AI graduates from pilot to production

For several years, broadcast technology vendors have described AI primarily in terms of what it might do.

Across the Industry Insights roundtable conversations on cloud production, that framing is shifting.

Specific products are deployed at named customers, and the conversation is moving from internal proof-of-concept to operational use.

“As AI moves from experimentation to production, companies are able to apply language models and agentic architectures to better handle unpredictable demand and disruptions, enabling greater virtualization of remote production, as well as improved customer engagement and content localization,” said Ian McPherson, head of business development at TMT Insights, in the Industry Insights roundtable on cloud production.

The progression McPherson describes is consistent with how technology cycles typically mature in broadcast: a period of vendor demonstrations, followed by selective deployment with reference customers, followed by gradual normalization across workflows. AI appears to be moving into the second phase.

A named deployment

The most concrete example surfaced in the roundtable conversations came from AWS, which described production deployments tied to specific broadcasters.

“Currently, we’re seeing a continued move toward personalization, and, more recently, vertical video as Gen-Z audiences consume 88% of streaming content on their phones. With this in mind, we recently launched AI-powered solution AWS Elemental Inference, which broadcasters including NBCUniversal and Fox Sports Digital are using to optimize live video feeds for 16:9 viewing,” said Nina Walsh, global leader, business development, GTM and solutions, media and entertainment, games and sports at AWS.

The deployment is narrow in scope but it represents the kind of bounded, measurable AI application that broadcasters tend to accept first. The function is well-defined, the success criteria are observable and the failure mode is recoverable.

Advertisement

That pattern matches how AI adoption is moving in adjacent industries: starting with tasks that involve transformation or classification rather than judgment, then expanding from there.

What is driving the deployment

The reasoning behind the AWS product launch sits at the intersection of audience behavior and production economics. Gen-Z consumption habits are pushing broadcasters to deliver content in formats that traditional production chains were not built to handle at scale. Reformatting by hand is costly. Reformatting algorithmically is a fit for machine learning models.

Walsh tied the broader role of AI in media directly to cloud adoption, arguing that cloud infrastructure is what is broadening access to AI tooling and accelerating innovation across the industry.

The argument is that AI’s deployment timeline in broadcast is closely tied to cloud infrastructure adoption. Models, training data and inference services live in cloud environments. Broadcasters with established cloud workflows can access those services without rebuilding their stacks. Those without have to absorb infrastructure costs and AI deployment costs simultaneously.

The organizational angle

Walsh also described an organizational shift that AI is accelerating. Production, engineering and IT teams are working together earlier in the workflow as new tools come into the evaluation pipeline.

“Engineering and IT are evaluating new AI tools and ways of working, then production is testing them, and vice versa. Cloud-based creative studio Untold Studios is a great example of this,” said Walsh.

The cross-functional pattern matters because it determines how quickly AI capabilities reach actual production. In hardware-centric workflows, evaluating new tools often required dedicated procurement cycles and integration projects. In cloud workflows, teams can pilot AI services on existing infrastructure and decide whether to scale them based on real outcomes.

That speeds the path from experimentation to production but raises a separate question about how broadcasters evaluate model performance, output quality and the operational implications of automation.

The forward-looking view

While the current AI deployments tend to be narrow and task-specific, vendor commentary on the future of cloud production is more ambitious. The most expansive framing came from Globecast, which described agentic AI as part of the operational layer in a mature cloud environment.

“A mature cloud production environment will be defined by seamless integration between cloud, edge, and on-prem systems, with an orchestration layer assisted by Agentic AI. Multi-agent AI will play a larger role in optimizing performance and even cost in real time. Ultimately, the goal is an infrastructure that is both highly adaptable, predictable and operationally hyper-responsive, supporting a wide range of production needs that can easily manage dynamic changes and complexity in real time,” said Chris Pulis, CTO at Globecast.

Pulis described a future in which AI agents make ongoing operational decisions — scaling resources, balancing cost against performance and responding to dynamic changes without human intervention. That description is several maturity steps beyond a deployment such as AWS Elemental Inference.

The maturity gap

The gap between current AI deployment and the agentic future shows up most clearly in how vendors describe AI’s role in the workflow. Guillaume Aubuchon, VP of product management at Avid, framed the role conservatively, with professionals retaining decision authority.

Advertisement

“AI will increasingly act as a co-pilot across the media lifecycle, guiding decisions and automating routine work, while professionals remain in control. The organizations that succeed will be those that make the shift,” said Aubuchon.

The co-pilot framing is more limited than the multi-agent orchestration framing. Co-pilots assist; agents act. The vocabulary used by different vendors signals different positions on where AI sits on the autonomy spectrum and how soon it can be trusted with operational decisions in a live environment.

Both framings can be true at once. Broadcasters are likely to deploy AI as a co-pilot for tasks where errors are easy to catch and correct, and to test agentic systems in lower-stakes parts of the workflow before extending them to live production.

One reason AI deployment in broadcast has moved cautiously is that content rights and data privacy create constraints that do not apply equally in other industries. Walsh described the issue in operational terms.

“With the rise of AI and use of external tools, content owners want to make sure any data they’re sharing remains private and isn’t used to train base models,” said Walsh.

That requirement shapes how AI services are integrated into broadcast workflows. Vendors that can demonstrate that customer content is segregated from model training pipelines have an easier path to enterprise deployments than those that cannot. The question is not only whether the AI works, but whether the deployment respects the rights structure of the content it touches.

Where the curve sits today

Putting the roundtable commentary together, broadcast AI deployment sits at a specific point on the maturity curve.

Bounded, task-specific applications are running in production at named broadcasters. Cross-functional teams are evaluating additional services. Vendors are describing more ambitious agentic architectures as the destination rather than the current state.

The gap between current and future is narrower than it was a year ago but wider than vendor marketing tends to suggest.

The next phase of broadcast AI is likely to look less like the autonomous orchestration described in forward-looking statements, and more like steadily widening use of the kind of task-specific deployments AWS Elemental Inference represents — bounded functions, observable outputs and clear lines of accountability.

Advertisement