How AI is used
The record is built by an automated pipeline in which a language model writes the entries. This page explains what that means for what you read here.
The “AI generated” label
Pages that show entries carry this icon with a short label, and so does the footer of every page. It is the icon published by the European Commission for labelling AI-generated content, used here as provided for by Article 50 of the EU AI Act. It marks text written by AI that is not reviewed by a person record by record.
What the AI does
Daily or weekly, the pipeline downloads the vendor sources listed on Sources and keeps only the parts that changed. A language model (Claude, by Anthropic) reads those parts and proposes entries in a fixed format:
- the feature name and a short, factual summary, in its own words;
- the release state (announced, preview, beta, GA, deprecation, retirement), the vendor's date and its precision, and any date the vendor promised;
- the capability it belongs to, from the published taxonomy, and its significance (major, notable or minor);
- market signals (partnerships, acquisitions, funding, leadership changes) when a source reports one;
- a doubt flag when the source is unclear, which publishes the entry as under review.
The model works only from the source text it is given. It is instructed to state facts without opinion or marketing language, and every entry links to the source it was read from.
What the AI does not do
- It does not choose the sources, the platforms or the taxonomy. These are decided by the author.
- It does not publish on its own: every proposal must pass the automatic checks below.
- It does not write opinion, rankings or recommendations. Interpretation, when there is any, is on the author's site and is labelled as his view.
How the output is checked
- Automatic rules on every record before publication: data format, a capability from the current taxonomy, dates consistent with their precision, states that only move forward, sources on an allowed vendor domain. A record that fails is not published. Details in Methodology.
- No per-record human review. The author reviews in batches, not every record. A record without corrections has not necessarily been checked by a person.
- Accuracy audits on random samples, checked one by one against their sources; the latest results are in Methodology.
- Doubts stay visible: entries under review are published with their source but left out of the Release Matrix and side-by-side views.
- Corrections: every change to a published fact reviewed by the author is listed in Corrections, with the value before and after.
- Anyone can report an error with the “Report an issue” button on every record; reports are checked against public sources.
The model's own confidence score is not published: it is a self-assessment, not a calibrated measure.
What is not AI-generated
The counts, charts and coverage figures are computed from the records by fixed rules, not written by AI. The explanatory pages (this one, Methodology, About, Sources, Status, Taxonomy, Corrections, Privacy) and the site's code were drafted with AI assistance and are reviewed and published by the author, who is responsible for them; they carry no label on the page.
What the AI receives
The pipeline's model receives only public vendor pages and the author's public MarTech Watch. Nothing about readers is sent to it.