[Skip to content](https://www.innguma.com/en/why-innguma-ai/#contenido)

Watching + AI

# Watching and intelligence in the era of LLMs: from the ephemeral answer to the lasting asset

**In short** 

Innguma does not compete with ChatGPT, Claude, Microsoft Copilot or Gemini: it gives them the governed data they need. An AI assistant searches, summarises and writes; Innguma keeps the watch as knowledge infrastructure, with declared sources, continuous capture, human validation, linked sources and history. Through the Model Context Protocol (MCP), those four assistants work on that knowledge and, if your organisation enables it, govern the watch cycle. AI answers. Watching expands the context.

-   **Analyst: judgement and decision** Interpretation, expert validation and strategic decision.
-   **AI: analysis and processing** Classification and prioritisation in Innguma; summary and drafting with the organisation's AI.
-   **Innguma: infrastructure and governance** Systematic capture, traceability, storage and governed distribution of the data.

Comparison of approaches

|  | Generalist AI agent | Innguma + AI |
| --- | --- | --- |
| Data traceability | Variable | High |
| Structured history | Low | High |
| Compliance (National Security Framework, ENS / GDPR) | Low | High |

Qualitative assessment based on the analysis in this document.

## Can an AI assistant replace a watch system?

Current AI assistants search the web, cite sources, summarise and can prepare something similar to a newsletter. To deny it would be a mistake. The error lies in comparing a visible output, the newsletter, with a complete watch system. Before the newsletter there must be defined information needs, a source map, a capture system, filtering and classification criteria, expert validation, a repository, distribution to each recipient and continuous improvement. The UNE 166006:2018 standard formalises that process as a management system.[\[1\]](https://www.innguma.com/en/why-innguma-ai/#ref-1)

## Searching is not watching

A search answers a question; watching sustains continuous observation. An agent rewrites the query on each run and depends on what its search engine returns at that moment, so two runs of the same brief can give different results. A configured capture reviews the same sources continuously under the same rules: that provides consistency and makes it possible to know what was watched, on what criteria, and what was discarded.

## From the ephemeral answer to the lasting asset

An AI produces a useful synthesis at a given moment. If that synthesis stays in a chat or an email, it is lost as structured knowledge. A lasting asset means that the watch stops depending on one person, one prompt or one-off search, and becomes part of a common base: declared sources, validated signals with their link to the original, tags, comments and history, which remain in the organisation even when teams or AI models change.

## The black box of sources

Was the full summary of Spain's Official State Gazette (BOE) consulted, or only the news items that mention it? Was EUR-Lex or a press summary reviewed? Were Espacenet or PubMed entered? Can the same search be repeated in three months and give the same result? These sources are systems with their own structures and formats: the BOE, for example, offers its daily summary and consolidated legislation through an open-data API in XML and JSON.[\[2\]](https://www.innguma.com/en/why-innguma-ai/#ref-2) Integrating them operationally and sustainably, and being able to say which sources each watch covers, is watching work, not writing work.

## What does public evidence say about AI assistants?

The problem appears when a model is asked to be at once a search engine, a capture system, a repository, a source of truth, a classifier, a writer and a distribution channel. Independent studies are specific:

-   The European Broadcasting Union (EBU) and the BBC evaluated in October 2025 more than 3,000 responses from ChatGPT, Copilot, Gemini and Perplexity, in 14 languages and with 22 public broadcasters from 18 countries: almost half had at least one significant problem, and around a third had serious problems of source attribution.[\[3\]](https://www.innguma.com/en/why-innguma-ai/#ref-3)
-   The Tow Centre at Columbia University published in March 2025 a test of 1,600 queries to eight AI search engines about articles from 20 publishers: taken together, they answered incorrectly more than 60% of the time.[\[4\]](https://www.innguma.com/en/why-innguma-ai/#ref-4)
-   OWASP ranks prompt injection as the first risk in its Top 10 for LLM applications (2025), including indirect injection: hidden instructions in websites or documents that the model processes.[\[5\]](https://www.innguma.com/en/why-innguma-ai/#ref-5)
-   NIST describes confabulation, erroneous content expressed with confidence, and automation bias, which leads to excessive trust in what the system proposes.[\[6\]](https://www.innguma.com/en/why-innguma-ai/#ref-6)

## AI risks and how they are controlled

Innguma also uses AI, to filter, group duplicates and classify. The problem is not using AI, but using it as a source of truth without control.

| Risk | How it shows up | Control with Innguma |
| --- | --- | --- |
| Hallucination | The AI concludes without verifiable evidence. | Each signal links to the publication it comes from, and the connected AI works on what your team has validated. |
| Source omission | The agent only sees what its search prioritises. | The source map is declared, configured and reviewed with each organisation. |
| Attribution error | The AI cites wrongly or mixes sources. | The signal keeps the link to its original source. |
| Indirect injection | A website or document alters the agent's behaviour. | What comes in goes through filters and human review; what the connected AI can do depends on the capabilities your organisation enables. |
| Loss of memory | The result stays in chats or emails. | What has been validated stays in the Library, with tags, comments, source and history. |
| Data exposure | Sensitive queries in ungoverned tools. | Innguma does not train models on its customers' data, and processes data in line with the GDPR. |

## Security, data use and compliance

Three levels are worth distinguishing. In consumer AI, data may be used to improve the models depending on the user's settings: Anthropic, for example, asked users of its Free, Pro and Max plans in 2025 to choose whether their conversations are used for training, and excluded its commercial services and the API from that change.[\[7\]](https://www.innguma.com/en/why-innguma-ai/#ref-7) Enterprise AI reduces that risk with contractual guarantees, but does not by itself solve capture, the source map, the history or distribution. Enterprise AI connected to Innguma combines those guarantees with a governed watch layer: Innguma's watch and intelligence system is certified under UNE 166006:2018, and the systems supporting its managed watch services are certified under the ENS at MEDIUM category.

