David Sugg has spent 26 years building and operating media supply chains at major studios. He's led large technology programs, managed teams of hundreds, and learned most of what matters by being in the room when things went wrong. He now consults on technology implementations, supply chain strategy, and operations, with a particular focus on how technology actually performs in production environments.
There’s a common question at every major media company, usually directed at Content Operations: “How long will it take and how much will it cost to service this deal?” Territory launch for the streaming platform. New sales deal with the ink still drying on the contract. Celebrity passes away, we need a retrospective. The answer is almost always some version of a mad scramble research project that crosses multiple asset repositories, people, rights, and clearance paths. It always takes longer than management, sales or the platform wants and the answer is rarely a happy one.
None of this is unsolvable. With enough people, time and money every one of those questions gets answered, and the deals get serviced. Answering that way has two costs, and we’ll explore the second one here. The first is that everything is slower and more expensive than it needs to be, which is the cost people notice and, unfortunately, budget for. The second is easy to overlook: when an answer takes weeks, the question gets asked less, and the ideas that depend on a cheap answer, the experiment on a territory, the plan that needs three tries to get right, mostly stay ideas. The times we did find a way to run that loop, through some combination of people and technology that could not always be repeated, produced some very large deals that would not have happened otherwise.
The expensive part was not the research. It was the ideas that stayed ideas because finding out cost too much.
Take Latin America
Every launch I was part of ran into a version of the same question: can the streaming platform launch in Latin America by June, and what will it cost? Senior leadership watches that one, because the success of the platform is treated as existential. Every deal has dates, though, and a missed delivery on a $100M sales deal gets the same attention.
Answering it means knowing which titles are available in the territory, under which rights, in which windows, and whether the materials exist to support the launch: the Spanish-language dubs and subtitles, the versions each platform spec needs, the artwork. The harder version of the question comes up when someone tries to plan exploitation across the whole business: the streaming platform, consumer products, licensing, ad sales, sports crossovers, broadcast integrations. Either way the answer is spread across several rights systems and a dozen or more supply chains, each built to acquire, warehouse and move its own content, and none of which agree on what a title is. So the plan gets built on what people remember and then checked afterward.
A dozen supply chains can each tell you what they have. None of them can tell you what the company has.
The shelf, and why it stayed that way
The content was not lost. It sits on what I will just call the shelf: every file the company holds, the state each one is in, and the data about it, spread across every repository that ever arrived with an acquisition or a business line.
The shelf is not new. Twenty-five years ago it was literally tapes on a shelf with some type of record that told you generally what was on them, and you did not really know until you pulled the tape, unless you serviced it often enough to remember or someone had been careful with the inventory tool. Files replaced the tapes, causing the problem to proliferate. Each supply chain’s asset management system knows its own content to the depth its operators recorded, and that depth depends on how rich the metadata was when the content arrived and whether anyone was funded to maintain it afterward. Ask whether there is a Spanish dub of a title for the Latin America launch and the answer may be sitting in another supply chain’s system, under another title record, at a frame rate the record does not mention, in a version nobody in your group knows about. The same holds further up the stack. There are several versions of a title and several edits of each. There are several rights systems. The title systems were often built as part of an ERP, so their records are financial instruments, not a description of what the content is. Asset management systems are part of any fix for this and I am not blaming them. They manage what they were given.
The shelf stayed that way for a reason I wrote about in May, in When the Human Layer Is the Control Layer. When knowing what was on the shelf got too big for the systems of the time, the industry gave the job to people. People were good enough at it that the better processes, and the tools to support them, never got built. It was always easier just to put more people on the problem. Much of that work could have been automated years ago with tools that existed long before large language models. It was not, because people were the stopgap, and a stopgap that works does not get replaced. Every proposal to fix the data ran into the same wall: clean it all first, with an army of people, and then build. That is the part that has changed. Implemented well, a language model can read the accumulated, inconsistent record as it is. The cleanup still happens, but it happens as the work runs, on the records the work touches, and that removes it as the reason to never start.
People were the stopgap, and a stopgap that works never gets replaced. What is new is a tool that can read the shelf as it is.
What a common language does
The previous piece, Own the Ends, argued for normalizing the two ends of the supply chain, intake and ordering, and leaving the middle alone. It closed on a rule: define one language that everything outside uses to reach the supply chain, and apply it wherever something connects, not only at the two ends. That piece called those points the edges; what matters is the language spoken at them. In that model, three things get normalized: how content is accessed (a title is requested, materials are requested, materials and the data about them are submitted), how demand enters the supply chain, and how tasks are managed once demand becomes work.
That language does not have to be an industry standard. In this model, it is a common language, a data contract, that every tool speaks whenever it connects to the supply chain. Every system that wants to participate in the content ecosystem must be able to read and write that lingua franca. How a system stores and processes data internally becomes irrelevant, because it is expected to speak the same canonical forms when it produces or consumes information from the ecosystem.
Data governance is part of that work. Deciding what a title is, what counts as a material, and what “ready” means for a given recipient is the governance, and in the common language it gets written once. Today it is negotiated system by system every time two of them have to talk. The common language is also how every title and every material gets one identity that every system is held to. It does not resolve the fifteen title records that already exist for one work; it is how that mapping gets enforced once it is made.
