"AI Mania Is Eviscerating Global Decision-Making" is an essay published on 18 July 2026 by Nikhil Suresh, of the Melbourne-based data consultancy Hermit Tech, on the firm's blog. It argues that enterprise enthusiasm for AI has degraded organisational decision-making itself: that professions of belief in AI have become a condition of advancement in large organisations, that demonstrations override stated buyer requirements, and that non-AI projects are relabelled to pass an internal test. The essay is written from the author's own consulting practice and from what he describes as roughly 300 conversations with professionals over the life of the blog, ranging from people in niche service industries to Fortune 500 executives (Source: hermit-tech.com).
The essay is a first-person account whose sources are deliberately anonymised — the author states he will "file the serial numbers off any stories." Its claims are the author's attributed observations rather than measured findings, and the anecdotes cannot be traced to named organisations. It follows his earlier widely circulated essay "I Will Fucking Piledrive You If You Mention AI Again."
The title appears in two forms: the on-page heading hyphenates "Decision-Making," while the URL slug, the author's own LinkedIn post and most secondary references use "Decisionmaking."
Summary of argument
Project outcomes
The essay's central empirical claim is that every AI project the author's team has observed is failing: "we have seen 0% success in a year and a half," covering both projects the firm was asked to participate in and projects observed in passing during unrelated work. The firm states it has rejected all AI implementation work, describing AI as "absolutely a gigantic bubble" and saying every current contract would be unaffected by an OpenAI collapse other than through second-order recession effects.
Suresh separates two causes. Some failures are not AI-specific: companies are, in his account, poor at running software projects generally, and AI projects carry the ordinary failure modes plus novelty risk. Others are limits of the technology. He identifies the internally-facing chatbot as the most common instance, arguing internal deployments see little uptake because organisational documentation is poor and a model "can only know things that have been written down and made accessible." In both internal and customer-facing cases he argues project leaders avoid tracking whether the tools are used at all, or track metrics that are easily gamed — illustrated by a vehicle-support voice bot that promised a callback which never came, a failure that would not register as an error in any metric.
He concludes that "almost every report at a company about 'massive AI productivity gains' is untrue as a matter of brute fact," allowing that genuine gains exist but are the exception. This bears on AI and Productivity, where self-reported enterprise gains are a recurring measurement problem.
Belief as a condition of employment
In every observed business above roughly 500 employees, Suresh reports that advancement and increasingly employment require repeated professions of belief in AI's transformative power — not proposals for using AI, but what he calls "religious profession." He reports executives asserting that "AI is changing everything" while conceding their organisation uses no LLMs, and one case of an executive who had never used any AI tool producing an AI-centred technical strategy for an organisation with more than $2 billion in revenue.
Two behaviours follow from adoption mandates. The first he calls AI-washing: engineers who complete work competently without AI report having used it, because managers are dissatisfied otherwise. The second is gaming of usage metrics, where staff measured on token leaderboards — with higher consumption scored as better — set models prompting themselves in loops and disregard the output; the essay quotes an anonymous engineer describing exactly this. Suresh states that the only people he knows to have been fired over the matter are those who voiced doubt about the strategy. See Shadow AI for the inverse pattern of undisclosed use.
Demonstrations and purchasing
The essay describes the firm demonstrating Snowflake's Cortex natural-language query layer to prospects who were lukewarm on its main offerings. Suresh characterises the accuracy of the tool from memory of a presentation by Snowflake staff as approximately 92% under ideal configuration — which he glosses as likely best-in-class for such tools but equivalent to a chief financial officer having one number in ten be wrong. He reports that every lukewarm prospect shown the demonstration wanted to buy immediately despite being told it would not do what they wanted, and that other considerations, including larger non-AI value, were set aside. The firm declined the sales and withdrew Cortex from its demonstrations, reasoning that "doctors don't walk around showing off cool pills that they'd never prescribe." He notes that a two-hour build was better than anything the prospects had previously seen, including at a publicly listed company already promoting its AI usage.
The coordination problem
The essay's mechanism for why implausible claims go uncontradicted comes from an anonymous Fortune 500 executive. Executives at that company's customers were claiming very large productivity gains; a vendor executive contradicting them would undermine the customer's credibility, be read as an attack, and risk an enterprise contract. Because the vendor is itself a large buyer, its own suppliers face the same constraint. Suresh frames this explicitly as a coordination problem: cooperating preserves jobs, defecting risks being fired by embarrassed peers and replaced by someone who will not defect, and simultaneous honesty cannot be coordinated. He reports that board members at S&P 500 companies to whom he presented admitted skepticism while saying their positions depended on demanding AI investment.
Relabelling of non-AI work
Suresh argues that a substantial share of the apparent surge in AI projects consists of non-AI projects with an AI element added to pass an internal test. His worked example is an Oracle-to-Snowflake database migration in which the vendor added a preliminary phase to have a model translate SQL dialects; the phase failed on permissions grounds, the translation was completed by hand, and the work was reported upward as an AI-driven success because a small portion had been machine-translated first. He states that projects driven by a model as the sole mechanism, which can therefore visibly fail against specific numbers, are rare outside startups.
He also reports that companies have publicised hiring policies requiring staff to demonstrate they attempted to use AI before requesting headcount, and adds that reporting an attempt that did not suffice risks being labelled "bad at AI."
Recommendations
The essay closes with practical advice in two registers. For readers pursuing an unrelated objective inside a captured organisation, it recommends raising concerns one-on-one rather than in groups; using an anonymous 1–10 poll of project success chances, where the bimodal result observed (roughly half rating 3/10 and half 8/10 on a project three years late) can demonstrate to a chief executive that information is being withheld; involving front-line users, which in one case revealed staff were unaware they had been issued AI licences at all, undermining the associated productivity claims; and not contesting broad claims such as "AI is changing everything" before trust is established. For readers simply trying to remain employed, it advises accepting that meaningful pushback is unlikely, limiting AI news intake, and beginning a job search early when facing large volumes of machine-generated code review or management communication.
Reception
Cory Doctorow summarised the essay's argument in a Pluralistic post on 1 August 2026, quoting the 0% success figure and the Snowflake Cortex example; that post is how the essay entered wider circulation two weeks after publication (Source: pluralistic.net).
Relationships
- supports: AI Bubble Debate — first-hand consulting account of enterprise AI project outcomes
- supports: Professional Services AI Adoption — describes buyer behaviour and vendor incentives in professional-services engagements
- related: AI and Productivity — argues self-reported enterprise productivity gains are largely unreliable
- related: Shadow AI — documents the inverse, disclosure-driven distortion ("AI-washing")
- related: Snowflake — Cortex is the named product in the demonstration example