This section covers the AI stories where the number, the claim or the announcement needs checking before it is repeated. We read the primary document, say what it actually measures, and mark how far each claim can be trusted.
Every factual claim in this section falls into one of three evidence levels. Where an article mixes them, the wording makes the level clear instead of leaving it for the reader to guess.
| Level | What it means | Typical source |
|---|---|---|
| Verified | We read the original document or data and the claim holds as stated. | Filings, regulatory texts, published datasets, vendor documentation, peer-reviewed papers |
| Reported | A credible outlet or company states it, but we could not check the underlying material ourselves. | News reporting, executive interviews, press releases |
| Estimated | A projection, survey extrapolation or modeled figure. Method and sample size matter more than the number. | Analyst forecasts, market-size reports, vendor-sponsored surveys |
Corrections are made in the article itself with a dated note. For the full process, including how we choose sources and handle disagreements between them, see our methodology page.
These are the checks we run before a figure appears in an article. They work just as well on a slide deck, a press release or a social post, and none of them requires specialist knowledge.
A single survey is a data point, not a trend. We treat something as a trend only when independent sources with different methods point the same way, and we say so when they do not.
The section sits between the more specialized categories on the site. When a story has a clear home elsewhere, we link to it instead of duplicating it. These are the neighboring areas most often connected to news coverage here.
If a number in an article needs translating into budget terms, the AI Cost Calculator and the AI Model Matrix let you test it against your own usage. Unfamiliar terms are defined in the AI Glossary.
When there is something worth reporting, not on a fixed schedule. Each article shows its publication date and, when it changes materially, an update date. We would rather publish less often than repeat an unchecked claim.
No. We cover launches when they change what readers can do, what it costs, or what the evidence says. Routine version updates are better handled by the vendors’ own release notes, which we link to when relevant.
Mostly because of definitions and samples. One survey may count any employee who has tried a chatbot, another only systems running unattended in a business process. Our analysis of AI agent adoption statistics walks through a concrete case.
Sometimes it is the only data available. When we use it, we label the funding source and the evidence level, and we avoid presenting it as independent confirmation.
Use the contact page. A link to the primary source is the most useful thing you can include with a correction.
Five reputable sources, five different percentages, one phrase — “in production” — doing all the lying. By Tom Morgan · Updated September 8, 2026 · 13 min read THE SHORT VERSION: “AI agents in production” gets quoted anywhere from 11% to 80% because each number answers a different question — pilot interest, app-level embedding, or … Read more