All 6 articles in this series are now available in our free eBook ‘Content Operations In The Multi-Screen Age’ – download it HERE.

All articles are also available individually:

AI in its various guises is intruding upon all aspects of content operations and dissolving boundaries between disciplines across the workflow pipeline that were previously largely separate. Distinctions between content preparation or generation and subsequent processing, distribution and playout have become more blurred, with increasing intersection.

We focused on transcoding in Part One of this content collection which inevitably brought in AI, but here we attempt to draw the bigger picture, examining how the various components and disciplines under the umbrella of content operations fit together.

AI is also blurring boundaries between programming content and the mechanisms designed to liberate it and make it more readily available to viewers, ideally with some personalization. Broadcasters are starting to experiment with this, such as Luxembourg’s RTL, which more than almost any other video service provider has made a major commitment to bring AI into the heart of its content workflow.

A Central Pillar Of Content Strategy

RTL’s CEO Stephan Schmitter recently asserted that AI was no longer a sideshow for broadcasters but was becoming a central pillar of content strategy. At least it is becoming so at RTL.

RTL had noted the experience of other video service providers such as Magenta TV, the pay TV arm of Deutsche Telekom, which ran into teething troubles recruiting AI for dubbing a crime series early in 2025. The synthetic AI-generated voices in German were lifeless, robotic and devoid of emotion, leading to the dubbed versions being pulled, leaving just the original language version with subtitles.

Quite correctly Magenta TV has been undeterred by this experience over AI in general, continuing to experiment with the generation of short films alongside more traditional formats for narration. But that experience at a fellow service provider helped confirm Schmitter’s view that AI should be employed at RTL to enhance and complement human creativity, rather than supplant it. 

This itself is a challenge, as AI extends multiple aspects of content workflow. To meet that challenge, RTL has set up an internal AI Academy to train staff on how to apply the relevant AI tools productively across multiple disciplines. It has also been forging partnerships with academia and tech firms around the world to support this effort with an attempt to gain some early mover advantage.

One such partnership was announced in June 2025 between RTL and US firm Perplexity AI, a company backed by Amazon’s Jeff Bezos which gained global notoriety for its disruptive and surprise $34.5 billion bid for Google’s Chrome browser. RTL was interested in the ability of the firm’s eponymous conversational search engine to interrogate the content of programs made by its online news outlets in Germany, such as NTV and Stern.

RTL Deutschland is now investigating how these tools designed to bridge the gap between traditional search and AI-based program generation can unleash content. The aim is to make content more readily available to different demographics through searches that probe deep into the content beyond just available metadata, or even audio recognition. The broadcaster said that from now on it would be intimately involved and engaged with Perplexity AI’s ongoing R&D to help direct developments towards content liberation, dissolving the lines between generation and search.

There is also an ambition, as Isabella Thissen, Managing Director Digital at RTL News put it, to distinguish NTV and Stern from the torrent of AI-generated content, as trustworthy brands. In other words, RTL is recruiting AI to combat associated challenges of the technology, as well as for competitive reasons.

Considered Applications

Other broadcasters are similarly engaged with standing out from the crowds of alternative content generated by myriad social media and other platforms, often with less considered application of AI. Among them is the BBC in the UK, still maintaining its reputation as a technical innovator despite increasingly stringent financial controls as the clock continues to tick down on license fee income.

The BBC is cautious over application of AI for primary content generation on account of its editorial and ethical guidelines, but is engaged in a range of pilots to produce summaries and process external sources of material, especially news items, for stylistic and formatting consistency. The broadcaster is just starting to test two key pilots publicly, according to its Executive sponsor of Generative AI Rhodri Talfan Davies.

These are ‘At a Glance’ summaries and BBC Style Assist, with the former applying Generative AI (Gen AI) to help journalists create summaries of longer form news articles. Gen AI has the ability both to generate and translate content across multiple media, including text, audio, video and graphics. That same capability can be applied to reduction or summarization, taking account of video, audio and graphics, text or captions embedded within the material.

At this stage, the BBC is concentrating on the creation of short and accurate bullet-point summaries that can be scanned and genuinely represent the longer form content. The summaries should grab attention without being sensationalist or distorting the message embodied in the original content. The BBC has already tested such summaries with users and found they increase viewing levels, especially among younger audiences.

Again, in line with its standards, the BBC insists that journalists will continue to review every summary and edit as necessary, although with the expectation there will be diminishing need for manual tweaking. “We will also make clear to the audience where AI has been used as part of our commitment to transparency,” the BBC stressed.

Style Assist

The BBC applies a similar process for its Style Assist as it does for “At a Glance” but focused in this case on reformatting to house style without reduction, rather than summarizing. The BBC receives a lot of its national news content from a public news partnership it funds called the Local Democracy Reporting Service (LDRS). This provides content of local relevance that is often of wider interest, which has in the past required manual conversion to house style and standards. This has limited the amount of content that can be made available, which has become a greater constraint in the era of online streaming where there is no effective limit on the number of items that can be disseminated.

The content can be trusted and so does not need inspection on that count. The workflow then begins with a trusted story taken perhaps from the LDRS, which is submitted to the BBC’s content system, where Style Assist generates a draft revised for BBC style and tone. The story is then reviewed by a senior journalist before being published on the BBC News website and app.

