AI Fragility: Why I’m Not Pro-AI or Anti-AI, but Pro-Thoughtfulness
tl;dr
AI is neither savior nor threat. The differentiator is thoughtfulness: critical thinking, problem solving, and control over how the work gets done.
The bulk of the business world is being split into two camps: pro-AI and anti-AI. I’m neither. I’m pro-thoughtfulness. AI is here, and we can’t stick our heads in the sand. But whether straight-up fear of AI or FOMO is driving behaviors, none of that will yield good results.
There is plenty of rapid evolution to come, and plenty of on-ramps to get caught up to speed. I think the risk of waiting before any serious focus is low. The risk of going too fast is high. But whether you pay attention to the news and to case studies, you experiment and learn, or you focus massive resources, the risk of breaking things is very real.
This is why I wrote about the handbrake in an AI strategy. The accelerator will affect your speed, and the steering your direction, but the handbrake is paying attention to the risks.
AI is just a tool
No matter how powerful, a tool is still a tool. Tools need purpose, users, knowledge, and more to be useful. The same is true about AI. This multi-tool may be powerful, but that has people confused, thinking that it’s different from a tool. It is not.
The core of this is critical thinking. Without critical thinking, you lack the ability to guide AI, the ability to assess AI outputs, and the courage to protect your unique insights. Of course, critical thinking has been important throughout the lifespan of human existence, but there is a very real threat to taking this capability for granted.
Problem solving is also vital, as that’s most of what we want AI to help us with. You need problem-solving capabilities and a process. This is why learning that problem solving is more than a tool or template is so vital, as you can’t just replace your template with AI. Even OpenAI CFO Sarah Friar shared on a CNBC interview that the vast majority of chats are limited to a single prompt and response. Problem solving begins with good problem selection and clear problem statements, but goes further through understanding cause and effect, ideation, experimentation, and more.
Having control over your data is almost a given, although AI has simultaneously required us to improve our control of data and allowed us to extract more value of our messy data. But beyond data, maintaining control of your standard work and processes is just as important.
This ultimately suggests that you should have very good awareness and control over HOW the work is done. That was already beneficial before AI, but without good process standards, the “black box” is more like a black hole, with no ability to diagnose, assess, get consistent outcomes, or learn and improve. Brittany Irwin, Applications and AI Engineering Manager at NFI Industries, states (on the People Solve Problems podcast1) that having good control and definition of your standard work may be just as critical as having control over your data. However, because it is easy to turn over the “how” to AI, many are losing control over the “how.” And the AI companies are encouraging it, perhaps not with malice but certainly ignorance, essentially prioritizing their compute power.
Does AI replace people, and should it?
There are examples of AI effectively replacing humans, but most of these cases are where the humans weren’t doing a particularly good or distinctive job at that task. In most cases, AI is improving value, quality, speed (which is different than capacity), or handling tasks that weren’t being done before.
The news would have you believe that tens of thousands of employees have been laid off and replaced by AI. If you believed all press releases and earnings calls at face value, that would be true. But there is little evidence that these lost jobs are actually replaced with AI. In fact, some high-profile cuts are being reversed: Commonwealth Bank rehired staff and acknowledged that an AI-driven role reduction had been a mistake. 2
In some cases, companies have been a bit bloated with overhiring or are struggling with their markets and are using the “positives” of AI as a cover for bad news. In other cases, executives think that the pressure created by removing the resources will accelerate adoption and application. This isn’t an invalid method, but it doesn’t mean AI is doing the work of those individuals.
Part of this equation is that AI generally reverts to the median answer, but this does not get you above-average results or unique results. The ability to take conventional wisdom (which is what the AI trained on) and say “this isn’t right; there’s a better way or a different way” is where breakthroughs happen. The Wright brothers began designing their wings based on the published lift tables of aviation pioneer Otto Lilienthal and the long-accepted Smeaton coefficient, but as they started experimenting, they determined that the experts at the time were indeed wrong, building a wind tunnel and recalculating the figures from scratch.3 This required courage, and maybe a bit of arrogance too, to decide to rebuild existing knowledge from scratch. Both the decision and the unique work that followed helped bring the aviation age into existence. Aristotle would refer here to phronēsis, or human judgment, which is the kind of fork-in-the-road decision and the selection of the just fork that AI isn’t likely to master.
