pawprints in desert dust

Here are a few highlights from this morning's read – moments where I had to look away from my screen and read out a passage to my wife so that we could share a "holy shit" moment.

A person worked for a division that "pivoted" to re-engineer its software to create interfaces that support AI agents. When it became apparent that only ten users had touched this expensive new technology, they "pivoted" again to support "agentic workflows." Why did they double down on AI agents after discovering such yawning market indifference for "agentic"? "Because every company has to do something agentic now."

Suresh describes this as a literal religious mania. In the 500+ employee businesses Suresh studied, the only people who were promoted – or even spared from being fired – were people who professed "religious declarations of faith" about "the transformative power of AI." Employees who voiced honest, informed objections to AI in the workplace were passed over for promotions or targeted for layoffs.

This has created a situation in which everyone – "boards, executives, employees, vendors, consultants" – has a strong incentive to lie about how much AI is delivering for their companies. Suresh says he's seen announcements from publicly traded companies about their AI triumphs that he knows for a fact never took place.

Suresh says he's never seen a successful enterprise AI project: "Every single one – we have seen 0% success in a year and a half." Not one of their clients would face a business challenge if OpenAI went out of business tomorrow. The problem most companies struggle with is that they're "terminally bad at running software projects effectively." Adding AI to the mix doesn't solve this problem – it just adds a whole new range of ways that software deployment can fail.

Chatbots don't help. The internally facing chatbot that's supposed to help employees figure out how to navigate the business sucks because it is only as good as its training data – the business's documentation of its own processes. Businesses suck at documenting their processes. Customer-facing chatbots also suck. They either can't solve your problem, or, when they seem to solve your problem, the "solution" goes nowhere.

Suresh recounts his sole positive customer service chatbot experience: a Mitsubishi chatbot with a natural sounding, responsive voice politely took all the details of an automotive failure and promised him a callback. That callback never came, but Suresh is certain that Mitsubishi has logged this as a chatbot success story, even though the experience convinced him not to buy a Mitsubishi car.

Suresh and his team at Hermit Tech now have a policy of not even asking about ongoing AI projects. They've learned that by the time an AI project has begun, no one will discuss it honestly until it reaches a crisis point.

Suresh says he frequently encounters people who reflexively utter the AI catechism: "AI is changing everything." But when he presses these people for details, they admit that their organization "does not currently use LLMs for anything, and indeed, that they cannot name a single thing that has changed other than they get some use out of ChatGPT."

This shear ("AI is changing everything"/"Well, OK, we're not using AI for anything") is so extreme that Suresh once met an exec who confessed to crafting an AI-centered AI strategy for a $2b/year business, even though that exec "had never even used ChatGPT or any AI tool in their life."

Some people have privately admitted to Suresh that they've embraced AI in order to earn a career-boosting corporate reputation for "thought leadership." But many other people (especially nontechnical people) sincerely believe that AI is about to "change everything." As Suresh says, if you're in business with a liar, you might be able to reason with them in private – but you can't reason with a true believer.

The true believers are in charge. Suresh points out that it would be very weird for the CEO of an engineering firm or a hospital to mandate "specific procedures or building techniques without explicit agreement from the professionals on staff." But when it comes to AI, business leaders will confidently demand that the skilled professionals who perform the business's core functions use AI, even if those professionals don't think it will help.

As an aside: I remember the dotcom era, when the business press was full of articles about the conflict between CEOs and a new workforce that demanded the right to use the web on the job. Today, the business press is full of articles about the conflict between the workforce and CEOs who demand that they use AI.

Suresh describes workers who feel they have to "AI wash" their work: "They just do the work, the same way they have for decades, and say Claude did it." To add verisimilitude to this sham, they write circular processes in which one chatbot prompts another, and then the process repeats itself in reverse, for the sole purpose of consuming AI tokens to score a high rank on corporate "token leaderboards."

How to account for this wildly, expensively irrational corporate leadership? Suresh places the blame in the hypnotizing, mesmerizing power of the AI demo. For example: Hermit Tech is often engaged to set up a database product called Snowflake for its customers. Snowflake has a useless, expensive AI bolt-on called Cortex, that Snowflake itself describes as being 92% accurate under ideal circumstances (that is, at least 8% of the time, it will mislead you, perhaps very badly).

Suresh describes sales meetings with execs who were lukewarm on the idea of retooling with Snowflake, but who were very interested in Cortex. Against their better judgment, Suresh and his team provided them with a Cortex demo, carefully explaining that this AI tool could not satisfy their requirements. Without fail, this resulted in the previously lukewarm customers insisting that they be allowed to purchase Cortex immediately. Sales prospects who'd been unmoved by a pitch for new technology that would result in millions in savings were hypnotized by demos of a product that was described as unsuitable and unreliable.

