Nobody ran an AI rollout at Buyers Edge. Its 1,200 employees did it themselves.

Take the per-seat rate at both ends, multiply by 1,200 heads and twelve months, and full coverage costs between $288,000 and $576,000 a year ^[11]. That is the visible half of the bill and the easier half to manage, because a per-seat invoice at least arrives in one place. Duplication does not. When one employee swears by one model and someone two desks away insists a competitor is better at the same task [c3b], the second licence is a line on somebody's expense claim, not a decision anyone made.

The accuracy problem generates no invoice at all. John Davie's account is that nobody was tracking outputs, their accuracy, or how they surfaced in critical business decisions ^[4]. He reaches for a global figure of $67.4 billion in losses from consequential hallucination errors in 2024, attributed to a Four Dots report citing AllAboutAI ^[9], which is a number to treat as an order of magnitude rather than a measurement of anything. His own version is more damning and needs no third party: as usage climbed, he asked which tools were earning their keep and how carefully people were checking what they leaned on, and nobody could answer him ^[5].

Then there is the asset that never reached a balance sheet. The refined prompts and the workarounds that actually moved work along lived and died inside third-party tools, with no way to capture them or pass them around ^[7]. That work was done on company time, on company output, and stored in an account the company does not hold. If his workforce resembled the population Microsoft and LinkedIn surveyed, roughly 936 of the 1,200 would be working through personal accounts ^[12]. That is an extrapolation across two different populations rather than a count of his own staff, and he does not claim otherwise.

Read the closing third with the byline in view. Davie says nothing offering the flexibility and oversight he wanted was available on the market [c6b], so his organisation built CollectivIQ, which he also runs ^[1][10]. The diagnosis survives the conflict of interest; the market conclusion is the oldest premise in enterprise software and should be checked against actual vendor pricing before anyone repeats it. What is worth keeping is the smaller finding. The persuasion work leaders had budgeted for never arrived ^[3]. The governance work they had not budgeted for is now the job, and it starts from a position where the employer knows less about its own tooling than any individual employee does.