Minimising the costs of AI-powered content discovery
FinOps could play a key role in shaping how the television industry harnesses AI – and it is already helping Liberty Global decide whether and when to use large language models (LLMs) to improve content search and recommendations.
Chris van der Linden, director of entertainment platforms at the pan-European Pay TV operator, recently highlighted the importance of a strong FinOps function during The Media Leader’s ‘Achieving leadership in AI-driven content discovery’ webcast.
This goes hand in hand with clear total cost of ownership (TCO) figures when establishing use cases for AI across the platform operator’s user experience (UX).
Van der Linden (pictured above at Connected TV World Summit 2026) is responsible for Liberty Global’s product strategy and is also the ad interim director of the AI roadmap at the company.
Speaking on the webcast, he noted the need to manage AI-related costs, emphasising the importance of determining whether a large language model (LLM) should be used to resolve viewer voice search content queries or other forms of recommendation, including those found in a personalised content rail on an operator home screen.
He warned: “Personalising every content rail using LLMs that have been trained on the entire Internet could be very costly.
Avoid breaking the bank
“If you shoot every personalisation request to an LLM that is going to break the bank.”
Based on descriptions from the FinOps Foundation (a nonprofit trade organisation that is part of the Linux Foundation), IBM and Microsoft, FinOps is an operational framework and form of cultural best practice that uses collaboration across engineering, operations, finance and business teams to achieve financial management discipline and accountability when harnessing cloud technology.
It is also designed to maximise performance and business value from the cloud.
FinOps is a portmanteau of ‘finance’ and ‘DevOps’. According to the FinOps Foundation, this function may also be known as ‘cloud financial management’ and ‘cloud financial engineering’ among other terms.
It says the expression could also cover spending across technology categories such as SaaS, licensing, and data centres.
Liberty Global is clearly tuned to the dangers of allowing AI token costs to run out of control.
Picking the right models
Van der Linden declared: “You have to think about [large language] model choice.
“The very sophisticated models cost more, and that is where a lot of your money could go, but you don’t always need the latest version of a model, so you can use fewer tokens with alternatives.
“You need good FinOps because [large language] models change, and hyperscalers can increase or decrease pricing on certain models.
“A model can be suddenly decommissioned, forcing you to switch models.”
He outlined the hybrid approach his company is taking to improve discovery.
“Content personalisation was possible before LLMs, so there is no point using them whenever personalisation is needed.
“Depending on the use case, it may be perfectly fine to use standard algorithms in the back office.”
Liberty Global is harnessing AI for its SuperSearch conversational voice search capability, which launched at the Swiss Pay TV operator Sunrise in summer 2025 and has since been implemented at Vodafone Ziggo in the Netherlands.
Both operators are part of the Liberty Global group.
Liberty Global companies already provide voice search via their remote controls, but SuperSearch takes this to another level by using an LLM – whenever necessary – to better understand what users are looking for.
AI provides natural language understanding and is able to infer what people want based on what they say, even if their request is short of detail or inarticulate.
Search results are presented in the operator UI as a collection of ‘posters’ – which is more visually appealing than a list.
“SuperSearch uses Gen AI to understand the intent behind a customer query,” van der Linden confirms.
That intent is then translated into keywords that are entered into the operator recommendations system, which decides what content to present to the consumer, harnessing the (often rich) metadata associated with content.
Liberty Global has taken architectural decisions on when to use LLMs.
First, an LLM is not being asked to make the recommendations itself.
An alternative AI-powered content discovery approach – which has been pitched to the industry but is not used at Liberty Global – uses LLMs to make recommendations that are then checked by the operator.
Commands, not requests
The SuperSearch design also prevents an LLM from being asked to process voice requests that are actually simple commands – like ‘Netflix’.
In this case, the platform knows to open the Netflix app as fast as possible and let the viewer browse the content themselves.
However, if you tell the remote control to find the film with Tom Cruise where he has a brother he didn’t know about, the LLM will be drawn into the process (and figure out you want Rain Man).
During the webcast, Rafal Fagas, chief technology officer at the metadata services provider Media Press Group, also addressed the need to minimise AI costs in content discovery.
One of his company’s services is ‘metadata enrichment’, which expands on the programming information offered by content suppliers.
To help with this, Media Press Group uses agentic AI to automatically conduct deep searches of the Internet to find what reviewers and other commentators are saying about content, among other things.
Fagas explained the benefits of a hybrid AI approach that can use a locally hosted and smaller language model or a third-party large language model, depending on the scale of the specific task.
His company has invested in CPU processing in data centres, so can run its own ‘medium language models’ locally. It offloads the majority of its metadata enrichment processing there.
The local models are fit-for-purpose even if they require some additional prompting and come with more operational complexity.
Fagas explained, “This approach allows us to heavily limit the cost of using AI.”
More powerful models, locally
He added: “The technology gets better and better, so we will be able to run more powerful models locally using the same hardware resources.”
Fagas believes this hybrid approach is future-proofed.
“We are observing a trend for higher prices among the major large language model providers, and as AI companies come under pressure to show profits, we will probably see the cost of frontier models rising again. We can accommodate that.”
Ben Gidley, senior director, product management for OpenTV ENTera (the platform operator UX, discovery and personalisation solution) at Nagravision, also addressed the cost of improving content discovery.
He pointed out that whether AI-supported or not, all advanced UX techniques need an ROI justification.
A/B tests are part of the answer. For example, one content recommendation rail shown to a viewer could make use of standard collaborative filtering (for content recommendation) while another harnesses enriched metadata, too.
The impact can be weighed against any extra costs.
As we reported previously, this webcast investigated the future of content discovery, with the focus on platform operators. It highlighted a trajectory towards more personalised home pages for each viewer, who is served more of what they want with less navigation effort.
Predicting viewing uplift
The audience answered a poll: “What is a realistic uplift in engagement (time spent viewing) for operators who master AI-powered content discovery?’
30% of answers selected an 11-20% uplift and another 30% chose a 6-10% uplift figure.
22% of answers said 21-35% uplift, and 13% stated 36-100% uplift.
4% of answers chose ‘Under 5%’.
Renáta Fülöp-Árvai, head of TV and content services at the Hungarian Pay TV provider One Hungary, said her vote is for an 11-20% uplift in viewing time.
One Hungary is aiming for a 15% increase in engagement over time, in fact, but has only just started its journey in AI.
Van der Linden is far less bullish. He thinks customers are already finding their way to the content they want, one way or another.
The big gain, therefore, is not total viewing time (which he expects to rise only modestly on a platform-wide basis, albeit with wins for some content assets).
Instead, the win comes from reducing customer annoyance.
“We can reduce the irritation factor by making it easier to find what you want to watch next,” he declared.