I promised in the last article that I would show the working rather than make promises. So here it is.
Everyone in media services is talking about AI, and I get it; the capabilities are exciting, and some of them are genuinely impressive. But most of the conversation I hear at conferences and read online stays at the level of “AI will transform content servicing.” That is not insight, it is a press release. The useful questions are more specific: which tasks, for which content, with what safeguards, and who is still accountable when something goes wrong?
We have spent the last eighteen months at EIKON trying to answer those questions. Not by theorising about it, but by building things, running them against real content, and being honest with ourselves about what worked and what did not.
What AI can actually do right now
Let me be straightforward about the state of play, because I think the industry would benefit from a bit less hype and a bit more honesty.
Speech-to-text is impressive; for clean audio in major languages, it is remarkably accurate. But we do not work with clean audio very often. Overlapping dialogue, accents, background noise, mixed languages, domain-specific terminology; that is what real content sounds like. Accuracy drops off quickly once you move away from ideal conditions.
Machine translation gets a lot of attention, and for certain applications it has come a long way, but we do not use AI for translation. Our clients need broadcast-quality localisation where tone, nuance, cultural context, and timing all matter. That is human work, and I do not see that changing any time soon. The gap between “technically correct” and “ready to broadcast” is bigger than most people realise.
Automated QC is where things get more interesting for us. AI can detect technical issues at speed: sync errors, level violations, format problems, missing elements, and it is fast and consistent. What it cannot do is tell you a mix sounds wrong, or that a grade looks off, or that a subtitle reads awkwardly in context. That kind of judgment still comes from experienced operators who have been doing this for years.
A loudness meter can tell you a mix is within spec. It cannot tell you the mix sounds wrong. That distinction matters more than most people think.
Where we use it, and where we do not
EIKON works on premium content for major studios and streaming platforms; that is what we know. The quality bar is as high as it gets, the creative constraints are tight, and the delivery specs are demanding. Our clients chose us because we meet that standard, and we are not going to compromise it by automating things that should not be automated.
So the question for us has never been “how do we automate the workflow?” It has been “which specific tasks within a human-led workflow get better with AI assistance?” And the honest answer is: quite a few, as long as you pick the right ones.
AI is brilliant at the heavy-lifting analytical tasks: technical QC, content analysis, metadata extraction, scene detection, transcription where conditions allow it. These are tasks where speed and consistency at volume make a real difference, and where you can objectively verify whether the output is correct. We get excited about this, because the tools are improving fast and the results are tangible.
What AI does not do in our workflows is produce output. We are not using generative AI to create deliverables. Our AI tools analyse, detect, extract, and validate, but the creative and editorial work, the actual output that reaches our clients, is produced by people. Translation is done by translators, editorial QC by experienced operators, and compliance review requires human judgment. These are the things our clients are paying for, and they are the things we are good at.
I want to be clear about a few other things, because I think they matter.
We do not use cloud-based AI services; everything runs on secure, private infrastructure, and no client content passes through public endpoints, ever. We do not apply AI to any client’s content without their explicit permission. These are not terms and conditions buried in a contract, they are how the platform is built. When you have spent a decade handling the most sensitive content in the industry, you learn what trust actually looks like at an engineering level.
How the workflows actually work
This is the part I find most exciting, and where I want to be specific about the principles without giving away how we have built it.
The standard approach to AI in our space is linear: run content through one tool, get an output, pass it to the next tool, review at the end. It is straightforward, but it is limited.
What we have built is different. Multiple specialised AI processes run in parallel against the same source material, each focused on a different dimension: transcription, speaker identification, scene detection, on-screen text recognition, descriptive indexing. Each one produces a structured output, and each output gets reviewed by a specialist operator before it is confirmed. The confirmed results combine into what is essentially a detailed map of the content, structured and machine-readable, that informs everything downstream.
The people are still doing the skilled work; they are just working with much better information than they had before.
The downstream effect is where it gets really interesting. Once content has been properly indexed like this, every subsequent process gets faster. Audio description scripts can be built from confirmed content descriptions instead of raw footage. Subtitle timing locks to confirmed scene boundaries, and translators start with accurate, structured source material rather than working from scratch.
I should say where we are with this honestly. Some of these workflows are in production, some are in advanced testing. The pace of development is faster than anything I have experienced in this industry, and the tools are getting better almost weekly. We are building and proving this right now, and the results so far have been genuinely encouraging.
Why our data matters
There is one more piece to this that is easy to miss. EIKON has been handling premium content for over a decade. That means we have years of operational data about what “good” looks like across thousands of content types, language pairs, and delivery specs.
That matters because AI on its own does not know what good looks like for a specific client, a specific language pair, or a specific content type. It does not know that subtitle timing conventions differ between Japanese and German, or that a particular studio has specific terminology requirements for their localised content. We know, because we have been doing the work.
That knowledge is what lets us deploy AI with confidence, and it is the difference between running an AI tool and running an AI workflow that actually meets the standard.
What comes next
AI is going to change this industry. I am more certain of that now than I was a year ago, because I have seen what is possible when you apply it carefully and in the right places.
But it has to be deployed by people who understand the work, the quality standards, and the security requirements that come with handling studio content. Demos are easy; delivery at broadcast quality, at scale, across dozens of languages, on private infrastructure, with full auditability; that is the hard part.
In the next article, I will describe how EIKON-iQ is evolving from the integration layer I described in Article 1 into something more ambitious: an intelligent orchestration platform that coordinates AI and human expertise across all of this at scale. The technology is only useful if the architecture can support it.
AI-powered servicing, human-perfected quality.
More In This Series
Article 1
The Integration Problem Nobody Talks About
The media supply chain has modernised dramatically, but there is a gap where automation meets human judgment. EIKON-iQ was built to close it.
Article 3
From Integration to Intelligence: Real-World AI Use Cases
When AI and human workflows run side by side across multiple content tiers and client requirements, how do you manage that complexity? How does a platform evolve from connecting systems to making intelligent decisions about how work flows?
Article 4
The Future of Content Servicing Is a Partnership
What does trust actually look like when AI is touching studio content? How does the relationship between content owners and service providers fundamentally change, and what does the governance framework need to deliver?
