Chequeado: Nunchi, an agent that reads the comments no human could

Problem

Chequeado watched metrics but was not reading the comments on its website or social media. The scale made manual reading impossible at roughly 10 Instagram posts a week, averaging a thousand comments each; around 10,000 comments a week from Instagram alone. Website comments are far fewer but "valuable" and need curation. The value in the comment section went unextracted.

Solution

Nunchi listens to, organises and prioritises audience conversations and turns them into editorial insights. Comments are collected by n8n (an open-source workflow automation tool), sent to an AI agent on the OpenAI Platform for categorisation, and stored in a PostgreSQL database. The results surface as Slack alerts and a dashboard, used both to check the AI's categorisation and for broader analysis. This includes spotting how one topic is generating political debate while another draws positive engagement.

Each alert summarises the signal, suggesting an editorial action such as adding a methodological clarification in the comments. It includes real comments alongside the AI's reading, which lets staff check the system against its source and learn to trust it. An external n8n specialist set up the initial workflow while training the team to support the tool themselves. Free and self-hosted, n8n runs on Chequeado's existing cloud. The product team took it from ideation to an implemented Minimum Viable Product (MVP).

Results

Initially worried about AI producing simplistic categories for Spanish-language comments, the team received positive results from the tool. One or two pieces have been published off the back of it, correcting or clarifying something the comments flagged.

Insights

The early trap was trying to do too much by extracting too many aspects from each comment. Narrowing down to fewer dimensions helped the team achieve better progress. Another lesson is timing. Defining the right moments to involve the newsroom, balancing showing a prototype too early against too late for useful feedback.

No one is assigned to verify every categorisation; the team keeps an eye on it and reads the raw comments when editorial flags a bad output. That lighter validation is accepted because the output is internal-only. Anything audience-facing would require stricter processes. The roadmap points into using more platforms as data sources, new use cases on the accumulated database such as trends and impact measurement, adoption in other areas of the organisation, and possibly opening the tool to other newsrooms.

Cost

Application Programming Interface (API) usage runs roughly USD 50–100 a month at current comment volume, expected to rise if other platforms are added. The n8n licence costs nothing self-hosted. Internal build time was roughly 10 to 15 hours a week.

Replicability

The stack is inexpensive using open-source collection, a consumer AI API, a database and Slack. Narrowing the scope is important. Over-scoped categorisation is the predictable trap, and raw comments beside AI output are what build newsroom trust. The categorisation Chequeado settled on is based on sentiment in six registers (analytical, critical, frustrated, hostile, ironic and positive), intent, and signals including support for the outlet, criticism of it, and conspiracy theories, with a reviewer field built into the interface. The build pays off mainly for newsrooms whose comment volume is beyond manual reading; volume and platform mix set both the value and the running cost.