Is AI economical?
This is part of the argument and why layoffs appear to come from all directions. A computer doing the work of a human doesn’t sleep or take breaks or need a 401(k).
First, many of these layoffs that were articulated as for AI weren’t completely valid. For many of them, stating that they had become bloated with resources doesn’t tell a very good story, but there were many companies that, subjectively, just had way too many employees. Saying that the economy isn’t strong isn’t very popular and, to be fair, requires nuanced explanation, as the top-line economy may be strong but many segments are struggling mightily. For these companies, saying that they don’t need 3,000 employees because they have invested in AI, true or not, leads to a nice little “they’re forward-thinking” stock bump. Were both things true (don’t need 3,000 employees, and they invested in AI)? Yes, they were true. But it doesn’t mean one was actually the cause of the other.
As written in the Wall Street Journal 4
“Companies that are too quick to lay off workers on the assumption that AI can do their jobs risk wrecking their future competitiveness in two ways. The first is that they can lose critical institutional knowledge. The second is that they risk hurting their own talent pipelines. While it may be tempting to replace junior engineers with AI, doing so means that when senior engineers move on, a company will no longer have the humans required to review the work of those AIs.”
For other executives, the layoffs are a bit of a forcing function. This is the equivalent of the burning of the ships. If you’re worried that your employees aren’t adopting AI fast enough, take away their other ways of getting work done, like having a team of employees. This is a messy path not based on doing the math or knowing the outcome, but has proven an effective strategy.
In most cases, this is “forward-thinking” virtue signaling and hoping for the stock bump that comes with being an innovative company. This corruption of terminology does more than manipulate investors; it removes credibility of more essential core strategy and values espoused by the respective executives.
But employee costs are only part of the equation. The equation only balances if the costs of AI make sense as well. Companies are finding very quickly that enabling AI doesn’t fit well into budgets for two reasons. First, there is no one best tool you can simply give employees access to. Even the frontier models are best at different things, so companies are ending up signing up for multiple enterprise licenses of frontier models. On top of that, there are many specialty software companies built on top of frontier models that solve specific problems for them. The software stack is getting expensive, and none of it has replaced the enterprise software tools that they are already paying for and committing resources to service.
For those using token-based plans, especially in the coding space, the upper limit of spending proves both uncapped and hard to forecast. If there is anything CFOs hate more than rising costs, it’s unpredictable costs. There are both legitimate and questionable drivers of this. The most questionable is what’s known as tokenmaxxing.
Tokenmaxxing is the AI equivalent of the person who schedules extra meetings and sends too many emails in the spirit of being “productive.” Productivity is simply output divided by input, and if you assume the ratio is fixed (it is not), then just generating more input will equal more output. Tokenmaxxing is working in a way that ensures that you consume as many tokens as you can, which, of course, you pay for.
Employees at Meta generated a tokenmaxxing leaderboard until its obviously questionable behavior impact led them to take it down. Developing agents that run multiple agents in parallel can certainly do more at once, but they also cost more, and companies are noticing.
Microsoft dropped most of its Claude Code contract, reportedly over cost, directing engineers to GitHub Copilot CLI by mid-2026.5
The starkest numbers come from Uber, which deployed Claude Code to roughly 5,000 engineers and saw individual engineers spending between $500 and $2,000 per month on tokens, burning through its annual AI coding budget in four months.6 When much of the industry is trying to move forward unleashed, it’s still a telling story. This is all leading companies to explore many cost containment strategies, according to The Economist, including downgrading model selection, open-source models, and spending caps. A fundamental difference in this world is that for most software, the incremental user costs nothing, but with AI, that’s no longer true.
Part of the problem isn’t the behavior of the individual but the capabilities of the tools. Being more capable requires them to think hard. As you can see from this chart (I generated with Mistral), while the cost per token has dramatically decreased, the number of tokens per task has also risen dramatically and appears as if it will continue to rise. Goldman Sachs projects that token consumption could grow 24 times by 2030, reaching 120 quadrillion tokens per month, driven largely by agentic AI. 7
The cost per task has turned the corner and is starting to rise.