Everything that connects speaks that language. Say a post-production group brings in a new tool and wants it to push materials into the supply chain and pull them back out. They do not integrate with your APIs and you do not integrate with theirs. They read and write the common language, over a shared messaging infrastructure. The systems you already have connect the same way, through an adapter that speaks the language on their behalf. Nothing inside them has to change, which is what leaving the middle alone means in practice.
This is also how the common language reaches the shelf. Each repository speaks it through its adapter, so there is one way to ask every one of them what it has and what state it is in, and one way to order work on what comes back. What comes back is only as good as what that repository recorded. A common language does not clean the shelf. It gives you one way to ask, and one record of what was asked for, what was ordered, where it went and what it took: the operating record. Those are two different things. The operating record tells you what happened; the repositories tell you what exists, to the extent they ever knew. What agents need before they can do useful work is demand that is legible, task state that lives outside someone’s head, and orchestration that runs without the expert in the room, and the common language is how that gets built. The control layer in most media companies has been people; this is where it becomes something written down.
A common language does not clean the shelf. It makes it something a tool can be pointed at.
Agents speak it too
Now point a tool at the shelf for the Latin America launch, with an expert running the workflow. The expert is the operations person with the leather notebook, who has always known where things were and has never had a tool that could reach all of it. The agent does the part that used to take weeks. It asks every repository, in the common language, what exists for each title on the launch slate and what state each material is in. It reports what it found and where the answers disagree or are missing: the same title under three records, a Spanish dub whose record says nothing about its frame rate, a rights entry for Mexico that contradicts the contract. It proposes the orders that would close the gaps it can see. The expert resolves the ambiguities the record cannot, and decides. An agent cannot infer rights that were never recorded or the condition of a material nobody inspected. It can make every one of those unknowns visible in an afternoon, which nobody could do before.
That gives the expert back the loop the shelf took away. Ask whether Latin America could launch in June and an answer comes back in a reasonable timeframe, with the gaps listed. Ask again with Brazil in and Argentina out. Try the exploitation plan against the catalog before the meeting, find the titles with a rights problem, and walk in with a plan that already routes around them. The runs that fail cost an afternoon or, at most, a few days. The combination of people and technology that produced the deals I mentioned can now be built on purpose.
The reason that agent can be trusted on the shelf is that it connects the same way the post-production tool did, in the common language. Connected that way it has an identity of its own rather than a borrowed login, a permitted set of messages (the same request-a-title, request-materials and submit-materials capabilities everything else gets), and an audit trail that is kept for everything said in that language. A finite message set is a large part of least privilege: anything an agent says outside the language is rejected, and the systems on the shelf answer in an expected form. Which titles, which recipients and which rights conditions a given agent may act on is a policy enforced on the same connection, and it is the expert’s policy.
The work the agents take is the data moving and the human automation, which at many companies is most of what is left. Agents placed inside a processing system for a middle goal are a different case and not this piece’s. Seth Hallen at Hallstone Ventures put the requirement this way after Own the Ends: before autonomous agents can reliably manage complex media workflows, companies need consistent data, clear business rules, and an accurate record of decisions and transactions. The common language is how all three get written.
Connecting in the common language is what gives an agent an identity, a permitted set of messages and a record. The expert is who it reports to.
The question gets cheap
Back to Latin America. With the common language in place and agents working the shelf under the people who understand it, the launch question becomes a query: what exists for each title, what state it is in, what has to be ordered, and what each of those orders took last time. The exploitation plan is the same query run across the slate. How long and how much come from the operating record. What the same orders took last time, from the same suppliers, is the estimate, and it is already written down because it was written as the work happened. The common language did not answer the question. It made the tool that answers it buildable and governable, which is the difference between an agent you can put in front of your catalog and one you cannot.
The faster answer is the smaller, but still important, gain. The larger one is that the question gets cheap enough to ask early and often, and that is the reason a normalized supply chain is worth more than the operating savings Own the Ends described. The questions the business has been paying for by the week get asked the way they should have been all along, and the ideas that depended on a cheap answer stop staying ideas.
The industry is in the middle of its largest change in decades, and working out what comes next depends on understanding the business you already have. This piece is about how the question gets asked. How much of what sits on the shelf is actually knowable, after decades of thin records, is a different problem, and the archive, where the AI business cases concentrate, is the next piece.
Sources and further reading:
David Sugg, Own the Ends (Oct 11, 2026); The Supplier Inside Your Supplier (Sept 18, 2026); The Model Is a Supplier (Sept 13, 2026); When the Human Layer Is the Control Layer (May 23, 2026).
Seth Hallen (Hallstone Ventures), LinkedIn, October 2026, on Own the Ends.
Andy Beach and Lorenzo Zanni, Our IBC 2026 Analysis (Sept 25, 2026), where the agent identity, skill-audit and least-privilege questions were raised.
On standards for a common language, the nearest existing specs, none of which covers the whole of what this piece describes: MovieLabs, Asset Ordering, Delivery and Tracking (TR-META-AOD); MovieLabs, Ontology for Media Creation, a shared vocabulary that describes itself as “not a schema”; MovieLabs, Interoperability in Media Creation (2023), which notes that “in many cases, however, no standards exist for such APIs”; SMPTE, ST 2021 Broadcast Exchange Format (BXF); EBU and AMWA, FIMS; and, from the music industry, DDEX.
