A deep dive into agentic AI in media: adoption stats, real use cases, multi-agent infrastructure, governance, IP risks, and the ROI reshaping the industry.
Media companies stopped experimenting with agentic AI in 2025. They started running it in production. According to Google Cloud's 2025 ROI report, 54% of media and entertainment executives whose organizations use generative AI have also deployed AI agents in production, putting the sector ahead of most other industries on the agentic curve. These aren't chatbots. They're autonomous workflows that ingest, reason, and act across newsrooms, playout, ad ops, and rights.
This piece walks through what "agentic" actually means for media operators, where adoption sits today, which use cases have made it into production, the infrastructure powering them, and the governance work required to keep any of it defensible. The throughline: the returns show up when agents are wired into a real workflow with guardrails, not bolted onto a CMS.
What Agentic AI Means for Media (and How It Differs From Generative AI)
Generative AI drafts a headline when you ask. Agentic AI decides a headline needs drafting, drafts it, checks it against your style guide, publishes it to three platforms, and flags the ones that need a human editor. The shift is from prompt-response to outcome-pursuit.
From Prompts to Autonomous Workflows
Publishing consultancy Persistent frames the change as shifting from static processes to intelligent, adaptive publishing operations where systems pursue outcomes rather than wait for instructions, evaluating manuscripts, preparing audio adaptations, identifying rights opportunities, and coordinating downstream actions. A generative feature answers a question. An agent watches for a trigger, decides what to do, uses tools, and reports back.
Under the hood, most production agents run a ReAct loop: reason about the input, call a tool, observe the result, reason again. Our runtime wraps each agent with a system prompt, tools, knowledge bases, and optional MCP servers, so every step in the loop (tool calls, retrieval events, streamed tokens) is inspectable after the fact. Logging matters more here than for a chatbot, because every decision the agent makes has to be defensible in a legal or editorial review.
How Multi-Agent Systems Work in Media Production
A single agent handles a discrete task. Real media workflows chain many of them. Amagi's Newspulse, for instance, is an agentic system that extracts distinct storylines from live and file-based news and reframes them into publish-ready clips for social and digital destinations. Story detection, clipping, vertical reformatting, and policy checks all run as specialists coordinated toward one outcome: first-on-social publishing.
Multi-agent orchestration in media usually follows a supervisor pattern: one router, several specialists, bounded step counts to keep runs from spiraling. In Powabase, that's the shape of our orchestrations primitive. A coordinator delegates each step to the right entity agent and merges results before responding.
The State of Adoption: What the Data Shows
Executive sentiment turned into budget in 2025, and the survey data now backs that up.
Executive Sentiment and Budget Allocation in 2025
Google's survey shows how deep adoption has already gone relative to other verticals. That 54% production number covers everything from single-task creative assistants to sophisticated multi-agent systems combining advanced models with access to tools and data. CSI's IBC 2025 review framed the technology as the era-defining theme for the industry, with vendors across playout, MAM, and ad tech describing a future of "invisible teams of AI agents quietly orchestrating the industry's machinery".
The Pilot-to-Production Gap and Agentic Maturity Levels
The averages hide a gap. Cross-industry uses (security operations, software development, general creative assistance) show much higher adoption in media than industry-specific ones. Google's report notes plainly that content management, creation, and monetization use cases show lower adoption than horizontal ones. That's where the pilot-to-production gap actually lives.
A rough maturity ladder for media teams looks like this:
| Level | What's in production | Typical owner |
|---|---|---|
| 1. Assist | Single-prompt drafting, summarization | Individual journalists, producers |
| 2. Single-task agent | Metadata tagging, ad-break detection, transcript cleanup | Ops team on one workflow |
| 3. Multi-agent workflow | Newspulse-style clip pipelines, scheduling + rights + QC chains | Platform team |
| 4. Governed operating model | Cross-department agents with audit, HITL, policy controls | CTO / legal / editorial jointly |
Norsk's team put the risk plainly at IBC: there's real value in agentic systems, but only if you put in the work, or you end up with clever demos that are unreliable in real operations.
How Media Companies Are Using AI Agents
Three domains have moved fastest: newsrooms, the content supply chain, and monetization.
Publishers and Newsrooms
Publisher-side experiments are now live across major mastheads. Digiday's reporting notes Reuters has benefited from early testing of Thomson Reuters' generative tools and is preparing to embed agentic AI more deeply into newsroom workflows, following its parent's 2024 acquisition of agentic AI startup Materia. The patterns showing up repeatedly:
- Manuscript triage and first-pass evaluation
- Research assistance and source aggregation
- First-pass fact-checking against internal archives
- Format adaptation from long-form to newsletter to social
- Rights lookups against contract corpora
For publishers who want the mechanics of routing multi-step editorial work between agents, we go deeper in our overview of enterprise AI workflow automation and the use cases seeing the fastest returns.
Content Supply Chain Automation
The content supply chain is where agents are most visibly replacing manual work. Astro, Malaysia's largest media and broadcasting company, worked with TCS to deploy autonomous agents built on Amazon Bedrock and the NOVA multimodal model that analyze visuals, detect UI elements, and generate semantic metadata including titles, genres, and categories across its web and mobile apps. Metadata had been throttling personalization and time-to-market; now it doesn't.
On the scheduling side, Mediagenix has launched agentic capabilities for linear, FAST, and VOD schedule optimization, rights validation, and title management with governance and human oversight preserved throughout. Agents handle the pattern-matching bulk (metadata, ad-break detection, compliance checks) while humans handle exceptions and sign-off.
Monetization, Ad Optimization and Revenue
Ad-break detection is a particularly clean agentic use case: a bounded task with clear success criteria, high volume, and expensive human labor if done manually. Amagi's stack pushes further into monetization by reasoning through workflows to enrich content, generate artwork, schedule channels, and optimize ad placement end-to-end. The revenue argument is direct: better break placement, faster time-to-monetize on new content, less waste on manual QC.