To their credit, Hermit Tech refused to sell these customers Cortex, and stopped doing Cortex demos altogether. Suresh describes the experience of "the total 180°, that shift from ice-cold to red-hot buying frenzy" as "deeply unsettling." What's more, the Cortex demos that Suresh and co performed were, by his account, pretty uninspiring. The thing that these demos had going for them is that they showed AI actually doing something marginally useful, to execs who'd already spent millions on AI without having anything to show for their money. The spectacle of AI that does something galvanizes corporate leaders who feel like they're the only bosses who can't find a revolutionary use for AI in their businesses.

This is the situation up and down the corporate org-chart. Suresh has a reader whose title is "Head of AI" at a billion-dollar firm who tells him "their job is totally fraudulent but it was the only promotion pathway remaining at the organisation." This exec is hardly alone. They're part of a cohort of executives at companies that have publicly announced "100x" productivity gains, but who confessed to Suresh that nothing of the sort has happened.

Why did these companies make these claims? Because their customers were making the claims. How could you hope to sell to a company that had 100x'ed its productivity with AI unless you, too had 100x'ed your productivity? If, as a vendor, you walked into a boardroom and said that this wasn't a plausible claim, you'd be calling your sales prospect a liar, with real consequences: "getting enterprise contracts cancelled because you wanted to opine on something that doesn’t really matter to your organisation’s mission is a great way to get fired."

With the state of the industry dominated by froth, lies and mutual destruction pacts, it's no wonder that companies are deploying "totally gameable metrics such as 'money spent on AI'" as a means of evaluating employees and divisions.

Between true believers and people who must find ways to plausibly tout their AI usage, there is now a gigantic market for "AI solutions." At best these are just traditional tech consulting contracts, like migrating a database from Oracle to Snowflake, with some kind of ornamental AI usage around the edges so that the person who commissions the work can claim to be "procuring AI-enabled services" for the business.

This isn't a harmless frippery: contracts are delayed and work is put off until the work can be made "sufficiently AI" to attain the minimum degree of buzzword compliance. Worse: every fake AI project that produces real results (because it's not really AI) adds credibility to the AI true believers, who view these projects as proof that AI can do anything, and therefore demand to know why everything isn't being done by AI.

Suresh ends his essay with a long section on how to "navigate AI mania" – advice for how to smile and nod politely when you're confronted with AI bullshit, while steering clear of the worst consequences and avoiding needless fights. This looks like very sound advice for anyone in a corporate environment, but thankfully, that isn't me.

Rather than summarize that advice, I want to reflect a little on two questions that Suresh's essay raises but doesn't answer. The first is why? Why are people in power such easy converts to this religious mania?

I have my own theory. The most important discomfort that powerful people experience is having ego-shattering conflicts with subordinates who know how to do things they do not know how to do. The fact that you're "in charge" is hard to reconcile with the fact that the people you're nominally in charge of tell you that all your ideas are impossible, illegal, immoral, or lethal:

https://pluralistic.net/2026/01/05/fisher-price-steering-wheel/#billionaire-solipsism

Take that Cortex demo. Sure, Cortex is an expensive, unreliable way to address a Snowflake database. But (unlike Snowflake) Cortex is controlled via conversational, plain-language commands. With Cortex, a boss doesn't need to ask an underling to retrieve information from the company Snowflake system, an interaction that might come with unsolicited feedback about the technical or commercial incoherence of the boss's request. Cortex is the underling, except that unlike a human underling, Cortex never back-sasses you about your foolish questions. The fact that it grossly misleads you 8% of the time is a small price to pay for a life untroubled by uppity pismires who insist that your ideas be connected to base reality as they understand it.

The other question Suresh implicitly raises is, "How can you reconcile the failure of AI in the enterprise with the individual claims of skilled technologists who insist that AI is helping them do great work?" The answer is that these AI users are "centaurs" – experienced workers who are assisted by automation on terms that they set for themselves:

https://pluralistic.net/2025/09/11/vulgar-thatcherism/#there-is-an-alternative

Thanks to their skill and experience, these workers possess discernment, the ability to tell good code from bad, and (more importantly) good uses of code-generation tools from bad. They demonstrate the adage that worker-driven automation improves quality, while capital-driven automation improves throughput:

https://pluralistic.net/2026/07/28/hitl-ers/#ai-ai-oh

An automation technique that requires close supervision by skilled and experienced workers isn't going to be a raw productivity powerhouse. You don't "100x" your code this way, at least, not in the sense of firing 99 of your coders and having the remaining programmer pick up all their work. Rather, an automation tool that requires the continuous and conscientious exercise of discernment will let individual practitioners improve their work in extremely satisfying and useful ways. It's a way to spend more on operations in order to produce better outputs. It's not a way to cut your workforce, realize a gigantic savings, and still produce comparable goods and services at a far lower cost.

That is why some individual coders report such delight with their AI tools. They engage with those tools on their own terms, to improve their work in the ways that they, in their expert judgment, consider beneficial. No one ranks them on a "token-maximization" scoreboard. No one tells them they can't do a project if it isn't "sufficiently AI." When they set out to do a project, no one makes them prove that it couldn't be "done by AI."

As ever, the most important fact about a given technology isn't "what it does," but "who it does it for" and "who it does it to."