An essay titled AI Mania Is Eviscerating Global Decision-Making spent the middle of July 2026 climbing Hacker News and circulating through engineering circles, and it puts a blunt label on something many people inside large companies have been describing privately for two years. The piece, written by consultant Nik Suresh and published on his blog, argues that corporate AI mania has done more than waste budgets. It has made honest conversation about failing projects a firing offense, which means the people best placed to correct course have every incentive to stay quiet. The claim is uncomfortable, the anecdotes are vivid, and the response was large enough that it is worth reading the argument carefully rather than through the filter of a headline.
- The essay claims every enterprise AI project its author observed over roughly 18 months has failed to deliver, while being reported internally as a success.
- The central mechanism is social, not technical: dissent about AI is treated as heresy, so executives who privately doubt their own strategy keep repeating it.
- Vendors say they cannot contradict a customer executive who claims 100x productivity gains without risking the contract.
- The author draws on roughly 300 professional conversations across companies from Fortune 500 scale down to niche firms.
- Simon Willison amplified the piece and the Hacker News thread drew hundreds of points, signalling how widely the frustration is shared.
What the corporate AI mania essay actually argues
Suresh writes from the vantage point of a data and machine learning consultant who sits in rooms where AI strategy gets set and then watches what happens when the slide deck meets reality. His starting claim is deliberately absolute. Across the projects his team has observed over a stretch of roughly a year and a half, he says, not one has worked as sold. Internal chatbots go unused because the underlying documentation is too thin for them to answer anything useful. Customer-facing bots frustrate the people they are meant to help, and he reaches for a concrete example, a callback system he associates with Mitsubishi that simply did not function as promised. Where success metrics exist at all, he argues, they are either not tracked or trivial to game. The line that anchors the section is stark: every single AI project he has watched is failing.
It would be easy to dismiss that as one jaded contractor generalising from a bad run of clients. Suresh anticipates the objection by describing the scale of his sample. He puts the number at around 300 professional catch-ups, spanning organisations from Fortune 500 giants down to small specialist firms. That breadth matters to his case, because his real subject is not the technology. It is the behaviour that forms around the technology once a company has publicly committed to it.
The essay's sharpest move is to reframe the problem as one of decision-making rather than engineering. A model that underperforms is a normal, solvable situation. A company where nobody is allowed to say the model underperforms is a different kind of failure, because the correction mechanism itself has been switched off. That is the eviscerated decision-making of the title, and it is why the piece reads less like a product review and more like an anthropology of a workplace under pressure.
The anecdotes that made the essay travel
Arguments about culture spread when they carry stories, and this essay is dense with them. The one that Simon Willison singled out first when he linked the piece involves an executive who, by Suresh's account, produced a technical strategy for an organisation with more than two billion dollars in revenue that was built entirely around AI, having never personally used ChatGPT or any comparable tool. The detail lands because it inverts the usual assumption that strategy follows understanding. Here the commitment came first and the understanding was optional.
A second story captures the defensive absurdity the essay says the mania produces. An engineer describes checking out a parallel copy of a Go codebase and instructing an AI to rewrite the whole thing in Zig, not because anyone needed a Zig rewrite, but as a way to look busy on an AI initiative and hold onto a job.
Willison called out that vignette as a small masterpiece of misaligned incentives, and it is. The engineer is not lazy or dishonest in any ordinary sense. He is responding rationally to a system that rewards the appearance of AI activity and punishes anyone who suggests the activity is theatre.
The third anecdote is the one Willison said he enjoyed most, and it explains the mechanism behind the whole phenomenon. A skeptical executive at an over-enthusiastic vendor was asked why obviously inflated claims kept going unchallenged internally. The answer was not that the sales material was cynical, though it partly was. The deeper constraint was that the vendor's own customers had executives publicly promising 100x productivity gains. If anyone on the vendor side said out loud that such gains were not plausible, they would be undercutting the customer executive's credibility, which would be received as an attack, and could end an enterprise contract. Puncturing the fantasy, in other words, was a fireable act on both sides of the table.