Finding The Content

Broadcasters face various challenges extending the scope of such capabilities across their content archives, not least computational scale. In practice, making content more searchable requires the generation of more comprehensive metadata that allows finer grained categorization, offering scope for greater targeting and personalization.

The big tech companies have all been working on this, an example being Microsoft’s Azure AI Content Understanding, which now claims ability to transform unstructured video into more structured forms that are searchable. Aimed initially more at enterprises it is attracting growing interest from broadcasters. 

This employs multimodal data ingestion, where text, images, audio, and video, are combined in a unified system for collective analysis and processing, aiming to derive categories more nuanced and finer grained than would be possible using any single mode by itself.

An Existential Question

An important point is that broadcasters are rarely in full control of the whole video pipeline, being also reliant on capabilities within networks, Content Delivery Networks (CDNs), and target devices for the ultimate viewing experience. AI plays into this too, as well as extending the scope of video transcoding, with potential for quality enhancement as well as just minimizing the impact of compression, along with limited bit rates or device capabilities.

This point arose with the advent of onboard processing within TVs as they became capable of upscaling content for display at higher resolutions. It became common for TVs to scale up from standard definition to HD, and then more recently from HD at 1080p to 4K.

TVs were capable of performing the basic upscaling so that the content filled the full screen, but initially did so without any processing to take advantage of the higher resolution. The result was that the upscaled content, such as a movie, would tend to appear flat. In essence, pixels were just being replicated to blow up the picture without being recalibrated to make it look more vibrant, exploiting the higher resolution.

Now AI techniques are entering to make upscaling more natural, so that the video looks closer to what it would if it had been captured at 4K resolution in the first place. For best visual effect this involves some content generation that is inevitably somewhat speculative, leading to the almost existential question of whether what is being presented is real or not.

In practice, broadcasters do not need to be too concerned with such niceties but do need to prepare for the growing interaction between simulation, enhancement and transcoding under the auspices of Gen AI. There are also differences between live and on demand or archive content, with obvious compromises that have to be made for the former.

Different Approaches

For archive content, an array of AI-assisted techniques is now available for enhancing old material, from coloration to ray tracing for adding more realistic shadows and illumination within frames by analyzing apparent sources of light within them. For live content, a balance must be struck between quality optimization, latency and available compute power.

There is also increasing reliance on the playing device for enhancing quality of live playback, as in AI-enhanced upscaling. This can also be applied to legacy content, which can be sharpened up considerably through real time upscaling if it was originally filmed in standard definition.

Leading streaming services such as Amazon Prime Video already employ what they call hardware acceleration to improve playback, reducing buffering and artefacts. This typically uses the device’s onboard GPU (Graphical Processing Unit), with maximum benefit in lower end devices such as budget smartphones.

AI in general can be a double-edged sword when it comes to the balance between quality and hardware resources. 

On the one hand, AI at the coal face is compute intensive because of the processing involved in inferencing on the fly for transcoding or other processes. But it can improve efficiency by focusing the compute power on important aspects of a problem.

Dedicated Equipment

The latter arises in transcoding because some parts of a frame, such as a face, matter visually more than others, like an expanse of blue sky where there is scope for greater compression. Machine learning is improving Region of Interest (RoI) processing where an encoder adjusts block level quantization parameters around those important parts of a frame, resulting in improved perceptual quality for a given bandwidth or compression level.

Quantization in the video context is the process of mapping the pixels in a frame to a smaller set through processes such as truncation and rounding of numbers.

This is all part of Content Aware Encoding (CAE), exploiting analysis between and within frames to stream video at a lower bit rate for a given quality. This is where the lines can be blurred between transcoding and enhancement, giving scope for improving either compression level or quality further by allowing scope for regeneration. This process is sometimes known as lossy to lossless compression, which relies on some generation or simulation, or else would be defying mathematical logic.

These developments are affecting the design of video processing equipment at the hardware as well as software level, including transcoders. The high computational load favors the use of dedicated ASICs (Application Specific Integrated Circuits) to execute specific highly intensive AI algorithms alongside the overall video processing software that still runs either in GPUs, or even more general-purpose CPUs (Central Processing Units). In some cases, GPUs themselves are employed for hardware acceleration, but ultimately the highest levels of performance for specific tasks involving a lot of parallel processing are best done in ASICs, or some other form of dedicated video processor.

Some of the largest content distributors, such as Meta, have developed their own in-house ASICs to alleviate the huge processing overhead of both their VoD and live streaming workloads. While this is beyond the scope of most broadcasters, commercially available encoders are increasingly incorporating ASICs within their designs.

There is a lot for broadcasters to consider at a time of rapid evolution in the scope and capability of AI across content workloads. Inevitably some are hesitant over committing too quickly, with few willing to take the plunge to the extent RTL has.

But broadcasters should at least evaluate the options available extensively in pilots, so that they acquire the knowledge and capabilities required when they do bring AI more extensively into their workflows, as the BBC is doing. 

All 6 articles in this series are now available in our free eBook ‘Content Operations In The Multi-Screen Age’ – download it HERE.