At many organizations, weak AI governance may cost you an unhappy customer or a fine from a regulator. At the BBC, it costs public trust, "the thing that the entire institution is really effectively built on," Nathalie Berdat tells host Satyen Sangani on AI Radicals.
Berdat is Director of Product Data at the BBC, where she arrived from commercial organizations including Aviva and Kraken to become the corporation's first Head of Data & AI Governance in early 2023.¹ The BBC is funded primarily by a flat annual licence fee, which brought in £3.8 billion of its £5.9 billion total income in 2024/25.² Any UK household needs one to watch live TV on any channel or streaming service, or to use BBC iPlayer at all.³ That means its choices get scrutinized in a way a commercial broadcaster's never are.
"There is less optionality for a license fee payer," Berdat says. That raises the bar on AI governance well past process compliance. "The governance isn't a compliance checkbox," she notes. "It's closer to editorial standards. It has to be defensible to a journalist: Why did the algorithm do that?" A regulator asks whether you followed the process. A journalist asks why the model produced that specific output, and keeps asking.
In this blog, we'll recap some of the key takeaways from the interview with Berdat.
The tell that trust in data had collapsed
Berdat spotted the symptom of data distrust within her first few weeks. The BBC's all-star metric is weekly active accounts (an average weekly count of signed-in accounts across BBC Online, not of people⁴), and the number moved depending on who calculated it. What struck her was how the conversation quickly shifted: "People stopped asking whether the number is right and started asking, 'So who produced that number?'"
Teams knew which sources to trust for which purposes and had absorbed the quirks. But the knowledge lived in individuals' minds rather than in systems. "Tribal knowledge is not something that we can scale, or that we can actually rely on when we make big decisions," Berdat says. And the BBC faces genuinely big decisions in the age of AI.
When two teams arrive at a meeting with two different charts for the same metric and nobody in the room objects, trust in data has already gone. The absence of any argument reflects the incredible complexity at play. Underneath a single metric sits thousands of tables and multiple versions of the same dataset, each with slightly different filters, exclusions, and interpretations, and no clear owner for any of them. This was the key challenge Berdat faced.
Why the most-used dashboards were the wrong place to start
The first attempt at prioritization failed, and Berdat is candid about why. Her team pulled usage statistics and targeted the most-viewed dashboards, on the reasonable theory that popularity signals accuracy and value. "But it was the wrong way to approach this," she says. Heavily used dashboards tend to serve daily monitoring and low-level analytics, but they don't represent where consequential decisions get made.
So the team went top-down instead, to the audience scorecard the board actually reviews. Those metrics do double duty as performance reporting and as inputs to strategic decisions, which made them both the highest-value target and the most visible proof point. As Sangani observed in the conversation, this top-down approach aligns the organization around a shared set of objectives while also solving the selection problem of what deserves attention first.
Weekly active accounts were the obvious place to start. With a flat license fee, the BBC cannot pull a pricing lever to simply grow revenue, so engagement and logged-in usage carry disproportionate weight. 23.6 million weekly active accounts now depend on getting that definition right, up from 22.8 million the year before.⁴
How to get teams to adopt data products
Berdat's team traced how each priority metric was calculated, found the competing sources, and rebuilt them into a single data product designed to be the certified source of truth feeding every weekly active account calculation. The build itself took a few sprints. The hard part came next.
"The difficulty is not so much building it, because we've had all the dimensions that could allow people to slice and dice as they wanted to. It is for people to let go of the old tables they're used to," she says. Those old tables were pulled together with a shoestring and broke regularly, and users still defended them.
Three arguments eventually won people over. The first was accountability. The new data product shipped with clear SLAs, embedded quality checks, and a named owner, so anyone who questioned a number knew exactly who to ask. The second was time. A dedicated team now maintained the product, which meant users would spend far less of their week repairing brittle pipelines. The third was proof rather than promise. Berdat's team rebuilt one pipeline that had been falling over every Monday morning, removing an outage people felt directly.
Credibility also came from knowing when to stop. The team made a deliberate call to leave residual discrepancies of roughly 0.1% alone, since they knew where those gaps came from and further reconciliation had hit diminishing returns.
Adoption took nine to twelve months before Berdat could point to solid evidence rather than commitments to migrate later. She frames the exercise as cultural work more than technical, and notes that in a large organization assembled from divisions that share data, agreement is the scarce resource. (Teams working through a similar transition may find Alation's write-up of the BBC's search metrics data product useful here, which covers the same operating model in more detail.)
Building reusable data products on a shared governance framework
Berdat says the BBC now runs a few dozen data products and is completing a migration to a new enterprise data platform that lets divisions build and share their own. She declines the data mesh label, since the BBC is not standing up end-to-end teams in every domain and does not consider that practical from a resourcing perspective. Instead, her team is giving domains their own spaces to build in, with data governance providing the necessary framework for enterprise-wide consistency.
