AI Feed: browsing the watch by what each signal is about, not by its words

What is the difference between filtering by words and browsing by concepts?

A filter answers ‘what comes into this watch?’. AI Feed answers a different question: ‘of everything that has come in, what is there about this technology, this entity or this type of event?’.

The difference matters because the same reality is named in many ways. A technology appears under its technical name, its acronym, a manufacturer’s trade name and its translation; a company appears under its registered name, its brand and the name of its subsidiary. The information retrieval textbook by Manning, Raghavan and Schütze therefore distinguishes between rules written by hand, such as Boolean queries, and automatic classification, which learns its criteria from labelled examples^[1]. Rules give control over the perimeter; classification orders what has come in, even when each text says it in its own way.

How does the AI identify each dimension?

Each dimension corresponds to a known task in natural language processing, with its typical way of going wrong. Knowing it helps you interpret what the view shows.

What is it for in daily work?

• Preparing a meeting with a competitor or partner: seeing at once everything that has come in about that entity, even though each outlet names it differently.
• Reviewing a technology over a period: what has been published about it in the last quarter, by type of event.
• Isolating a type of event in a place: for example, regulatory changes in one country, or agreements between companies in one region.
• Reading the tone of the conversation on a topic, knowing that it is an approximation and not a measurement.

These are illustrative examples of possible routes, not predefined views.

When not to use AI Feed

• As a substitute for the filter. The perimeter of each watch is defined by people; a dimension of the AI does not guarantee that everything relevant has come in.
• As a substitute for validation. A signal being correctly classified does not mean it matters to your organisation.
• As statistics. Counting signals by tone or by technology describes what was captured, not the market; it depends on which sources are watched.

If the AI gets it wrong: source and judgement

NIST’s profile for generative AI (NIST AI 600-1) describes confabulation, the production of erroneous content expressed with confidence, and automation bias, which leads to excessive trust in what the system proposes^[3]. The European AI Regulation (Regulation (EU) 2024/1689) has required, since 2 February 2025, those who use AI systems to take measures to ensure AI literacy among their staff (Article 4)^[4]. Regulation (EU) 2026/1744, in force since 27 July 2026, kept that obligation but no longer requires a specific ‘sufficient’ level^[5][6]. In practice, the criterion is the same: whoever uses the classification must know how it can go wrong.

In Innguma, each signal links to the publication it comes from, and people decide what is watched, what is published and who receives it. The AI processes. Your team decides.