Last week, I explored how the streaming industry’s next control point is shifting from content ownership to ownership of viewer intent – and how conversational AI is reshaping discovery across every layer of the television stack.
That shift is what’s increasingly being referred to as AI television: a landscape where discovery is no longer a search bar or a static recommendation list, but a conversational layer capable of interpreting what a viewer actually wants, across streaming services, TV operating systems, and general-purpose AI assistants alike.
Now I want to break down what that shift demands in practice: commercially, infrastructurally, and in terms of the trust frameworks that will determine whether AI-powered discovery can be genuinely relied upon at scale.
Discovery is becoming a revenue issue
Discovery is often discussed as a consumer-experience problem, yet it is increasingly becoming a commercial one. Every failed search is a lost commercial opportunity: a subscription that doesn’t renew, an ad slot that goes unfilled, a piece of content that never earns back its investment.
The better a platform gets at discovery, the more of those moments convert into actual viewing sessions – which is why the stakes here extend well beyond user experience.
The most direct connection lies in session initiation. Reducing the time between intention and playback increases the number of completed viewing decisions and reduces abandoned sessions – and as Gracenote’s research showed, nearly half of viewers said difficulty finding content could lead them to cancel a service entirely.
And the longer-term commercial implications are more significant still. Traditional performance data tells a service what people watched, whereas intent data reveals what audiences wanted but could not find. That difference is valuable – unfulfilled demand that can identify missing genres, underserved audiences, emerging formats, and opportunities for content investment that retrospective datasets wouldn’t surface.
Realising that value, however, depends on AI systems being able to understand content accurately and execute recommendations reliably. That begins with the quality of the information those systems rely on.
Metadata is the foundation recommendations depends on
Long before questions of orchestration or governance come into play, the system needs to know what it’s allowed to recommend. Without trusted metadata to rely on – accurate information about rights, regional availability, live schedules, and advertising restrictions – AI models may confidently recommend a title that is unavailable or misrepresents the viewer’s request.
Getting this right depends on more than just the model itself. It requires rights intelligence, behavioural understanding, and reliable playback execution working together – so that when a system tells a viewer something is available, relevant and ready to watch, it actually is.
Beyond metadata: the case for an AI operating layer
Metadata alone, however, is only one component of the infrastructure required to deliver dependable AI-powered discovery. Accurate metadata might provide a system with what it’s allowed to recommend, but it doesn’t tell the system how to act on that safely, consistently, and across every surface a viewer might use to ask.
Behind every recommendation, a single system may need to operate across different applications, devices, cloud environments, and commercial relationships to return a nuanced response that reflects viewer intent.
This requires more than a language model or a standalone recommendation engine. It demands an operating layer capable of connecting models, agents, proprietary data, content systems, and infrastructure while preserving control over how decisions are made – determining which models are permitted to access particular data, how rights and commercial rules are applied, how agent actions are recorded and audited, how sensitive audience information is protected, and how inference costs are measured and controlled.
Retaining control in a fragmented landscape
As discovery expands beyond individual streaming applications into television operating systems, general-purpose assistants and emerging AI agents, media companies will need to decide how much control they are willing to surrender to external platforms. A governed AI operating layer offers another option: allowing companies to participate in new discovery environments while retaining authority over their data, infrastructure, policies, and commercial logic.
At Stelia, we see this as the role of an enterprise AI operating layer – not replacing metadata platforms, recommendation engines or consumer interfaces, but connecting them within a secure and governed environment in which AI can be deployed, monitored and scaled. Such an environment can orchestrate different models and agents, control access to proprietary data, record how decisions are made, measure infrastructure and inference costs, and help move AI capabilities from isolated demonstrations into dependable production.
The future of AI-enabled discovery in streaming will depend on the companies able to operationalise their intelligence securely, transparently and at scale.
Trust is the unresolved constraint
But crucially, AI discovery is not yet a fully trusted replacement for conventional search. Neither metadata nor an operating layer can guarantee that the viewer trusts the recommendations they are receiving. In fact, currently, most don’t.
Gracenote’s 2026 research among US AI chatbot users found growing interest in using AI to find entertainment – but 75% said they verified AI-generated recommendations elsewhere. In order to build genuinely effective AI systems, recommendations must be accurate, current, explainable, and commercially neutral enough to deserve confidence and elicit trust from their users.
The more influence AI has over distribution, the more important transparency becomes. Paid placement, house-content prioritisation, and sponsored recommendations will need clear regulations to ensure recommendations remain earned rather than bought.
Five priorities for media companies
The focus is no longer on whether to participate in AI-powered discovery, but on where to position within it and which decisions can become genuine sources of competitive differentiation.
- Build content intelligence.
Metadata can no longer be ignored; scene-level, contextual and rights-aware understanding is becoming core infrastructure rather than a data hygiene exercise.
- Model intention as well as preference.
A household profile built on historic viewing tells a system what someone has liked. It says nothing about what they want right now and increasingly, that’s the gap AI is being asked to close.
- Prepare to be found outside your own application.
As discovery moves into systems, general-purpose assistants and emerging agents, content needs to be understandable and accessible well beyond the interface a company controls directly.
- Tie discovery measurement to business outcomes.
Whether a new interface improves completed starts, retention, or advertising effectiveness is a commercial question, and it deserves commercial rigour rather than being treated as a UX metric in isolation.
- Decide where to sit in the future value chain.
Media companies will need to decide whether they own the content, the application, the intelligence layer, the audience data, or some combination of those.
What media companies should be building toward
That final decision will shape their leverage, partnerships, and the ability to compete as the route to the viewer becomes less fixed and more contested.
It is the platforms that create certainty in an age of infinite choice that will become more powerful than the ones that simply create more options.
But what makes that certainty achievable and trustworthy is the infrastructure built to deliver it.