She shared one example of the kind of challenges her team is tackling: How many people work at the BBC? The answer is genuinely hard, because the organization runs on contingent staff. A cameraman works two days. A contractor comes in for two months. Linking those people to cost centers (via data products) makes it possible to measure the true cost of a program, and since commissioning one is a three-year process, return on investment matters enormously. "It's like Lego, right? If you have one data product that you can then build another data product on top of, it just unlocks so many possibilities," Berdat says.
Berdat defines governance more broadly than standards for ingestion formats and minimum metadata. At the BBC, data governance has a human dimension, as it covers who owns the data, who fixes a product when it breaks, where the data came from, and who controls access. That's why Berdat's firm on sequencing: "I never start with tools when it comes to data governance." She recommends that leaders first decide what you want to do with your data and what safeguards it needs, then map tooling requirements against that, because human data governance does not scale.
In practice, the BBC embedded governance experts directly into platform sprint teams, then ran a large procurement. The award notice is public: a £1.2 million, five-year contract concluded in October 2024, selected from 14 tenders, bringing in Alation for the BBC Product Group's data governance team, alongside a separate data quality and observability tool that Berdat says was procured in parallel.⁵ People, process, technology, in that order.
When AI touches editorial judgment
The BBC has used AI behind the scenes for years. Berdat points to machine-learning models that watch live natural history camera feeds and flag the moment an animal appears, and the corporation has used AI to transcribe live football action since early 2023.⁶ The governance complexity arrives when AI gets close to expression.
Recommendations present the clearest case. "Because we are making a judgment call about what content someone can see, and at the BBC that's right up against editorial values like impartiality, range of perspective," Berdat says. A retailer recommending a product answers to an engagement metric. A public broadcaster recommending content needs an editorial justification, so editorial signs off on automated experiences before they reach production.
Recent BBC audience research points the same direction. Ipsos research for the BBC found that generative AI is welcome when it helps behind the scenes and resisted when it replaces human creativity, voice, or editorial judgement, with tolerance in news narrow and functional.⁷ 58% said AI in media makes them nervous and 78% prefer online news written by humans; 43% said fact-checked outputs would make them more likely to trust AI use in news.⁷ Independent research suggests the BBC is right not to assume it gets the benefit of the doubt: across six countries, only 33% think journalists always or often check AI outputs before publishing, and in the UK just 25% do.⁸
The BBC publishes transparency guidelines about where AI is used, and now ships AI labels, a hexagon icon paired with the phrase "How we used AI", at the top of content.⁹ The caution is evidence-led: research the BBC published with the European Broadcasting Union in October 2025 found that AI assistants misrepresented news content 45% of the time, through faulty sourcing, fabricated details, or outdated information.⁹ And no AI-assisted story is published until a journalist reviews and approves it.⁹
Style Assist shows what that looks like in production. The BBC receives hundreds of stories a day from the Local Democracy Reporting Service, a network of non-BBC journalists covering hyperlocal stories the corporation cannot reach.⁶ Rewriting one into BBC house style by hand can take around 30 minutes.⁹ Now, a model does the repackaging and a human journalist reviews every output before publication. The BBC ran it as a six-week pilot with news teams in Wales and the east of England, measuring time saved and the error rate left after editor review.
That review is not a formality: in trials, the tool sometimes muddled attributions and editors slowed down to fact-check more.⁶ Sangani called applying standards with AI, rather than generating content with it, a strong and emerging use case across Alation's customers, and a form of content governance.
What data governance needs to be in place before agentic AI
The BBC has no live agentic project, and Berdat is precise about why. "Before we would hand an agent autonomy, even to a small degree, we absolutely need clear ownership of the data it would touch. We need that documented lineage so we can trace a bad decision back to its source. We need machine-readable governance rules, not just a policy."
There is an organizational precondition too. Agents deliver value horizontally, across end-to-end processes, and lose most of their benefit when confined to one team. "Agentic doesn't stop to ask a department. It has to be able to run across," she says. Agents are also expensive, so deployment has to be aimed where the return justifies the cost. Sangani noted that the majority of enterprise data leaders he speaks with are taking the same cautious posture, unwilling to pay a premium for something unreliable.
The medium-term work is well underway. Berdat says the BBC has built a conversational layer over its gold datasets as a step toward low-code and no-code self-serve analytics, and that her team is templatizing model deployment, since shipping a model remains hard work. She also makes a point that deserves wider circulation: accuracy is expensive, and the right degree of accuracy varies by use case. On data lineage and machine-readable policy, the groundwork for AI agents is the same groundwork that made the metrics trustworthy.