Why dissent became dangerous
The essay's most transferable idea is that skepticism about AI has taken on the texture of heresy inside many organisations. Suresh argues that career advancement now frequently requires a kind of religious profession of faith in AI's world-changing power, and that voicing doubt can be career-ending. He describes cases where high performers who delivered results without leaning on large language models were sidelined, because their success embarrassed a narrative rather than confirmed it.
To show how deep the private doubt runs, he points to anonymous polling. When he surveyed teams without attribution, the responses split into two camps at once. Some rated a project a three out of ten while others gave the same effort an eight, and projects running years behind schedule were still receiving generous scores from people who, in a named setting, would presumably have stayed diplomatic. The bimodal pattern is the tell. It suggests that a large share of participants privately think a project is failing, and that the group average that leadership sees is smoothing over a genuine split the room is not allowed to voice.
These are one consultant's claims, drawn from anonymised anecdotes rather than published survey data. The essay is persuasive as a description of incentives, not as a controlled study. Read it as a hypothesis about corporate behaviour that many practitioners recognise, not as a measured failure rate for enterprise AI.
Game theory does the rest of the work in Suresh's telling. An executive who has staked their standing on an AI strategy cannot easily admit it is not paying off, because admitting failure invites being labelled incompetent and replaced by someone more willing to keep the story going. So the rational move for each individual leader is to keep professing confidence, even when a quiet majority of them privately expect the effort to disappoint. The collective result is an organisation that has lost the ability to update on evidence, which is a fairly precise definition of broken decision-making.
Demos, AI-washing, and the machinery of over-purchasing
Beyond the culture, the essay describes concrete mechanisms that convert enthusiasm into spending. One is the demo. Suresh recounts pulling a Snowflake Cortex demonstration out of his standard pitch because it was working too well as a sales tool. A chatbot hitting around 92 percent accuracy under ideal conditions turned lukewarm prospects into eager buyers, even after he explicitly warned that their use case was unsuitable. The demo bypassed the caveats entirely. People saw a number that felt like magic and stopped listening to the conditions attached to it.
The other mechanism is relabelling. Projects that have little to do with modern AI get an AI badge after the fact so they can ride the budget wave and be counted as wins. His example is a database migration where the actual work amounted to some straightforward SQL translation, later billed and celebrated as an AI-driven win. This AI-washing is quietly corrosive, because it inflates the apparent hit rate of AI initiatives with successes that had nothing to do with AI, which makes the genuine failures even harder to see.
Along the way Suresh name-checks figures whose more grounded posture he seems to respect, including Mitchell Hashimoto, Thomas Ptacek and the team at Fly.io, and the consulting elder Gerald Weinberg, whose advice on navigating dysfunctional organisations threads through the survival section. He also nods to the WordPress world and Matt Mullenweg. The references are not endorsements so much as reference points, a way of gesturing at what sober engagement with these tools can look like when it is not being driven by fear.
How the piece landed, and why that matters
The reach of the essay is part of its significance. It surfaced on Hacker News and gathered hundreds of points and a long comment thread within a day, then got a further boost when Willison, one of the more widely read independent voices on practical AI, linked it and pulled out its best passages. Willison introduced it as an entertaining perspective crammed with spicy anecdotes from anonymous sources, which is a fair description, but the volume of the response says something the anecdotes alone do not. A lot of technical people read this and recognised their own workplace.
That recognition is the real data point here, even if it is soft. The specific numbers in the essay are unverifiable by design, since the sources are anonymous and the sample is a consultant's memory rather than a dataset. But the shape of the complaint, that AI has become a topic on which honesty is expensive, matched enough lived experience to travel fast. When a critique spreads because readers keep saying some version of yes, that is exactly what my company is like, the critique has captured a pattern even if every individual figure is impressionistic.
