APB+ManagementNews & Events

Authority without accountability: Automating agentic AI without whole of tech industry support is very disconcerting & dangerous

By Joe Tan

In recent years, the industry’s AI conversation has focused on what ends up on screen – synthetic voices, generated artwork, AI-assisted editing. It has been a comfortable conversation because the cost of getting it wrong is mostly aesthetic. A quieter shift has been under way behind the output, in the parts of the operation viewers never see: ingest, quality control metadata, scheduling and playout. There, software has begun to do something it did not do before. It chooses. 

That difference in stakes explains why operations lagged the creative discussion. “Broadcast operations have a much lower tolerance for error,” said Vijay P Sankar, Director of Product Marketing at Amagi, and Zubair Kolia, the company’s Senior Vice-President and General Manager for Managed Services.

They told APB+, “If an AI tool produces poor artwork, you can discard it. If it makes the wrong scheduling, rights, or ad-break decision inside a live workflow, viewers may see the result and the broadcaster carries the consequences.” 

Operations is also the harder problem because no decision sits in isolation. A system may need to understand content, rights, schedules, channel policies and distribution requirements before acting, with that context usually spread across several systems. Bounded autonomy is already real: agents enrich metadata, generate captions, prepare artwork, identify ad breaks and build schedules. Amagi’s Smart Scheduler, for example, assembles schedules around rights and programming rules while a programmer reviews and publishes them. 

How far that extends beyond individual tasks is another matter. “Agentic AI today is much more mature in well-defined tasks than in autonomous, multi-vendor operations,” said Ali Hodjat, Senior Director of Marketing at Telestream. For Sankar and Kolia, the question worth asking has changed. “The real question is no longer whether the AI can perform the task. It is whether the broadcaster can understand why a decision was made, control what the system is allowed to do and intervene when something falls outside the rules.” 

An increment, or a break

Broadcast has absorbed a long run of incremental change, making it tempting to file agentic AI as the next instalment – automation, then orchestration, now agents. Amagi’s Sankar and Kolia resist that reading. Earlier increments were technological rather than operational: 720p to 1080, 1080 to high dynamic range (HDR), now ST 2110, each absorbed by buying new hardware and carrying on with the same workflows and organisational charts. 

“The picture quality got better, but the way we work didn’t fundamentally change.” Agentic AI, they argue, belongs in another category. “I’d call it a revolution, not an evolution, because it changes operations end-to-end.” 

Technologically, the line still looks clean. Automation follows an instruction; orchestration connects several instructions into a workflow; an agent assesses a situation and decides what to do next. The third step breaks with the other two because software is choosing between valid courses of action instead of executing a predetermined path. 

Take a scheduled asset that has not arrived. Automation raises an alarm; orchestration triggers a predefined fallback; an agent can weigh the schedule, available content, rights windows and repetition rules, recommend a replacement and explain its reasoninรg, escalating when nothing clears the threshold. That bounded discretion is the dividing line. As Sankar and Kolia cautioned, “a conversational interface on top of a rules engine does not make a system agentic.” 

That discretion runs the length of the chain, from ingest through QC, metadata, rights enforcement and scheduling to playout. “It means we have to take a genuinely different look at how the whole pipeline is designed.”

The high-stakes end

If risk rises with consequences, compliance and quality control ought to be the last places autonomy arrives. Interra Systems argues the reverse. “Compliance and QC are where autonomy delivers the most value, precisely because the stakes are high,” said Bibhas Kumar Samanta, Sales Head for India and MENA at Interra Systems. 

High-volume, repetitive checks – video and audio faults, captioning, standards compliance – are what automated systems handle more consistently than a person watching for 12 hours straight, he added.

What matters is the sequencing. “We would recommend treating autonomy as a ladder, not a light switch,” Samanta said. Start with humans reviewing every automated result, then hand over lower-risk decisions progressively while keeping a person accountable for anything ambiguous or high-impact. “Get that sequencing right, and then instead of holding back the autonomy, it becomes the driving factor for delivering the value.” 

QC also offers a rare test of the old promise that automation frees skilled people for higher-value work. Two decades ago, teams caught faults by eye and ear in real time; software now does the bulk of it, and the operator’s job has moved to calibrating thresholds, reviewing uncertain cases, and reading patterns across a growing library. “It’s a shift from doing the check to trusting, tuning, and validating the QC system that does it,” Samanta said. The engine underneath has to be accurate enough to deserve that trust. 

