do across every function in the company.

Managers are the critical variable in whether any of this works. They create the conditions where experimentation is safe, ensure findings are shared across teams, and redesign roles as the frontier shifts. Without management buy-in, self-automation becomes a personal hobby practiced by the most proactive 10%, and those people were going to figure it out anyway.

Limitations

The cost comparison uses midpoint estimates for MBB fees, which vary widely by engagement scope and are not publicly disclosed. The 26% "scaled value" rate inverts BCG's 74% figure, which conflates "failed entirely" with "succeeded but couldn't scale." The Klarna case is a single company, and other top-down efforts may have succeeded quietly. The $50/month per-employee license cost understates real expenses when including integration, training overhead, and the opportunity cost of employee time spent experimenting. Most critically, this argument requires workplace cultures that support psychological safety around AI experimentation, and most do not have it yet.

The Strongest Counterargument

Self-automation asks turkeys to vote for Thanksgiving. A rational employee who discovers that 60% of their job is automatable has every incentive to hide that finding. This is not hypothetical, because it is exactly what happened during the early spreadsheet era, when clerks who could do in an hour what used to take a day simply pretended it still took a day and used the remaining seven hours however they pleased.

The counterargument: AI capabilities are visible, public, and advancing on a monthly cadence that makes concealment temporary at best. Management already knows automation is possible, and the only question is how much and where. The employee who gets ahead of it controls the narrative, retains the irreducible core of their role, and positions themselves as the organizational expert on the human-AI boundary in their domain. The one who hides gets restructured when a consultant draws the wrong map and no one with actual knowledge is left to correct it. But this counterargument only holds if the organization has built the structural incentives described above. Without them, the turkeys are right to be cautious.

The Playbook

Start today, not next week. Pick the most tedious thing you did yesterday and try it with AI right now, because you don't need a framework or a planning phase to run your first experiment.

Test aggressively. Over the next two to four weeks, take your ten most time-consuming recurring tasks and try each one with AI. Perform the task both ways, compare the output side by side, and record what worked, what required editing, and what was unsalvageable.

Map the frontier. Tasks where AI saved time go into the "augment" bucket. Tasks where it failed or actively degraded the work go into "human core." That human core, backed by evidence from your own experiments, is your value proposition in any restructuring conversation.

Recalibrate continuously. New model releases happen quarterly and each one shifts the frontier in ways no one can predict from the changelog alone, which means the map you drew last month is already slightly wrong and the only way to maintain its accuracy is to keep walking the boundary.

Document everything. Share it. This is not optional. Publish your findings internally: "I tested AI against 10 of my core tasks. Here is what it can do, here is what it can't, and here is why." That document makes you the resident expert on AI applicability in your domain, which is precisely the person no restructuring plan can afford to eliminate, because eliminating you means losing the only reliable map of what works in your corner of the organization.

The Bottom Line

Klarna tried top-down automation and is now rehiring. Seventy-four percent of enterprise AI projects fail to scale. The consultants' maps don't match the territory, and the alternative is not to wait and hope. It is to run the experiment yourself, with full knowledge of every edge case and judgment call your role requires, and let the failures tell you something the successes never could: exactly where the boundary is between what AI can do and what still requires you. Try to automate everything you do, and what's left is the job.