A new essay from pivotal.substack.com argues that data quality isn’t a property of the data at all, but an emergent, use-case-dependent ladder — a framework that cuts straight to the pricing disputes roiling the training-data trade.
Every buyer of training data has run into the same wall: ask six vendors, six annotation shops, or six in-house data leads to define “quality,” and you get six incompatible answers, then hand the same corpus to those same six people and get six different scores. That definitional mush sits underneath every negotiation between frontier labs and the annotation shops, synthetic-data vendors, and data brokers who sell into them — because if nobody agrees on what quality is, nobody can agree on what it’s worth. It’s this exact problem that pivotal.substack.com takes aim at in the first of a two-part essay published July 10, 2026, arguing that the standard definitions — ISO 8000’s “data that meets its stated requirements,” ISO 25012’s fifteen-attribute checklist — are either tautological or incomplete.
the elephant problem
The essay’s opening move is to strip data of any inherent worth. “I begin with an assertion: data has no innate quality. Quality is a purely emergent phenomenon, conditional entirely on use case,” the author writes, adding, “In that essay, I argued that data has no intrinsic value; instead, the value of data is the value of what can be done with it.” That’s a direct challenge to any pricing model — for training corpora, labeled datasets, or licensed text — that treats “clean” or “accurate” as a fixed, transferable attribute rather than a claim that only holds relative to a buyer’s specific application.
a ladder, not a checklist
The essay’s organizing device is four ordered, dependent levels: granular (unit-level accuracy, provenance, well-formedness), aggregate (corpus-level coverage, deduplication, representativeness, drift), fitness-for-purpose (does the data actually answer the question, can you legally and operationally use it), and business-outcome (did it change a decision, was that change worth it). “Quality is a ladder. The lower rungs enable the higher ones; the higher rungs justify the lower ones,” the essay states — and crucially, a dataset can ace the first three levels and still fail the fourth, because instrumenting how the data actually gets used sits outside the dataset itself.
what incumbents already know
None of this framework emerges from a vacuum. ECCMA, the nonprofit that led development of ISO 8000, has spent years certifying “Master Data Quality Managers” on the premise that quality master data is simply data that “meets stated requirements,” according to a 2018 release on Newswire.com — precisely the tautology pivotal’s essay singles out. The UK government’s own Data Quality Framework, per GOV.UK, already leans toward “fitness for purpose” as the practical anchor, while warning that “data quality is more than just data cleaning.” And the stakes of getting this wrong are not abstract: a 2025 IBM Institute for Business Value report found that 43% of chief operations officers name data quality issues as their top data priority, that over a quarter of organizations estimate losses exceeding $5 million annually from poor data quality, and that 7% report losses of $25 million or more, according to IBM. The same IBM research ties this directly to AI adoption, with 45% of business leaders citing data accuracy or bias concerns as a leading barrier to scaling AI initiatives.
the pricing implication
For the training-data economy specifically, the ladder framework reframes a question that annotation shops, RLHF vendors, and data brokers have been fudging for years: what exactly are frontier labs paying for when they buy “high-quality” labeled data? If pivotal’s argument holds, granular cleanliness — the thing most quality-assurance pipelines actually measure and the thing IBM notes gets described via “dimensions such as data accuracy, completeness, timeliness and consistency” — is necessary but insufficient. A perfectly labeled, deduplicated, well-provenanced dataset can still be worthless if it doesn’t move an eval score or change a model’s downstream behavior, which is the business-outcome test that IBM’s own data suggests most enterprises aren’t equipped to run, given that 79% of organizations are adopting AI agents, per a PwC survey cited by IBM, often without the instrumentation to trace value back to the data itself. That gap between granular quality (cheap to certify, easy to checklist) and business-outcome quality (expensive to prove, easy to dispute) is exactly where pricing power in data licensing deals gets contested — and where sellers who can only demonstrate the former will keep losing negotiating leverage to buyers who insist on the latter.
The second essay in pivotal’s series, promised for the following week, turns explicitly to “data quality in an AI world” — the piece worth watching for anyone pricing training corpora, since it’s where the abstract ladder presumably meets synthetic data, RLHF labeling, and the actual mechanics of what labs will pay per token. If the framework travels intact from theory to term sheet, expect it to sharpen — rather than resolve — the argument over what “quality” data is actually worth.
I begin with an assertion: data has no innate quality. Quality is a purely emergent phenomenon, conditional entirely on use case. Readers of How to Price a Data Asset will recognize this line of thinking. In that essay, I argued that data has no intrinsic value; instead, the value of data is the value of what can be done with it.
