Token Leaderboards and Blind Mandates: AI's Hidden Governance Risk

A post from commentator Simon Willison has been circulating widely this week, curating anecdotes from consultant Nik Suresh's essay "AI Mania Is Eviscerating Global Decision-Making". Suresh, who runs the Ludicity blog and works at the data consultancy Hermit Tech, describes patterns he says he has observed across roughly 300 client conversations. The essay is framed as opinion and anecdote, not a study — but the specific incidents it describes map cleanly onto governance failures that security and compliance teams are already supposed to be watching for.

Strategy without hands-on knowledge

The anecdote getting the most attention: an executive responsible for a business unit with more than $2 billion in revenue reportedly admitted to never having used ChatGPT or any AI tool personally, immediately after producing a technical strategy document built entirely around AI adoption. Suresh also describes organisations running internal "token leaderboards" — dashboards that reward employees for raw AI token consumption — which he says pushes staff toward performative prompting loops rather than verified, useful output.

A third detail is the more serious one for anyone thinking about governance: employees, under pressure to demonstrate AI usage, reportedly relabel work they did themselves as AI-generated output to satisfy the mandate.

Why this is a governance problem, not just a culture problem

None of this requires a jailbreak, a prompt injection, or a compromised model to cause damage. It's a provenance and accountability failure that happens entirely inside the organisation's own reporting chain:

• Strategy set without capability. Decision-makers directing AI adoption without first-hand experience of the tools' failure modes are poorly positioned to size the actual risk they're signing off on.
• Metrics that reward the wrong thing. A token-volume KPI creates an incentive to *use* AI performatively rather than to use it *correctly*, which is the opposite of what a usage-monitoring control is supposed to achieve.
• Broken provenance. If staff are mislabeling human work as AI output (or vice versa) to hit a mandate, any later audit of what an AI system actually touched — a requirement under frameworks like ISO/IEC 42001 — is built on false data from day one.

Suresh also describes a vendor-side dynamic worth noting: customer executives promoting implausible productivity gains, and vendor staff who privately doubt those figures staying quiet because contradicting a customer's narrative risks the account. That pressure loop discourages the kind of honest capability assessment that a governance programme depends on to function.

What this means for security and governance teams

None of these are hypothetical, and none of them are new failure modes — they're the same shadow-IT and unverified-metric problems security teams have chased for two decades, now wearing an AI label. The practical response is the same as it's always been: an AI usage policy that specifies what tools are approved and for what data; usage metrics that measure verified, reviewed output rather than raw volume; and a governance record — mapped to a framework like ISO 42001 — that can actually be trusted when someone asks what an AI system did and on whose authority. A mandate handed down by leadership that has never used the tools it's mandating is, on its own, a signal worth escalating.