Factory: where watch signals are processed, classified and validated

Which stage of the watch process does Factory cover?

Within the operation of the system, UNE 166006:2018 distinguishes source control, search, processing, value-adding, distribution and storage^[1]. Factory brings the first four together in one place:

1. Capture. The sources of each watch are monitored continuously: news and websites, patents, regulation and Spain’s Official State Gazette (BOE), tenders, competitors’ websites and specialist sources.
2. Filtering. Each watch’s criteria decide what comes in and what stays out.
3. AI processing. Innguma’s AI groups duplicates and classifies each signal by category, technology, entity, type of event, place and tone.
4. Review and validation. The analyst discards the noise, validates what is relevant and records why it matters.

What is validated goes to the Library, which keeps it with its source and from which newsletters are sent.

Why group duplicates before reviewing?

A company’s press release or an official announcement is reproduced in dozens of outlets, with minimal changes to the headline or paragraph. The information retrieval textbook by Manning, Raghavan and Schütze notes that, by some estimates, up to 40% of web pages are duplicates of others^[2]. If each copy arrives as a separate entry, the analyst reads the same thing several times, and the genuinely new news item is buried.

Not all repetitions are alike, and it is worth distinguishing them because they are detected in different ways:

The third case is the most useful for the analyst and the most delicate: grouping well avoids reading the same agreement three times; grouping badly can hide a nuance that only one of the outlets provides. That is why grouping organises, but each publication’s source remains available.

What does classification add when reviewing?

Classifying each signal by technology, entity, type of event, place and tone makes it possible to review in blocks rather than in order of arrival: first regulatory changes, then competitors’ moves, then everything else. It also helps detect gaps, such as a technology that has not had a new item in weeks.

Classification is an aid, not a verdict. In information retrieval a distinction is drawn between rules written by people, which are transparent but costly to maintain, and automatic classifiers, which learn their criteria from examples and scale better^[3]. Each makes mistakes in a different way, which is why Factory combines explicit filters with AI classification and leaves the final word to a person.

Why is validation done by a person?

NIST’s profile for generative AI (NIST AI 600-1, July 2024) describes confabulation as the production of erroneous or false content expressed with confidence, and also warns of automation bias: the tendency to place excessive trust in what the system proposes^[4]. In watching, the natural control point is review before publication: this is when someone with knowledge of the business checks the source and decides whether the signal changes anything for the person who will receive it.

At Innguma, the AI processes; people decide what is watched, what is published and who receives it.

How to review well in Factory

1. Review with the information needs in front of you. The question is not ‘is it interesting?’ but ‘does it change anything for whoever will receive it?’.
2. Start with the types of event that weigh most. A published rule or a patent granted to a competitor usually has more consequences than an interview.
3. Open the source when in doubt. Each signal links to the publication it comes from; the original resolves what a headline leaves ambiguous.
4. Write down the why when validating. A one-line comment (‘affects the packaging line; alert Quality’) is kept with the signal in the Library, and gives context to whoever reads it months later.
5. Fix the source of the noise. If you always discard the same thing, the adjustment lies in the filter or the source, not in reviewing faster.