Surviving an organisation that cannot listen
The essay does not pretend that a single contributor can talk a mania-struck company out of its convictions, and the advice Suresh gives for people stuck inside one is notably defensive. For those trying to nudge things in a better direction, he suggests avoiding group dynamics wherever possible, leaning on anonymous polls to reveal the hidden skepticism, and pulling ground-level users into the conversation because they tend to answer honestly about whether a tool actually helps them. He also warns against ever challenging the broad claim head-on. Arguing that AI in general is overhyped reads as an attack on everyone who has staked their standing on it, so the practical route is narrower, questioning a specific project's specific results rather than the whole faith.
For those who conclude the culture will not change, the counsel turns to self-preservation. Accept that meaningful pushback is not available, he advises, and consider structuring work around fixed endpoints, the kind of contract with a defined finish line that a consultant can complete and leave, rather than an open-ended role that requires daily participation in the pretence. He suggests limiting how much AI-hype content you consume, on the grounds that a steady diet of it distorts your own sense of what is normal. And he is candid about the exit ramp, recommending that anyone asked to review huge volumes of low-quality AI-generated code as their main job should quietly begin a search elsewhere before the situation curdles further. It is bleak guidance, and the bleakness is the point, an honest reckoning with how little leverage an individual has once a whole organisation has agreed not to look too closely.
Five lessons for leaders who want honest signal
Suresh closes with practical advice for people trying to change a mania-struck organisation from the inside, and the counsel translates cleanly into lessons any leader can act on without waiting for a bubble to burst.
First, separate the question from the crowd. The essay's whole diagnosis is that group settings suppress honesty, so the single most useful change a leader can make is to gather signal in private. One-on-one conversations and genuinely anonymous polls surface the split that a status meeting hides. If you only ever ask how the AI project is going in a room full of people whose bonuses depend on the answer, you will hear the answer their incentives require.
Second, treat unanimous enthusiasm as a warning, not a green light. A bimodal response, where half the team is quietly alarmed and half is delighted, is healthier than a room that agrees on an eight out of ten. Suspiciously smooth consensus around a hard project usually means dissent has been priced out, not that it does not exist.
Third, insist on metrics before badges. The AI-washing pattern only works because outcomes are measured loosely. Deciding in advance what a specific initiative must move, and refusing to relabel unrelated wins as AI victories, keeps the scoreboard honest and makes the genuine failures legible instead of buried.
Fourth, distrust the demo. A polished demonstration hitting a headline accuracy number under ideal conditions tells you almost nothing about your messy production data. Suresh pulled his best demo precisely because it short-circuited people's judgement. The disciplined response to an impressive demo is to ask what happens when the conditions are yours, not the vendor's.
Fifth, protect the people who tell the truth. The failure mode the essay describes is one where skeptics get punished and true believers get promoted regardless of results. Reversing that, by making it safe to report that something is not working and rewarding accuracy over cheerleading, is the only durable fix, because it turns the correction mechanism back on.
A mania that outlives the bubble
Suresh's own conclusion is not that AI is worthless. It is that the current dysfunction stems from leadership structure and incentives rather than from the tools themselves, which means the dysfunction will persist even after the market euphoria cools. When the bubble deflates, the companies that spent two years unable to speak honestly to themselves will still be companies that struggle to speak honestly to themselves. The habit of treating a technology choice as an identity, and doubt as betrayal, does not evaporate when the hype cycle turns.
That is the part worth carrying away from a piece that is otherwise easy to enjoy as a collection of war stories. The technology will keep improving, and some of the projects that failed in 2026 will work in 2028. The harder question the essay raises is whether the organisations running them will have kept the ability to notice the difference. A tool is only as useful as a decision process that can tell whether it is helping, and the argument here is that corporate AI mania quietly dismantled that process in a lot of places while everyone was watching the demos. Rebuilding it is less glamorous than any model launch, and considerably more important. For a related look at how open models are changing the cost side of this picture, our blog tracks the releases that keep resetting expectations.