The European AI Regulation sets the timetable: AI literacy has been mandatory since 2 February 2025; the transparency obligations of Article 50 apply from 2 August 2026; and, after the Digital Omnibus on AI, those for high-risk systems in Annex III move to 2 December 2027 and those in Annex I to 2 August 2028.[\[8\]](https://www.innguma.com/en/why-innguma-ai/#ref-8)

## Govern the watch cycle from your AI assistant

Innguma offers connectivity through the Model Context Protocol (MCP), the open standard for connecting AI applications to external systems,[\[9\]](https://www.innguma.com/en/why-innguma-ai/#ref-9) currently compatible with ChatGPT, Claude, Microsoft Copilot and Gemini. With the capabilities that Innguma makes available to each integration, and depending on what your organisation enables, from the assistant you can:

1.  Query the news in the inbox and ask for any work on them: summarise, compare, prioritise or prepare a briefing.
2.  Organise and add sources, both from Innguma's source library and new ones.
3.  Modify the filters of the watches.
4.  Audit the performance of sources and obtain suggestions for improvement.
5.  Publish content automatically, with summaries, insights, tags or additional fields.
6.  Edit or delete published news items.
7.  Search across published content.

What is automated, and what goes through human review first, is decided by your team. It is the criterion recommended by the MCP specification itself: that there is always a person able to deny the execution of a tool, and that confirmation is requested for sensitive operations.[\[10\]](https://www.innguma.com/en/why-innguma-ai/#ref-10) It also matches OWASP's recommendation to limit the model's privileges to what is necessary.[\[5\]](https://www.innguma.com/en/why-innguma-ai/#ref-5)

## The right architecture: three layers

Innguma acts as the layer of capture, governance, linked source, storage and distribution of the data. AI acts as the analysis layer: Innguma's AI filters, groups duplicates and classifies, and the organisation's AI summarises, compares and drafts on what has been validated. The analyst contributes judgement, validation and interpretation. AI does not disappear from the process: it is placed where it belongs. Your AI provides analytical capacity. Innguma provides context.

## When an AI is enough, and when Innguma makes sense

**An AI may be enough** when the watch is one-off, low-risk, with few public sources, with no need for a history or for knowing what was covered, and with no critical impact on regulatory, technical or strategic decisions.

**Innguma makes sense** when the watch must be recurring, shared by several people or areas, verifiable source by source, connected to official or specialist sources, with a history, distribution to each recipient and later use with the organisation's AI.

## A 30-day test, with no prior claims

The most honest way to compare is a repeatable exercise on EU and Spanish regulation and tenders, using official sources such as EUR-Lex, the BOE, TED (the European tenders journal) and the Plataforma de Contratación del Sector Público. The sources, the criteria and who evaluates are fixed beforehand, and at the end each approach is compared on:

1.  Coverage: how many of the relevant publications of the period were detected.
2.  Critical false negatives: which important publications were missed.
3.  Accuracy of what is delivered, and of its attribution to the source.
4.  Time between publication and its arrival with the decision-maker.
5.  Effort required from people to review and correct.

## Sources and references

1.  \[1\] [UNE. UNE 166006:2018, R&D+i management: watch and intelligence system (standard record). 2018](https://www.une.org/encuentra-tu-norma/busca-tu-norma/norma?c=N0059973)
2.  \[2\] [Spain's Official State Gazette (BOE). Open data: API for BOE summaries and consolidated legislation](https://www.boe.es/datosabiertos/api/api.php)
3.  \[3\] [EBU and BBC. News Integrity in AI Assistants. 2025](https://www.ebu.ch/research/open/report/news-integrity-in-ai-assistants)
4.  \[4\] [Jaźwińska, K. and Chandrasekar, A. (Tow Center, Columbia Journalism Review). AI Search Has a Citation Problem. 2025](https://www.cjr.org/tow_center/we-compared-eight-ai-search-engines-theyre-all-bad-at-citing-news.php)
5.  \[5\] [OWASP GenAI Security Project. LLM01:2025 Prompt Injection. 2025](https://genai.owasp.org/llmrisk/llm01-prompt-injection/)
6.  \[6\] [NIST. NIST AI 600-1, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile. 2024](https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf)
7.  \[7\] [Anthropic. Updates to Consumer Terms and Privacy Policy. 2025](https://www.anthropic.com/news/updates-to-our-consumer-terms)
8.  \[8\] [European Commission, AI Act Service Desk. Timeline for the implementation of the EU AI Act (with the amendments of Regulation (EU) 2026/1744). 2026](https://ai-act-service-desk.ec.europa.eu/en/ai-act/timeline/timeline-implementation-eu-ai-act)
9.  \[9\] [Model Context Protocol. What is the Model Context Protocol (MCP)? Official documentation](https://modelcontextprotocol.io/)
10.  \[10\] [Model Context Protocol. Specification 2026-07-28: Tools (User Interaction Model and Security Considerations). 2026](https://modelcontextprotocol.io/specification/2026-07-28/server/tools)

## Related

- [MCP · Connect Innguma with ChatGPT, Claude, Microsoft Copilot and Gemini](https://www.innguma.com/en/mcp/)
- [AI at Innguma · Monitoring with artificial intelligence](https://www.innguma.com/en/ai/)
- [What happens if the AI makes a classification error?](https://www.innguma.com/en/resources/faq/ai-errors/)
- [Intelligence Academy: open training in monitoring and competitive intelligence](https://www.innguma.com/en/resources/academy/)

[Propose a 30-day test](https://www.innguma.com/en/demo/)