Orkonerei FM: OrkoBot, a co-pilot for a six-person station
No-code custom ChatGPT agent, fed the station's mission and workflows — six users on two shared accounts.
Problem statement
Orkonerei FM is a community radio station run by six journalists, 24 hours a day — in the newsroom's own umbrella framing, the aim was "to minimize the workload of staff… as we are a small team, 6 journalists… the radio must go." Two pressures stood out: social media demanded a constant presence, but content creation sat with a single social media manager; and the Friday meeting to build the coming week's workplan — programmes, topics, who to interview, what to ask — "was taking a lot of time."
The solution
A custom agent created inside ChatGPT — no code, on a bought subscription — which the newsroom calls its co-pilot: OrkoBot. After in-person training in Nairobi, the team discussed as a group what they needed, chose ChatGPT, built the agent using what the training taught them, and fed it the station's own documents so it "can understand our mission and our vision and the tasks we need it to do" — then tested it and shared it with the whole team. The station's editor and social media manager led the build; he says plainly that he "didn't have the knowledge" to create such an agent before the training — the single thing he wishes he had known at the start. Six users run on two shared accounts, and the tool is now in regular use.
CONTRADICTION FLAG — carried from extraction: the sources are ambiguous on whether this is one prototype or two (interviewer and notes suggest two; the newsroom's own narrative and written Q&A describe one, "OrkoBot"). The extraction's resolution — one entry, two functional strands — is followed here.Results
On the social strand: roughly three to five posts a day, varying with news volume and platform, produced mostly by the social media manager. Over the fellowship period the station's following grew markedly — Facebook from 2,430 to 6,092 followers, TikTok from 2,000 to over 12,000, Instagram from around 300 to 785, and a YouTube channel launched in August 2025 passing 500 subscribers in nine months. The newsroom's own attribution is deliberately careful: "We believe AI-supported content workflows, combined with our newsroom efforts, have contributed" to the growth — and the same grant also bought a camera and funded the YouTube launch. On the planning strand, everyone uses it, since everyone runs programmes: scripts come out "easy, good and interactive", arranging news after fieldwork is faster, and it helps avoid re-covering stories already done. The friction was human, not technical: prompting was the main challenge — "you need to write every detail in the prompt" — and one colleague remains reluctant, using it only occasionally. Discovering that AI could make small mistakes, such as rearranging a script incorrectly, is what produced the station's mandatory review rule.
Insights
OrkoBot is embedded at two points in the workflow: story development (refining ideas, angles, interview targets, choosing the format — vox pop, package, narrative) and post-reporting script structuring before editorial review. The boundary is stated consistently: "AI supports our work, but editorial decisions, verification remain with journalists and editor." Adoption is uneven — journalists use it mostly for workplans, while the social function is essentially one person's.
Cost
Grant-funded, covering the AI and internet subscriptions, a camera for events and short video, and stipends for those building the prototype. It took roughly three months from training to the team using the tool. The dependency is stated plainly: "if it's not this project, I don't think… we can create the prototype" — and ongoing, "the use of AI is great, but it needs money… we are trying to have those budgets on subscriptions."
Replicability potential
This is about as low-barrier as newsroom AI gets: a no-code agent on a consumer subscription, grounded in the station's own mission and task documents, with six users sharing two accounts. The critical input is training — agent creation and prompt-writing were exactly the knowledge the builder lacked at the start. The honest caveats travel too: adoption stays uneven without individual buy-in, and the sustaining question for a peer station is less "can we build it" than whether subscription and internet budgets survive after the grant.