Where accountability stops

All three organisations arrive at the same fault line, and it is not a technical one. “An agent can be given authority, but it cannot carry accountability,” Sankar and Kolia said. In broadcast, those two things have already travelled together in broadcast. Agentic systems pull them apart. 

Their remedy is to assign the parts explicitly. The broadcaster remains accountable for what reaches air. The technology provider defines the system’s limits, making it actions observable and ensures it fails safely. The operating team sets the policies, permissions and escalation threshold “that determine how much rope the agent gets.” Without that clarity, they warn, a system acting across ingest, QC, metadata, rights, scheduling and playout “will otherwise accumulate authority that nobody consciously granted it.” 

Hodjat’s concern begins where most facilities actually live, with equipment from a dozen suppliers and each layer becoming more autonomous on its own schedule. “As AI agents begin to interact with multiple systems, accountability can’t stop at the boundary of any one application or vendor,” he said. The industry needs clearer on what an agent is authorised to do, how its decisions are audited, when human approval is required and who owns the outcome. His answer is not more intelligence: “Open interfaces, strong observability, and human oversight will be just as important as the intelligence of the agents themselves.” 

Who writes those rules is itself unsettled. Samanta’s view is that quality standards are set mostly by broadcasters and streaming platforms, who have to retain viewers, while regulation attaches to specific, measurable risks: loudness, photosensitive-epilepsy guidance, captioning, and AI-content labelling. Beyond censorship rules, he said, that is largely where regulatory reach stops across APAC, and he does not frame it as contest: “We don’t see this as regulators versus broadcasters versus vendors. In practice it’s collaborative.” 

On what replaces a human signature, Interra lands close to Amagi. “We believe the final go/no-go on any broadcast-critical decision should remain human. What changes is how that human is making the decision,” Samanta said. Trust, he added, “will come from systems that are explainable and production-proven, with a documented decision trail a human can stand behind, not simply automation and a black box.” 

What to be sceptical of

The distance between a convincing demonstration and a live facility is an old problem, and Hodjat thinks agentic AI has widened it. “Broadcasters should be sceptical of claims that suggest human oversight can simply disappear,” he said. “The important questions are less about whether an agent can perform a task and more about how reliably it performs, what context it understands, what systems it can access, and what happens when it gets something wrong.” 

Commercial pressure complicates the picture, since broadcasters are being asked to hand over operational control during a period of cost discipline. Those pressures can align, Hodjat said, but cost reduction alone is the wrong reason to pursue autonomy. The risk arrives when autonomy becomes a way of removing people from the process. “The objective should be fewer unnecessary human interventions, not the elimination of human judgement. Cost efficiency can be an outcome, but trust and operational reliability have to remain the foundation.” 

Sankar and Kolia located the same failure earlier, in how the question is framed. Too much of the debate opens with “can we automate this?” – which no longer narrows anything, because with agents the answer is increasingly yes. They would put two others in its place: if the system gets this wrong, how serious are the consequences, and how easily can they be reversed? 

The regional test

APAC broadcasters are routinely cast as followers of North America and Europe. “I think the follower label is too simplistic,” Sankar and Kolia said, suggesting it reflects where technology stories get covered more than where difficult operational problems are actually being solved. Broadcasters here work across multiple languages, markets, regulatory environments and distribution models, serving linear, streaming, mobile and social audiences without letting operating costs grow at the same rate. 

What outside coverage misses is the shape adoption takes. It rarely arrives as a large autonomous control room project, surfacing instead inside less visible workflows: localisation, metadata protection, regional channel versioning, scheduling, news and sports clipping, and operational monitoring. Nor is the region one market: Australia, Japan, South Korea, India and South-East Asia differ in infrastructure, regulation and attitude to risk, and broadcasters may automate localisation aggressively while staying cautious about editorial calls. “That is sensible risk management, not slow adoption.” 

Which leaves a choice the industry can make now or discover later. “Five years from now, success shouldn’t be measured by how many decisions we removed people from, but by whether we created systems that are more efficient while remaining observable, accountable, and trusted,” Hodjat said. 

Governance, transparency and the boundaries of human oversight are much harder to retrofit once autonomous workflows are embedded in daily operations. Sankar and Kolia put the warning in sharper terms: less human involvement in each step is the point of the exercise, but less accountability for the operation as a whole is the failure mode, and if the industry defaults to whatever the technology permits, “it will erode invisibly until something goes to air that nobody chose to put there.”

Show More

Related Articles

Back to top button

Subscribe to the latest news now!

 

    Close