When social media traffic becomes unpredictable and search algorithms keep changing, email newsletters have emerged as a vital lifeline for news organisations. But manually coordinating content swaps between publications is time-consuming work that many understaffed newsrooms simply can't manage. openDemocracy's solution? Build an AI-powered platform that automates the entire process.

The problem: Manual content sharing takes too much time

openDemocracy, an independent global media platform covering democracy and human rights, had been running a successful but labour-intensive programme: manually arranging content swaps with other organisations to cross-promote their newsletters.

"We would reach out to people and there'd be a back and forth. It was very laborious," explains the project lead, Matthew Linares, who serves as both journalist and coder at openDemocracy. "In the context of a challenging distribution environment where social media and search traffic is precarious, newsletters have become one of the key elements for journalistic organisations to maintain that link with readers."

The manual process worked – organisations would share each other's content with sign-up links, helping everyone grow their subscriber base and increase donations. But it simply didn't scale.

Building the solution: CopySwap's AI-powered matching system

CopySwap was designed to automate this "virtuous circle" of audience sharing. The original vision was straightforward: organisations would upload content they're happy to share, and AI would suggest the best matches for each publication's newsletter.

"The idea was that AI would suggest appropriate content to match newsletters from the pool of submitted content," Linares explains. "Users would submit their content and receive AI-enhanced suggestions of what would suit their newsletter."

Building the platform: Tools and team

The development team kept things lean and focused, building on CiviCRM as the foundation – an open-source CRM that was already managing their donations and email campaigns – and adding a JavaScript app for additional functionality. Their technical stack initially included Quadrant as a vector database for content matching and OpenRouter for accessing various LLMs, with a deliberate focus on using "relatively lightweight LLMs" to minimise environmental impact. The team composition was similarly streamlined, consisting of a project lead who was a journalist/editor with coding skills, an audience manager, a multimedia producer, and a freelance developer.

"I'm perhaps unusual in that I directly code and understand the technical architectures whilst being a journalist and editor," notes Linares. "I bridge the two worlds, which simplifies things."

The pivot: When AI became the content creator

Seven months into development with only five active users, the team discovered organisations weren't creating enough content for the AI system to function effectively. "Users were struggling to create multiple messages on top of their other work," Linares admits.

So they flipped their approach. Instead of matching existing content, AI would generate the messages. "We're deploying a function that takes their social media feed and uses an LLM to create suitable messages," explains Linares. "It's like having a new marketing team member that transforms existing output for different platforms."

Ethical considerations: Keeping AI sustainable

Unlike many AI projects that prioritise functionality over sustainability, openDemocracy made environmental concerns central to their development process.

"We're very aware that AI has a large environmental footprint," Linares emphasises. "We think it's critical to minimise LLM usage, pick the right models, use the smallest footprint models, and use them only where needed."

This meant researching carbon footprints and ensuring human oversight remained paramount. "If something doesn't need to be AI, then don't introduce it," he advises.

Human-centred design: AI as assistant, not replacement

CopySwap treats AI as a helpful assistant rather than an autonomous system. When generating content suggestions, the platform presents both raw social media posts and LLM-generated alternatives, clearly labelled.

"The human gets the final sign-off. It's really just a helper," Linares explains. "We make it easy for users to accept it, take inspiration from it, or reject it entirely."

Future developments might include allowing users to add prompts or keywords, putting them "even more in the loop" to guide the AI toward their specific needs.

The opportunities: Scaling beyond journalism

Despite challenges, the team sees significant potential. The platform could expand beyond journalism to include campaigning and research organisations, potentially supporting fee-based models for sustainability.

The international dimension also offers promise. The team has begun attracting users from different regions including South Africa, and their global focus aligns naturally with their multilingual publishing approach and international content strategy.

Lessons for newsrooms

• Question the need for AI: Resist the "hype cycle" and clearly ask if AI is truly needed for your problem. Be sure there aren't cheaper or better-tested alternatives; sometimes, "boring technology" is the most desirable solution.

• Prioritise the problem, not the tech: CopySwap identified a real need – helping small news organisations grow audiences and revenue through newsletter partnerships – then thoughtfully applied AI where it genuinely helped, pivoting when user behaviour revealed a better application.

• Be ready to pivot: Successful implementation requires understanding user behavior and being willing to pivot when the original vision or application of the technology doesn't match reality.

• Focus on user needs and ethics: Success isn't about using the most advanced technology; it's about deeply understanding user needs, maintaining ethical standards, and being flexible enough to adjust your approach.