La Silla Vacía, a Colombian political outlet known for editorial independence and investigative reporting, developed a centralised hub of AI-powered tools to improve newsroom efficiency and audience engagement. The hub hosts modular assistants tailored to specific editorial needs, such as:

• A style-editing tool (“FranBot”), aligned with La Silla’s manual but adaptable to other style guides

• An assistant to build X threads

• A tool for drafting the daily news round-up (“Duerma Informado”)

• A newsletter builder for exclusive “Superamigos” members

All aim to save time, ensure consistency, and strengthen audience connection.

The problem: Behind the idea

The idea for a centralised hub emerged during the team’s AI Lab discovery sessions, after realising that individual tools delivered in isolation often failed to integrate smoothly into editorial workflows. Multiple platforms, logins, and configurations created friction, and journalists were less likely to adopt tools that felt fragmented or required significant effort to learn.

“We saw that even when a tool worked well, it risked being forgotten or ignored if it wasn’t easily accessible,” said Karen De la Hoz, Product Manager at La Silla Vacía and project manager of this initiative.

Each of the tools of the Hub are purposefully selected, and addresses a concrete problem identified through user research and internal feedback. As De la Hoz explains, “Creating X threads was a pain point. One editor said our assistant does 85% of the work now.”

Building the solution: On the way to prototyping

The foundations for the hub were laid two years earlier, when La Silla successfully migrated its CMS with support from the same product and tech team. That project helped establish internal trust and effective communication channels with the newsroom, enabling smoother collaboration on AI integration.

Initial problem identification was participatory. Editors were invited to submit ideas, and the tech team also approached specific desks with targeted offers to collaborate. Each concept was validated with senior editors to ensure sufficient newsroom alignment and long-term utility.

The team worked in weekly sprints, with cross-functional collaboration at the centre. They iterated quickly through prototypes and conducted daily stand-ups to identify blockers and adjust priorities. Feedback loops were built into every stage.

Tools used

The project combines several LLMs and development platforms:

• LLMs tested: ChatGPT (including GPT-4, GPT-5 and Opus), Anthropic Claude (Sonnet and Opus), Google Gemini, and DeepSeek.

• Back-end and infrastructure: MongoDB, Vercel, Google Cloud Platform, SupaBase, WordPress API.

• Prototyping: V0 from Vercel was particularly helpful early on for rapid front-end deployment, although its quality declined after a later update.

The team conducted multiple prompt engineering cycles and attempted fine-tuning four times for one of the tools, though with limited success. They prioritised flexibility in tool integration, switching models or APIs depending on performance and cost constraints.

Team and skills

The project was led by a hybrid team with backgrounds in journalism, technology, and product development and included a product manager and editorial bridge, with a background in journalism, an external collaborator with experience in chatbots and civic tech, full-stack and a backend developer. The team also included a feedback coordinator and user liaison during the project implementation, after discovering how relevant it was. 

The challenges in the way

• Reassessing and restructuring the hub: Halfway through the development process, the team decided to rebuild the hub’s architecture to better support external users. While this change added technical complexity and delayed timelines, it was necessary to future-proof the product.

• UX design limitations: No one on the team was a UX specialist. Eventually, they organised a "Design Critic" session, studied its documentation, and revised their approach in a space facilitated by their consultant. This helped improve usability, but the process highlighted the need for better design resources in future iterations.

• Feedback volume and complexity: The team underestimated feedback volume and learned that collecting input without acting on it would frustrate users. "We decided we could only ask for feedback if we had the capacity to act on it," said De la Hoz. They created a dedicated feedback owner role and established a lightweight system using WhatsApp voice notes, which are transcribed into actionable documents, enabling quick adaptation while reducing user friction.

The opportunities: What’s ahead

The team sees strong potential to scale the hub to other newsrooms in Latin America and beyond. “We want to figure out how to make this technically and economically viable for others,” De la Hoz said.

Internally, the tools are already being used regularly. FranBot, the daily round-up tool, and the X thread assistant have all been integrated into newsroom routines. Their continued use by a sceptical and demanding editorial team serves as a strong indicator of impact.

Lessons for newsrooms 

• Feedback systems are as important as the tools: Without mechanisms to gather and act on user input, even the best tools will fail. 

• Adoption requires trust and experience: Past wins like the CMS migration helped build internal credibility. Journalists were more willing to collaborate on AI tools once they saw tangible benefits and trusted the team’s intentions.

• AI projects need cross-disciplinary integration: Success depends on merging technical development with product thinking, newsroom collaboration, and user-centred design. “A tech team alone cannot build the right tools. You need newsroom buy-in from day one.”