Enterprise AI claims, one quarter on: what held up and what aged

This publication runs on an unusual constraint: every article registers exactly one falsifiable claim, and each claim is tracked on a public 30 to 90 day review cycle, marked Holding, Partial, or Not holding as it ages. One quarter in, the ledger holds 236 claims. That is enough to ask a question most commentary never has to answer: of the confident things written about enterprise AI, which ones survived contact with the following months?

The headline split, as recorded at /holding/: 222 Holding, 8 Partial, 6 Not holding. Roughly one claim in seventeen has moved off its original verdict. That number is less interesting than the shape of the movement, because the failures are not randomly distributed. They cluster, and the cluster is instructive.

What broke: the framing, not the analysis

All six retracted claims share one property. The structural problem lived in the article’s slug, not in its argument. Each was an inheritance from the publication’s WordPress era, before the claim-tracking standard existed: a slug built around a specific unverifiable dollar figure, a fictional protagonist who never existed, a category-average return rate presented as if it were a benchmark, or a dramatised event framing that the body itself later had to disown.

In several of those cases the underlying analysis was defensible, or was rewritten against audited sources and moved back toward Holding before the slug forced a full retraction. The lesson is not that the ideas were wrong. It is that a headline making a specific, sensational, untraceable promise is a standing liability regardless of how sound the body beneath it is. A reader, an auditor, and a search engine all read the promise first.

That last reader is worth naming. In each retraction the correction log records the same secondary observation: the search engine’s own quality systems independently flagged the same URL the publication had already moved to retract. The correlation is not proof of a mechanism, and it is logged as correlation rather than cause. But the direction is consistent enough to state plainly: the framing that failed the publication’s internal standard is the same framing that failed externally. Hype register is not only an editorial preference to avoid. On this evidence it behaves like a measurable quality signal.

What aged fastest: pricing and model tiers

The eight Partial claims point at a different, faster-moving hazard. The most common reason a claim slipped from Holding to Partial was not a flawed thesis. It was pricing and model drift.

One claim rested on a vendor’s premium subscription being a single high tier; within weeks a lower tier appeared beside it and the premium model version had already advanced. Another rested on two named tool prices that both moved before the next review. In each case the analytical spine held. The numbers underneath it did not. The correction log updated the figures and kept the verdict honest, but the speed of the drift is the point: a claim anchored to a specific price or model name in this market has a half-life measured in weeks, not quarters.

The rest of the Partial set is the ordinary work of a young ledger: claims registered retroactively against older articles, where the spine is defensible but per-figure verification is deferred to the next scheduled review. They are marked Partial precisely so the deferral is visible rather than hidden.

What this means for tracking your own AI claims

The reader of this site is usually the one writing the internal memo, the board update, or the vendor business case. The same discipline that produced this ledger transfers directly to that work.

First, date every AI claim and attach a re-check cadence to it. A statement about a model’s capability, a vendor’s pricing, or a benchmark result is not a fact with a long shelf life. It is a dated observation. The claims that aged worst here are the ones that read as permanent when they were really snapshots.

Second, treat pricing and model-tier specifics as the most perishable thing in any AI document. If a recommendation turns on a particular subscription tier or a particular model version, assume that anchor will move within the quarter and write the recommendation so it survives the move.

Third, separate the framing from the analysis when you audit your own past calls. The failures here were almost never in the reasoning. They were in headlines and figures that promised more specificity than the evidence could carry. A claim that survives is usually one whose framing was already modest enough to be true.

The full ledger, including every correction with its date and reason, is public at /holding/. It is the one part of this publication designed to be read when it is wrong.