Berdat's advice to a data leader starting Monday is deliberately small in scope. Do not try to fix everything broken in your data landscape. Pick something strategically important enough to matter to whoever sponsors the program, because the grand rebuild does not survive a budget conversation. "Rebuild the foundation, start from scratch, and you go, 'Well, you won't see very much for the next two years, but when it's done, it's going to be amazing.' That doesn't sell," she says. Her concrete instruction: pick one metric that has more than one version and is causing visible pain for your leadership, and use it to demonstrate what good looks like.
The through-line of the episode is that the BBC's AI position is a consequence of its data position. Ownership, lineage, and machine-readable rules are what make agent autonomy conceivable, and they are the same things that made a single metric trustworthy. Teams that pull ahead will be the ones that did that unglamorous work early.
Curious how a governed data products marketplace and trusted metadata foundation can support your AI strategy? Book a demo with us today.
Sources & Notes
Every external claim on this page is independently verifiable. The public sources are listed here.
The BBC appointed Nathalie Berdat as its first Head of Data & AI Governance in early 2023. — The Stack, 13 December 2024 ↗ https://www.thestack.technology/bbc-hands-alation-data-catalogue-contract-amid-data-transformation-pressure/
Licence fee income of £3,843 million against total income of £5,900 million in 2024/25. — BBC Group Annual Report and Accounts 2024/25, presented to Parliament 15 July 2025 ↗ https://assets.publishing.service.gov.uk/media/6876519139d0452326e28ec7/BBC_Annual_Report___Accounts_24_25_final_accessible.pdf
A TV Licence covers an address for all TV channels, live TV on pay and streaming services, and all BBC iPlayer use, on any device. — TV Licensing, accessed 27 August 2026 ↗ https://www.tvlicensing.co.uk/check-if-you-need-one/topics/watching-live-online-and-on-mobileAverage weekly active accounts using BBC Online: 23.6 million in 2024/25, up from 22.8 million in 2023/24. The BBC reports accounts, not individuals, and reports a weekly average. — BBC Group Annual Report and Accounts 2024/25 ↗ https://assets.publishing.service.gov.uk/media/6876519139d0452326e28ec7/BBC_Annual_Report___Accounts_24_25_final_accessible.pdf
BBC "Data Catalogue" contract awarded to Alation Inc; £1,200,000 excluding VAT; five-year initial term plus an optional 24-month extension; contract concluded 8 October 2024; 14 tenders received; requirement sits with BBC Product Group's Data Governance team. — UK Government, Find a Tender, contract award notice 2024/S 000-040206, published 13 December 2024 ↗ https://www.find-tender.service.gov.uk/Notice/040206-2024
BBC Style Assist and the Local Democracy Reporting Service; six-week pilot with news teams in Wales and the east of England; attribution errors in trials; AI transcription of live football action since early 2023; Rhodri Talfan Davies, the BBC's executive sponsor for generative AI, quoted throughout. — Ron Schmelzer, Forbes, 29 June 2025 ↗ https://www.forbes.com/sites/ronschmelzer/2025/06/29/bbc-rolls-out-ai-summaries-and-style-tool-in-newsroom-test/
UK audiences accept generative AI behind the scenes and resist it where it replaces human creativity, voice, or editorial judgement; tolerance in news is narrow and functional; 58% say AI in media makes them nervous, 78% prefer human-written online news, and 43% would be more likely to trust AI use in news if outputs were fact-checked. Online survey of 2,000 UK adults aged 16–75, fieldwork April–May 2025. — Ipsos research commissioned by the BBC, published 10 February 2026 ↗ https://www.ipsos.com/en-uk/generative-ai-and-audiences-revisiting-uk-public-attitudes-ai-media
Across six countries, only 33% think journalists always or often check AI outputs before publishing them; the figure is 25% in the UK. Survey fielded by YouGov, 5 June–15 July 2025, approximately 2,000 respondents per country. — Felix M. Simon, Rasmus Kleis Nielsen and Richard Fletcher, Generative AI and News Report 2025, Reuters Institute for the Study of Journalism, University of Oxford, 7 October 2025 ↗ https://reutersinstitute.politics.ox.ac.uk/generative-ai-and-news-report-2025-how-people-think-about-ais-role-journalism-and-society
BBC AI labels ("How we used AI"); BBC/European Broadcasting Union research of October 2025 finding AI assistants misrepresented news content 45% of the time; no AI-assisted story published until a journalist reviews and approves it; roughly 30 minutes to rewrite a story into BBC house style by hand. — Erik P. Bucy and Milad Jalalian Ebrahimi, Nieman Journalism Lab, Harvard University, 29 July 2026 ↗ https://www.niemanlab.org/2026/07/how-three-newsrooms-are-charting-different-paths-for-ai-use/
- AI
- Customer Stories
- Data Governance
- Data Products
- Data Quality
- Digital Transformation
Keep reading
More from the data desk