That doesn’t take into account two factors. One is the number of tasks per person completed, which is also dramatically on the rise. This is why, while token costs drop 98%, AI bills have tripled or worse. The other factor is that, by all accounts, the price of tokens is still being subsidized to encourage growth. That means the true costs may be higher. How long can they continue to be subsidized? Well, let’s ask another question: as SpaceX, xAI (owner of Grok and soon Cursor), Anthropic, and OpenAI go public, how will increased transparency and shareholder pressure affect pricing?
The final piece of the economic puzzle that can’t be ignored is AI slop. AI slop may be most prevalent on your social media feeds, which just cost your attention and time, but inside the world of work, which we will call workslop, it is costly on tremendous fronts. People can generate more emails, PowerPoints, and reports, but to what end? Well, for other humans to consume. And unless that work content actually advances the cause, improves the insight, helps make better decisions, then it is likely to be waste. It was waste in creating it, in sharing it, and in consuming it.
The number of people experiencing workslop is staggering and frustrating. In one survey of 1,150 U.S. workers, 40% reported receiving workslop in just the previous month, and respondents estimated that roughly 15% of all the content they receive qualifies.8 Not only does it create waste, but it reduces trust and harms collaboration. The same research found that workers who received workslop viewed the sender as less trustworthy (42%) and, in many cases, less capable and less reliable. 9 Lack of transparency of when people used AI and for what part of the work leads everyone to question everything they receive: Was this AI-generated, and will I regret reviewing it? Each incident is not free to untangle either, taking an average of nearly two hours to resolve.10
Overall, will AI improve the economic frontier of work? That’s almost a certainty. But not likely in as many cases as it’s being shoved, and not nearly yet.
Gen Z isn’t “AI-native”
Many baby boomers and Gen X are chasing AI competencies for fear of being made irrelevant in a world where new employees come in as “AI-natives,” meaning they grew up in an AI world. As it turns out, that may be an invalid fear.
Of course, we’re not yet at a point where kids truly grew up in an AI world, but even those who had exposure to AI throughout high school or college are proving a very mixed bag between AI-native and AI-resistant.
Graduation commencement speakers were quite famously booed when bringing up the promise of AI, and then appeared quite surprised by the reaction, as if they had never spoken to a college kid on the subject. The backlash is about a lot of factors, not the least of which is companies hiring fewer and fewer new college grads, presumably under the assumption that they don’t need them because of AI.
But of course it’s more than that. Environmental impacts of construction, energy, and water usage are very real, and data centers aren’t going in remote places away from populations because that’s not where the energy is found. This resistance is real, from activists to local communities that simply don’t want data centers. That resistance is elevated by outside social media interference, such as from the CCP who clearly recognizes that friction on the US AI frontier gives an advantage to them.
While AI may create enough need for energy that we finally break through technological barriers to clean and abundant energy, the only way that happens is by first creating an energy crisis. This is one place where China has an advantage: building energy supply any time and anywhere they want because resistance isn’t allowed, but in an open market like the US, this could very well mean price hikes or even outages.
What’s the point if the outputs are bad?
Getting good outputs from AI isn’t trivial, nor is it consistent. While the reduction of hallucinations is appreciated, it also comes with AI tools better at hiding them when they do mess up. Producing content that sounds thorough and sophisticated is easy, but it will self-reward for the style regardless of the accuracy. When humans do that, we call it “baffling you with bullshit.” I guess we should call it the same thing when your AI does it. With most tools, when there are errors or problems, experts can identify them. AI fluency makes identification of errors more difficult, even for experts. And the risk is compounded by how often unverified output gets shipped: 69% of workers admit to passing along AI work without fully checking it first.11
Some AI just doesn’t work as planned. Starbucks dumped its AI inventory system after 9 months, which is long enough to work on making it better and, in this case, realizing it can’t get there. Good for them for making the bold and embarrassing call to pull the plug, but how much investment, lost revenue, and lost trust was the net outcome?
I recently had a tool build a dashboard for me for a topic I was researching, and because a certain tool was apparently not available at the moment (no explanation as to why), it just skipped the quality assurance (QA) step. Perhaps worse is that I only noticed this fact as I glanced over as the activity log was scrolling by, but there was no mention of that when it finally presented its work. The analogy of AI as an inexperienced but smart intern often seems valid, but you’d actually think even then some common sense would have you believe that the intern would either pause work or let you know that it couldn’t complete a key step. How many important processes do you have where that’s an acceptable risk to take?
