From AI pilots to newsroom practiceThe operating question1×0:53The adoption gap2:45La Nacion builds a controlled toolkit4:57Clarín measures the work7:19A small-newsroom operating model9:07What to act on now0:0010:380:00HostThe AI problem for most newsrooms is no longer access. People already have tools, accounts, and experiments running in different corners of the organisation.0:11AnalystThe harder problem is turning those experiments into a workflow that is safe, useful, and measured against an editorial or business outcome.0:22HostThat is the question for this week's briefing: what does it take to move from scattered AI pilots to newsroom practice, especially when the team does not have a large technology department?0:36AnalystThe short answer is structure. Recent reporting from INMA, plus the new Future Newsrooms Study from FT Strategies and WAN-IFRA, points to the same pattern: the limiting factor is usually people and process, not model availability.0:53HostStart with the survey. The Future Newsrooms Study collected responses from 448 people in newsrooms across 86 countries, alongside interviews with sixteen editorial and executive leaders.1:06AnalystIn Press Gazette's reporting on the study, 52 percent named cultural resistance or scepticism as the biggest barrier to wider AI adoption. Sixty-one percent cited a lack of internal technical skills or AI expertise, and 45 percent pointed to unclear use cases or strategic direction.1:28HostThose numbers describe a management problem. A newsroom can be curious about AI and still have no shared answer to three basic questions: which work should change, who owns the change, and how will we know it helped?1:44AnalystThe study also found that 61 percent of newsrooms offered no formal training for developing new skills. More than half had no AI expert in the newsroom, and only 14 percent of leaders were very or extremely confident that their current technology stack was fit for the future.2:04HostThat last figure matters for smaller publishers. It is easy to read AI adoption as a software purchase. The evidence suggests it is closer to an operating change, with training, ownership, and feedback built into the work.2:19AnalystThe study's official summary adds another warning. Sixty-four percent of newsrooms still design stories for a primary channel before adapting them elsewhere, while only 21 percent begin with a defined user need or audience group.2:34HostSo the first verdict is an immediate opportunity: define the job before choosing the model. A better prompt cannot rescue a workflow that has no clear owner or audience.2:45HostThe most useful case study this week comes from La Nacion in Argentina. INMA reports that the company has moved beyond isolated experiments and built an AI ecosystem used across six business units, including editorial, product and technology, data and analytics, and training.3:07AnalystThe toolkit is deliberately practical. It includes transcription, translation, style-guide correction, summaries, semantic archive search, automatic tagging, and a chatbot that answers questions from the newspaper's own archive.3:25HostThe important design choice is the boundary around the tools. La Nacion centralised its model connections behind one backend and hosted its automations on company servers. The point was to reduce repeated integration work and keep company data inside a controlled environment.3:45AnalystThe newsroom also kept a human review step. Journalists can edit or discard what the system produces, and nothing is published without editorial control. The tagging tool is limited in a useful way: it selects from existing tags rather than inventing new metadata.4:04HostThat is a good example of a small publisher's first AI platform. It does not need to begin with an autonomous reporter. It can begin with a few repeatable services around the archive and the CMS, where the risk is manageable and the benefit is visible.4:22AnalystThe reported results are modest but concrete. La Nacion says its proprietary correction editor saves about 700 US dollars a month in licensing costs. Its tagging system expanded the number of relevant tags per article from two to five, and nearly twenty training sessions reached 242 active users.4:44HostThose are publisher-reported results, so they should be treated as operating evidence, not an independent performance evaluation. But they are still more useful than a generic claim that AI made the newsroom more efficient.4:57HostGrupo Clarin offers the second case study, and it pushes the conversation from tools to measurement. INMA reports that the company produces about 600 stories a day, reaches 400 million monthly unique users, and serves 750,000 subscribers.5:17AnalystAt that scale, pageviews alone are a poor definition of value. Clarin developed an internal score from one to ten that combines pageviews, reader engagement time, subscription conversions, and new-audience acquisition.5:33HostThe score is not just a dashboard. It can influence homepage placement and the dynamic paywall, so the measurement feeds back into editorial and product decisions.5:43AnalystThat is the part local publishers can borrow without copying the scale. Pick two or three signals that reflect your actual strategy. For a regional publisher, that might be engaged minutes, newsletter registrations, return visits, or a subscription start. Then use the score to decide what gets more attention.6:03HostClarin also embeds AI directly inside its CMS for headline and deck suggestions, editing assistance, search and distribution support, transcription, translation, and clipping. That reduces the incentive for staff to send sensitive work through separate consumer tools.6:22AnalystIts governance model has four parts: clear rules for acceptable use, a multidisciplinary AI committee, organisation-wide training, and controls for data quality, privacy, compliance, and security. AI ambassadors in different business units help identify use cases and surface problems.6:45HostThe lesson is not that every publisher needs a committee with a grand title. The lesson is that someone has to make decisions across editorial, product, technology, and commercial teams. Without that connection, pilots stay local and the organisation learns very little from them.7:04AnalystClarin's own phrase is useful here: "Journalism leads and AI executes." That is a sharper test than asking whether a tool is impressive. If the editorial purpose is unclear, the tool is probably early.7:19HostLet us turn the evidence into a workable sequence for a local or regional newsroom.7:26AnalystFirst, choose one workflow where the current pain is repeated, visible, and low risk. Archive search, transcription, tagging, translation, or first-pass headline options are easier starting points than automated publishing or unsupervised reporting.7:45HostSecond, write the decision rule before the prompt. State what the tool may do, what it must never do, and which person signs off. A one-page rule is more useful than a long policy nobody reads.7:59AnalystThird, give the pilot a newsroom owner. That person does not need to be an engineer. They need enough time to collect feedback, document failures, and keep the tool tied to an actual desk or audience need.8:14HostFourth, centralise the parts that should be consistent. Shared access, model settings, data handling, prompt templates, and logging should not depend on every reporter making a separate choice.8:27AnalystFifth, measure one operational result and one editorial or audience result. Time saved can be useful, but it is not enough. Pair it with fewer correction errors, faster publishing of verified updates, more archive reuse, stronger newsletter conversion, or another outcome the newsroom already values.8:51HostFinally, train through the work. La Nacion's twenty sessions and 242 active users are a reminder that adoption is a practice, not an announcement. Show people a real task, let them test it, and keep the review loop visible.9:07HostHere is the editorial verdict. The immediate opportunity is to build a controlled internal assistant around one repeatable workflow, with a named owner and a baseline measure.9:20AnalystThe trend to watch is the shift from separate AI tools to shared services inside the CMS, archive, analytics, and product stack. That is where governance becomes part of the daily work instead of a document on the intranet.9:37HostAnd the industry noise is the next chatbot purchase treated as a strategy. A newsroom can add another model and still have no training, no data boundary, no success measure, and no person accountable for the result.9:53AnalystFor the next thirty days, a small publisher could select one workflow, interview the people who do it, set a before-and-after measure, create a human review checklist, and run the pilot with a small group of users.10:07HostAt the end of that month, keep it only if the work is measurably better or newly possible. Otherwise, stop it and record what failed. That discipline is part of AI governance too.10:21AnalystThe practical shift is simple: stop asking which AI tool the newsroom should try next. Ask which editorial job deserves a safer, better-supported system around it.10:33HostThat is this week's briefing. The sources and further reading are in the episode notes. Thanks for listening.