Episode Transcript
[00:00:00] When Context becomes Infrastructure why the next phase of global content requires more than Better Prompts By Agustin Defino Deluchi Generative AI Gen AI was supposed to be the answer, and in many respects it delivered. Language models now produce fluent, contextually plausible content across dozens of languages at a scale and speed that would have been unimaginable five years ago.
[00:00:27] Organizations have built workflows around this capability. Budgets have been reallocated. Headcounts have been restructured. The industry moved fast and it captured real gains.
[00:00:38] So why isn't the friction going away? The same content evaluated as correct by every established measure performs radically differently across markets. Prompts that work well in one campaign context fail in another, not because the language is wrong, but because something upstream is missing.
[00:00:57] Quality assurance QA processes designed for source based pipelines are being applied to generative outputs that do not share the same logic. Review cycles grow longer, disagreements about ownership multiply. The tooling keeps improving, but the operational friction does not resolve. The instinct is to look for a better model, a smarter prompt, a more refined workflow. But that instinct is pointing in the wrong direction. The problem is not generation.
[00:01:27] Generation is largely solved. The problem is that the systems doing the generating have no stable, structured understanding of the environments for which they are generating. They produce content without owning any of the consequences that follow.
[00:01:41] This is not a generation problem.
[00:01:44] It is a representation problem. Two systems running in parallel Most organizations in the language industry are currently operating in a hybrid state that nobody designed. Source based pipelines built around translation memory, style guides, glossaries, and structured review exist alongside generative systems that operate on an entirely different logic. One paradigm assumes a source artifact, the other bypasses it. One is optimized for consistency and control, the other is optimized for speed and scale. These two systems are structurally incompatible. The friction that practitioners experience every day disagreements over review scope, confusion about quality ownership, difficulty comparing outputs across workflows is not a process failure. It is the predictable consequence of running two fundamentally different operating models in the same organization applied to the same content problems.
[00:02:44] The industry's instinct has been to reach for tooling better Computer assisted translation integrations machine translation post editing workflows Generative aware QHX these adjustments help at the margins, but they do not address the structural mismatch. The problem is not procedural it is architectural. What LLMs actually lack large language models are pattern engines of remarkable sophistication.
[00:03:16] Given a well constructed prompt with sufficient context, they will produce output that is fluent, plausible, and often genuinely good.
[00:03:25] This is not a small achievement, but there is a structural limitation that no amount of prompt refinement can overcome. LLMs have no persistent model of the world for which they are generating.
[00:03:36] Every generation call starts from scratch. The model has no memory of how a previous campaign performed in a given market. It carries no understanding of how a particular audience segment responds to different rhetorical registers. It cannot reason about the regulatory constraints that apply in one jurisdiction but not another, or track how those constraints have evolved. It produces output without owning any of the consequences that follow.
[00:04:02] The standard response to this limitation is to inject more context into the prompt. Audience Personas, market guidelines, tone of voice, documents, brand glossaries. This works up to a point. Injected context is better than no context. But there is a fundamental difference between context that is provided at generation time and context that is maintained as a persistent system. One is a workaround, the other is infrastructure. The distinction matters because injected context is static. It does not update when market conditions shift. It does not incorporate performance signals from previous deployments.
[00:04:41] It does not model how the same content will land differently in London versus Lagos or how that difference has changed over time. And it requires someone, usually a skilled practitioner, to reconstruct it from scratch with every new generation task. The Measurement Problem Quality frameworks in the localization industry were built for a different era.
[00:05:03] Segment level evaluation, checking accuracy, fluency, and terminology compliance unit by unit was the right tool when content moved through structured source driven pipelines. It still has value, but it does not tell you whether the content is working. This gap is becoming visible in a way it was not before.
[00:05:24] When content was expensive to produce, organizations had an implicit incentive to evaluate it carefully. When content is cheap to generate at scale, the volume of outputs quickly exceeds the capacity of any segment level QA process. And more importantly, a piece of content can be linguistically correct by every traditional measure and still fail to achieve its intended effect in the market. The question the industry has not yet answered systematically is what does good look like at the outcome level? Not whether a translation is accurate, but whether the content is performing, whether it is resonating with the intended audience, whether the behavioral or commercial signal it was designed to produce is materializing in the market. Answering that question requires a different kind of system.
[00:06:11] Not one that checks outputs after the fact, but one that can model outcomes before content is deployed.
[00:06:18] What other industries already Know the challenge of operating intelligently in complex variable environments is not unique to global content.