The Infrastructure Behind Agentic Media Operations
None of this works without solid plumbing. Two pieces matter most: orchestration and connectivity.
Multi-Agent Orchestration and Runtimes
A media agent that transcribes a segment, drafts three social posts, checks brand safety, and schedules distribution is really four agents behind a coordinator. Building that from scratch means writing routing logic, state management, streaming, retry policies, and observability before you ever touch a use case.
That's the layer platforms are increasingly collapsing. Our runtime handles the ReAct loop, session state, SSE streaming, retrieval events, and tool logging as first-class primitives, with compute placed next to storage so retrieval and agent loops stay fast. Frameworks like Agno give you a lightweight Python object model but leave deployment and observability to you; managed RAG products like Vectara handle retrieval well but don't run agents. For media teams whose workloads combine large document corpora (rights contracts, style guides, archive metadata) with multi-agent execution, keeping both in one runtime removes the joins where agentic pipelines usually break.
Why Model Context Protocol (MCP) Matters
Media stacks are notoriously heterogeneous: a MAM here, a rights system there, a scheduler somewhere else. Model Context Protocol (MCP) is the standard that stops every integration from being bespoke. Mediagenix has explicitly added MCP support to its agentic operating model, and we support MCP servers as an agent attachment so tools from external systems appear alongside built-in tools without custom glue code. For an industry with 30 years of vendor sprawl, one discovery protocol is the difference between agentic pilots and agentic platforms.
Governance, Trust and Human-in-the-Loop
Autonomy in a newsroom or a rights workflow without controls is a legal problem waiting to happen.
Guardrails, Audit Trails and Enterprise Trust
Every serious vendor has converged on the same principles: policy-driven guardrails, human oversight at decision points, and full auditability. Amagi describes Newspulse's controls as policy-driven guardrails in plain English, with autonomous runs or human-in-the-loop as configurable modes. Eidosmedia positions its AI Content OS as the governed layer that sequences, audits, and monetizes AI across the content lifecycle, rather than bolting features onto a CMS.
At the runtime level, human-in-the-loop needs primitives, not just intent. Our agents support hooks that pause a run, emit an approval event, and resume once a human signs off, so a publishing agent can wait for an editor before it pushes to production instead of sprinting through.
IP, Copyright and Legal Risk
Legal exposure is the reason content-specific adoption trails cross-industry adoption. The IAB's December 2025 playbook on AI, IP, and digital advertising transactions is explicit about the pre-transaction diligence media buyers now owe themselves: understanding whether a model provider will train on your prompts and inputs, and the confidentiality and retention terms that flow from that decision. For media enterprises with valuable archives, the wrong default here doesn't just leak IP. It trains a competitor's model.
The practical answer is a runtime that keeps your data on your own isolated stack, with clear boundaries about what leaves it. Powabase gives each project its own dedicated Postgres, Realtime, and Storage with no shared logical databases and SOC 2 / ISO 27001 assumptions intact, which is the substrate governance teams tend to insist on before they'll approve anything autonomous.
Vendor Platforms Powering the Shift
The vendor map splits into three groups. Media-native specialists like Mediagenix, Amagi, Eidosmedia, and Accedo ship agents pre-wired to broadcast workflows, MAM, and playout. Systems integrators such as TCS, Persistent, and Google Cloud's professional services build custom agentic solutions on hyperscaler foundations, as the Astro case demonstrates. AI application platforms, Powabase among them, provide the runtime layer where media teams build their own agents against their own data, without stitching together a vector database, an agent framework, and an orchestration layer separately.
The choice usually comes down to how differentiated your workflow is. If linear scheduling is your workflow, buy the specialist. If your differentiation is your archive, your editorial voice, or a proprietary distribution model, you'll build, and the platform question becomes whether your team wants to assemble RAG, agents, and workflows from four vendors or run them on one. Against pure-RAG products like Vectara, we position RAG as one of five indexing strategies alongside a full agent runtime, workflows, and Postgres in the same project, not a separate service to integrate.
The ROI and Business Impact So Far
The measurable wins cluster in three areas. First, throughput on metadata and tagging: Astro's motivation was scalability and time-to-market on personalization. Second, speed on social and digital publishing, which is Newspulse's first-on-social framing with automated clip generation. Third, reduction of manual toil in scheduling and rights, the core of Mediagenix's operating model.
These aren't "AI wrote our reviews" stories. They're the same team shipping more, faster, with fewer errors.
The economics start to make sense when you count agent action costs directly. Web search calls in Powabase run $0.02–0.04 per call depending on tier, with LLM inference billed separately. For a metadata pipeline processing tens of thousands of assets, that per-call transparency is what turns "agentic AI" from a line item into a unit-economics conversation.
The productivity thesis Digiday found in publisher pilots (testing agents specifically to lift output per journalist and per editor) is the same thesis showing up in broadcaster operations, ad ops, and rights. The specific numbers vary by shop; the trend line is consistent across every source above.
Where This Goes Next
The 2025 adoption wave established that agents work in production for bounded, high-volume media tasks with clear guardrails. The 2026 question is how deep they go into editorial and monetization work, where IP sensitivity has kept a lid on autonomy. Three things will determine that: MCP maturity, so agents can drive the media stack without custom integrations; governance tooling that gives legal and editorial teams real audit trails; and runtimes that keep data isolated by default.
If you're evaluating where to start, pick one workflow with a measurable KPI (metadata coverage, time-to-publish, ad-break precision), wire an agent to it with human-in-the-loop on the risky step, and instrument every tool call. That's the shape of every media agentic project that has actually shipped this year.