Leadership1 publisher2 min readPublished
An AI company chief executive demotes data governance to the second step of adoption
An Entrepreneur contributor who runs a data and AI company says he still believes enterprises need clean data and real governance. He now puts a decision about how staff think alongside AI ahead of it.
The Board Room · Leadership desk
What happened
- An Entrepreneur contributor who runs a data and AI company writes that after a decade of telling clients their AI problem is a data and governance problem, he now ranks data foundation as the second step.
- He points to MIT Media Lab research on cognitive debt that he says found participants using only ChatGPT showing the least EEG activity and brain-only participants the strongest.
- A Microsoft and Carnegie Mellon report he quotes found that the more confidence someone had in an AI answer the less critical thinking they did, with trust in the tool and trust in one's own judgment moving in opposite directions.
- His worry is a machine mean, where routine acceptance of first drafts pushes a company toward the model's most probable output and costs it variance, dissent and unconventional ideas.
Compiled by The Board RoomSomething wrong?How this is made
Why it matters
- cost If the expense lands on judgment before it lands on tokens, it is paid by the manager reviewing work and never appears in the AI line finance signs off.
- constraint The reordering keeps every day of the data programme and puts a prerequisite in front of work the author says most enterprises are already years behind on.
- decision A firm funding governance this quarter has to choose between holding that work while it writes rules on review and dissent, or running both and accepting first drafts in the meantime.
- contradiction The column asks for a thinking policy as step one and breaks off before specifying its contents, so a reader who accepts the sequencing reaches the board with a step one and no contents for it.
Token spend, GPU costs, model licensing, migration off outdated infrastructure and training for people are the items the column lists as the usual answer when someone asks what AI costs [17]. All of them have owners and invoices. The step the column puts in front of them has no owner and no invoice. The data work, meanwhile, stays where it was: the author writes that enterprises need clean data, real governance and operating models built for AI, and that most are years behind on all three [16].
Count the positions and the data programme lands third; a company following that order does its first data work as its third action [7]. Acknowledgement of the change is what he calls the zeroth step [6]. Deciding, deliberately, how people are allowed to think alongside AI is the first [5]. Data foundation, which he used to describe as the starting point, comes after both [4].
Both studies the column leans on measure individuals, and the author says as much: the cost is an individual-level one, and it shows up in judgment long before it shows up on a token bill [11]. The move from there to an enterprise mean is his argument. He gives it two routes. In the first, effort shifts from creation and execution to verification and integration of AI responses, so the baseline everyone works from is the model's [12]. In the second, which he calls an artificial hive mind, separate large language models converge on their outputs to open-ended questions [13].
The piece runs as contributor opinion, and Entrepreneur notes that the views expressed are the contributor's own [1]. It names the direction of the loss it describes and leaves the size of it open. That loss sits in the tail: variance, dissent and what he calls "weird, unlikely, correct-anyway ideas" [14].
The column breaks off mid-sentence before it says what a thinking policy contains, at "So, the first step is deciding, deliberately, how" [18]. It does leave a test for whether a firm has skipped the step. "Bolting AI onto existing thinking is the enterprise version of accepting the first AI draft," the author wrote [15].
What to watch
- Whether the MIT Media Lab EEG result holds up in a larger replication, since the column's individual-level case rests on it.
- Whether any firm publishes an actual policy on how staff review AI drafts, with a named owner and a review threshold, so the first step becomes auditable.
- Whether clients accept the reordering when it means holding a data programme they have already funded.