A weekend deep-dive on the second-order problems of the agentic era — AI slop, token economics, de-skilling, the vanishing pipeline of expertise — and the case that companies aren't the bumbling laggards the media paints, but are already solving these problems in public, one shared playbook at a time.
We've stopped asking if AI is a thing — and started solving the problems it creates.
A year ago enterprises were still debating whether AI was overhyped. Now the questions have flipped from 'if' to 'how': how to police AI slop with writing policies instead of bans, how to allocate scarce tokens like capital instead of SaaS seats, how to measure value per unit of intelligence, and how to keep building experts when AI eats the grunt work. New technologies solve old problems while creating new ones — and the encouraging story of this year is how fast, and how publicly, organizations are working the new ones.
The steam engine took nearly a century to show up in Britain's sustained productivity growth, electricity took roughly five decades, and computers close to a decade. The first stage of any technology revolution is infrastructure build-out — data centers, chips, power, talent — before gains diffuse across the economy.
Some areas deliver immediate, transparent, huge gains; others stay stubbornly stuck. And integrating agentic work with new human oversight creates a whole new set of work that, in the short term, fills in the time you won back — a transitional phase, not a wash.
Unlike traditional enterprise software, AI carries a meaningful marginal cost every time it's used, turning a one-time technology investment into a recurring operating expense. Firms have reported exhausting annual AI budgets within months, and EY's implication is clear: manage AI like any other capital allocation, where the value per token has to exceed its cost.
For AI's first few years, the equation was headcount times seat cost times twelve months versus value created. Agentic AI makes it less like software and more like a new type of labor — and the real surprise wasn't the shift itself, but how fast enfranchised employees could burn through expensive tokens.
The media discourse treating enterprises as incompetent bumpkins frantically slapping on usage caps is infuriating and wrong. Serious organizations are building complete architectures — different models and structures for different problems, different access levels for different people — and pairing caps with pathways to apply for more budget or demonstrate you deserve it.
The 'AI makes labor redundant' misconception was pushed by two groups: lab leaders (some now walking it back) and business leaders who needed a market-acceptable excuse for layoffs. That excuse isn't working anymore — and every story of companies re-hiring people they fired puts a dagger further into it.
Institutional and social immune systems are responding to the flood of terrible AI writing: AI detectors on Substack (about which NLW remains skeptical) and LinkedIn's new button letting readers flag that a post 'seems like AI slop.' The more of these systems emerge, the weaker the incentive to be lazy with AI writing.
Co-founder Varun Anand published Clay's official policy — originally for engineering, expanded company-wide: stand behind every idea and sentence; writing is thinking, so don't circumvent it; spend more time writing a document than readers spend consuming it; and longer is not better — if you generated a doc from a short prompt, consider just sharing the prompt.
Clay's policy doesn't brand anyone with a scarlet letter for using AI or even create subtle social pressure against it. It's an injunction to respect that the process of producing something is often as valuable as the output — and with 8,377 reactions on the post, expect it to show up at many more organizations soon.
In 'What Building an AI-Native Finance Function Taught Me,' Sarah Friar describes moving past static spreadsheets toward a real-time, reconciled, traceable view of the company's finances plus continuously updated forecasts. Getting there requires redesigning work around the decisions that matter — not just adopting new technology.
Friar's first lesson: access creates the most value when paired with structured experimentation around real problems — bottom-up experimentation plus top-down strategy. Intercompany hackathons have gone from sneer-worthy to a genuinely useful architecture showing up everywhere.
As organizations design more sophisticated ways to allocate scarce tokens, requiring everyone to prove ROI for every single use case will yield only the easiest-to-prove use cases — simple productivity enhancements, not the total reimaginings of work where the real leverage lives.
OpenAI research shows 40% of finance professionals' specialized AI use involves work outside traditional finance, and 22% involves engineering-related tasks. Friar says everyone on her team is building custom dashboards and tools with ChatGPT and Codex — work moving from static Excel and PowerPoint to live tools on the full context of the business.
Friar's scorecard asks four questions per workflow: Did AI complete work that mattered? What did it cost, including employee time, review, and rework? Was the result good enough to use? Did it help us move faster or make a better decision? Buying more seats or burning more tokens tells you almost nothing.
Section CEO Greg Shove's advice to CEOs: don't shrink the budget before you can see gains — invest in transformation and accept you won't have the full picture for one to two years. Pick a lighthouse team and 10× the investment, and avoid the 12-month stall: don't get skeptical, get specific about which teams are blocked and why.
BCG Global Chair Rich Lesser says the most common CEO question has quietly shifted from 'Which model should we use?' to 'Are we committing too much too soon to an evolving ecosystem?' The answer: organizations should own the harness where their 'enterprise cortex' — IP, data, business rules, codified process knowledge — lives, able to use any model or combination of models.
Per a BCG paper, the quiet erosion of judgment, critical thinking, and problem framing across a workforce can happen while adoption numbers look great on a dashboard — half of surveyed leaders already see it, and over 60% expect it to be a real threat within three to five years. Fewer than one in five employees feel confident using AI today; token usage isn't a proxy for adoption, confidence is.
Executives are twice as likely to increase investment in new technology as in employee training; 57% prioritize performance and efficiency while under 10% prioritize workforce training. The payoff for doing both: 37% of leaders who increased workforce investment reported 20%+ revenue growth over three years, versus 25% overall.
A paper called 'The Tragedy of the Cognitive Commons,' surfaced by Zara Zhang, names the bind: checking AI output requires deep expertise, deep expertise comes from years of grunt work, and grunt work is the first thing AI eats. Each company eliminating junior roles acts rationally; collectively, professions lose the ability to catch AI's mistakes. NLW's counter-question: is grunt work a law of nature, or just how expertise has always happened — and are there other ways to build it?
Over the past year, businesses moved from the not-particularly-useful question of whether AI is going to be a thing to the far more valuable questions of how to do it well. Keep asking those questions and sharing your answers in public — so everyone doesn't have to solve them alone.