[00:06:27] Manufacturing, financial services, logistics, and autonomous systems have all confronted it and all arrived at a similar architectural build an explicit structured model of the environment you are operating in. In manufacturing, these are called digital twins, persistent structured representations of physical systems that are continuously updated with real world data and used to simulate outcomes before physical changes are made.
[00:06:56] In AI research, the analogous concept is a world model, an internal representation that allows an agent to reason about consequences rather than simply respond to inputs. In financial services, risk models serve a similar function, not predicting the future with certainty, but structuring what is known well enough to reason about what might happen.
[00:07:17] Global content has not built this layer. The industry has invested heavily in translation, memory, terminology management, and increasingly sophisticated AI generation tools, but it has not built persistent structured representations of the markets it operates in, representations that can be continuously updated, queried, and used to reason about how content will land before it is deployed. This is the missing layer. Not a better model, not a smarter prompt, a better system.
[00:07:48] Context as a system the architectural direction that addresses this gap is what I refer to as context. As a system treating context not as a set of instructions appended to a prompt, but as a persistent operational layer, structured, continuously maintained, and capable of reasoning about consequences before content goes live. A context system in this sense would represent a target market's linguistic norms, cultural signals, regulatory constraints, behavioral patterns, and performance history in a structured form that persists across generation tasks. It would update when conditions change, new regulations shifting sentiment, emerging terminology.
[00:08:32] Rather than requiring practitioners to manually reconstruct that knowledge every time, it would allow organizations to simulate how content is likely to land before deployment, not just evaluated after.
[00:08:45] This is not science fiction. The component capabilities exist Structured knowledge representations, schema defined audience models, feedback loop architectures, LLM driven reasoning over dynamic data.
[00:08:58] What has not yet been assembled is the full system and more importantly, the industry framing that makes clear why it is needed. The distinction that matters is Context provided is a workaround for a system that does not maintain it. Context maintained is infrastructure. One requires a practitioner to reconstruct what the system should already know.
[00:09:21] The other accumulates knowledge over time, compounds it, and makes it available to every generation task that follows. What this means for practitioners for localization managers, language technologists, and the language service professionals that serve them, the shift toward context systems is not a threat to expertise. It is an elevation of it. The knowledge that practitioners currently apply informal intuitions about how a particular market responds. Accumulated experience with a client's tone, awareness of regulatory sensitivities in specific jurisdictions is precisely what a context system needs to be built on.
[00:10:02] That knowledge, properly structured and maintained, becomes a compounding organizational asset rather than something that lives only in individual heads and gets lost when people move on the role shift. This implies is real. However, the practitioners who thrive in this environment are not primarily reviewers of generative output. They are architects of the domain models that govern how generation happens in the first place. They move upstream from correcting errors after the fact to defining the constraints and representations that prevent those errors from occurring, from reacting to launch content to modeling what will happen before it goes live. This is not a smaller role, it is a more consequential one, but it requires a different orientation, from linguistic evaluation to system design, from segment level quality to outcome level effectiveness. For LSPs, the strategic question is similar Organizations that build structured context systems accumulate a behavioral knowledge base about the markets they serve, a knowledge base that grows more valuable over time and is difficult for competitors to replicate quickly.
[00:11:10] Those that remain at the output layer, optimizing prompts, managing review cycles, delivering files, will find that layer increasingly commoditized. The Horizon the evolution of global content infrastructure has followed a clear trajectory from human translation to translation memory to neural MT to gen AI. Each step brought new capabilities, and each step also revealed a new limitation that the next step addressed generation has been the headline capability of the current phase. The limitation it has revealed is representation. The systems that will define the next phase are not primarily about generating better content.
[00:11:51] They are about knowing more persistently, structurally actionably about the environments that content operates in.
[00:11:59] Content is no longer the constraint.
[00:12:02] Consequences Context as a system is how the industry begins to address that. Digital twins mirror reality world models simulate it context as a system acts within it, and that is where Global Content Infrastructure is heading.
[00:12:19] This article was written by Agostin Dafino Delucchi. He is co founder of Trilogica Global and the originator of the Context as a System framework.
[00:12:28] A former director of global data and AI at Microsoft, he holds two patents related to multilingual content systems and has over 25 years of experience shaping global content infrastructure. He is a guest instructor in The University of Washington's Localization Certificate Program. Originally published by Multilingual Magazine, issue 255 September 2026.
