Your AI Status Reports Are Politically Captured. Better Dashboards Make It Worse.

An executive who had “never even used ChatGPT or any AI tool in their life” authored the AI-centered technical strategy for an organisation with more than $2B in revenue. Nikhil Suresh, who runs the Melbourne consultancy Hermit Tech, says he watched it happen, and says the strategy moved forward because everyone downstream of it understood which answer was safe to give.

That scene anchors AI Mania Is Eviscerating Global Decision-Making, published 18 July 2026. Suresh carries an axe and shows it openly. He wrote the “piledrive” essay, his firm has rejected all AI implementation work offered to it, and he reports “0% success in a year and a half” across the AI projects he and his team observed. His base is roughly 300 catchups with professionals around the world over the life of his blog. One consultant’s field log, in other words. Read every figure here as testimony, never as measurement.

Testimony still earns attention when it describes a mechanism, and this mechanism should unsettle anyone who sells governance. It unsettles us, because governance is what we sell.

Belief became the loyalty test

Suresh names a scale threshold: “every sufficiently large business we have observed (say, with 500+ employees)” runs some version of the dynamic. Professing enthusiasm about AI protects a career. Voicing doubt endangers one. Once that price structure exists, the reporting line stops carrying information upward and starts carrying loyalty upward instead.

A career CISO he quotes, anonymous by request like several of his sources, put it this way: “there’s a cult-like atmosphere to it that you didn’t see with, say, the cloud… From talking to CISOs everywhere, I would say most of them are quietly skeptical but afraid to speak up.” Mitchell Hashimoto, of HashiCorp and Ghostty, opens the piece with a harder version: “I strongly believe there are entire companies right now under heavy AI psychosis and it’s impossible to have rational conversations with them about it.”

The block quotes in Suresh’s essay are testimony he gathered, not comments from readers. That distinction matters for how much weight they carry: these are sourced accounts, collected by a partisan, from people who asked to stay unnamed.

Four checks fail at the same time

A healthy organisation catches a bad bet through four separate paths: the honest status report, the usage metric, the procurement review, and the informal sniff test where a senior person asks around about who actually uses the thing. Loyalty pricing disables all four in one move, because all four require somebody to say an unwelcome sentence out loud.

Metrics go first. Suresh: “project leaders are very careful to avoid tracking basic metrics, such as whether the tools are being used at all, or they track metrics that are easily gamed.” He names three of the gamed variety, and each one is real enough that you have probably seen it: token leaderboards where higher spend scores better, money spent on AI as an employee evaluation criterion, and headcount requests gated on proving you tried AI first. We have argued before that agents will optimise the artifact instead of the outcome. Here the optimisation is deliberate and human, which makes it far more durable.

Procurement goes next. Suresh describes a chatbot demo his team assembled in two hours that outperformed everything the client’s technical leads had seen, including at an ASX-listed company already publicising its AI usage. Then the part that should stop a buyer cold: “every lukewarm client that saw the chatbot in action, even with us telling them that it was not going to accomplish what they wanted, wanted to buy it immediately.” Vendor incentives compound the effect, which is the structural bias we traced through AI productivity claims, and the reputational payoff for announcing AI you do not operate sits on the same shelf.

Accuracy claims survive this environment intact because nobody prices them. Suresh recounts Snowflake staff telling him verbally, from memory, that Cortex ran at roughly 92% accuracy at ideal configuration. Take the number as illustration rather than fact, and it still lands: at that rate, one in ten of a CFO’s numbers comes back wrong. A finance function would reject that error rate from a junior analyst in a week. From a platform with a keynote behind it, the rate goes unmeasured.

The reason nobody defects first

The interesting part is why the equilibrium holds even among people who privately agree. Suresh’s framing makes it a coordination failure among executives rather than a property of the technology, and that framing has teeth.

Consider a vendor executive who believes a customer’s public 100x productivity claim is implausible. Saying so implicitly calls a peer executive a liar in front of his own board. The truth-teller absorbs the entire cost of that sentence. The benefit, a market where claims mean something, spreads across everyone including his competitors. Every participant faces the identical payoff matrix, so the honest move stays permanently one step away for all of them. This is the same shape as the pace mismatch we mapped across Fortune 500 adoption, viewed from the incentive side instead of the calendar side.

Suresh also draws a distinction worth carrying into any negotiation: “you should avoid getting into business with a liar, but if you must, you can at least reason with them even if only in private. A true believer is much more threatening because they are impervious to even inducement by self-interest.” A liar responds to a changed payoff. A believer treats the changed payoff as a test of faith.

One point of his deserves emphasis because it cools the temperature. He places the root cause upstream of AI entirely, in delivery capability: “companies are terminally bad at running software projects effectively… Very few companies are so good at shipping software that they can afford the extra risk profile.” Delivery discipline is the prior condition. AI raises the variance of an organisation that already struggles to ship.

Where this leaves our own recommendations

Victorino sells governance work: assessments, control design, operating models. Every one of those deliverables assumes the inputs are honest. An assessment reads status reports. A control design consumes usage telemetry. An operating model depends on a steering committee where somebody can say “this is three months behind” and keep their job.

If the reporting layer is politically captured, a sharper dashboard produces a more precise, more confident, more expensively produced wrong answer. That is the uncomfortable conclusion, and we are going to leave it uncomfortable. Refusing the engagement solves nothing for the client. Pretending the instrument is clean is worse. What remains honest is to test the reporting layer before charging anyone for work built on top of it, and to say clearly when the test comes back bad.

The instrument that reads through capture

Suresh supplies one tool, and it is cheap enough to run this month. On a project already three years late, an anonymous confidence poll returned a bimodal result: roughly half the respondents rated it 3/10, while others clustered around 8/10.

The average of that distribution is meaningless. The shape of it is the entire finding. A team with shared information converges. A team where half the members hold a private assessment they decline to voice in the meeting splits into two peaks. Bimodality is direct evidence that information exists inside the organisation and has been withheld from the reporting line, and it produces that evidence without requiring a single person to volunteer as the truth-teller. The instrument absorbs the career cost that the prisoner’s dilemma imposes on individuals.

Do this now

Run one anonymous confidence poll on your largest in-flight AI initiative this week. One question: on a scale of 1 to 10, how confident are you that this project will deliver what was promised, on the timeline that was promised? Collect it outside the reporting line, with genuine anonymity, from everyone who touches the work including contractors.

Then read the shape rather than the mean. A tight cluster anywhere, high or low, means your organisation shares a view and you can plan against it. Two peaks means your status reports have been telling you what you wanted to hear, and every governance artifact built on those reports needs re-examination before it gets extended. Run the same poll quarterly and the trend in dispersion becomes a leading indicator of reporting integrity, which is the one metric that makes the others readable.

This analysis synthesizes AI Mania Is Eviscerating Global Decision-Making (Hermit Tech, July 2026), a single consultancy’s field observation rather than research. Every figure in it, including the “0% success” rate and the Snowflake Cortex accuracy claim, is the author’s own count or a verbally relayed report, and is presented here as testimony.

Victorino Group tests whether your AI reporting layer can be trusted before designing controls on top of it. Let’s talk.