And while models continue to get better, they also continue to be different as they get released in rapid succession. Building your business on a foundation of shifting sands is quite difficult, with each new release, policy shift, or pricing shift having you wondering what is now broken. This all has employees spending more and more time managing their AI tools, significantly cutting into any gains they are finding. One study found that employees spend about 6.5 hours a week managing AI, and that despite 87% of workers using AI, only 13% said it had significantly improved their company’s performance.12
But the outputs of AI aren’t the only problem, and perhaps aren’t the main problem here. Workslop is becoming more prevalent. Workslop is the workplace subset of the broader term “AI slop,” where content generated by AI doesn’t advance the work but often confuses people or generates more work. Not only does this cut into any potential productivity gains, but it hurts trust and relationships and, ultimately, coordination and collaboration across an organization.
The AI genie is out of the bottle
It’s here, and it will be a part of your life and work from here forward. Asking if AI belongs in your life is like asking if the internet or electricity belongs in your life. But what that future is like is up to us. It’s up to us how we think, work, make decisions, collaborate, take ownership, and more. I’m not pro-AI or anti-AI; I’m pro-thoughtfulness. Political theorist Hannah Arendt never had the chance to write about AI, but as she described the “banality of evil,” it was of people acting and simply just going along without engaging that inner dialogue to evaluate their intent and actions. That could lead to disaster, individually and collectively. Destiny is not written.
Endnotes
1. Brittany Irwin, ‘People Solve Problems’ podcast, NFI Industries. https://www.jflinch.com/brittany-irwin-ai-engineering-manager-nfi-industries-why-your-ai-rollout-isnt-really-about-ai/
↩ Return to text
2.Commonwealth Bank of Australia reversed AI-driven job cuts and acknowledged the decision had been a mistake, rehiring affected staff, as reported in mid-2026 coverage of
enterprise AI rollbacks. See Klarna and Commonwealth Bank AI rehiring coverage
3. In 1901 the Wright brothers built a wind tunnel and tested roughly 200 wing shapes, recording data on about 50, after concluding that Otto Lilienthal’s published lift tables and the long- used Smeaton coefficient (0.005, in use since the 1700s) were inaccurate; they recalculated the coefficient to about 0.0033.
NASA Glenn Research Center; Smithsonian National Air and Space Museum.
4.‘When AI is More Harm Than Good’, The Wall Street Journal, 2026. https://www.wsj.com/tech/ai/when-ai-more-harm-than-good-519a83e7?mod=hp_listb_pos2
5.Microsoft canceled most internal Claude Code licenses in its Experiences and Devices division, directing engineers to GitHub Copilot CLI by June 30, 2026; the stated reason was “toolchain unification,” with cost the widely reported rationale. The Next Web; Yahoo Finance
6.Uber deployed Claude Code to roughly 5,000 engineers; per-engineer token costs reached $500 to $2,000 per month and the company exhausted its 2026 AI coding budget in four months, per CTO Praveen Neppalli Naga. Yahoo Finance / The Information.
↩ Return to text
7.Goldman Sachs Research, “AI Agents Forecast to Boost Tech Cash Flow as Usage Soars,” May 20, 2026, projecting token consumption to multiply 24 times by 2030, reaching 120 quadrillion tokens per month. Goldman Sachs.
8.BetterUp Labs and the Stanford Social Media Lab, survey of 1,150 full-time U.S. desk workers, September 2025: 40% reported receiving “workslop” in the prior month, estimating about 15% of received content qualifies. BetterUp Labs.
9.Same study: 42% of workers who received workslop viewed the sender as less trustworthy, and about half saw them as less capable and reliable. BetterUp Labs.
10.Same study: resolving each workslop incident took an average of one hour and 56 minutes, contributing to an estimated cost of roughly $186 per worker per month. BetterUp Labs; The Next Web coverage.
11.Glean Work AI Index 2026 (6,000 full-time digital workers, US/UK/Australia, December 2025–January 2026): 69% of workers admitted shipping AI-generated work without verifying it. Glean Work AI Institute.
12.Glean Work AI Index 2026: workers spent about 6.4–6.5 hours per week “botsitting” (managing) AI tools; 87% of digital workers use AI, but only 13% said it significantly improved their organization’s performance. CIO Dive; Glean Work AI Institute.