Field Notes: Context as a System

For two years the industry has obsessed over models, prompts, and fine-tuning, and that work has paid off: content generation is faster, more fluent, and more scalable than ever. But teams keep hitting a ceiling. The output passes every linguistic quality check, yet how it actually performs across markets stays unpredictable, sometimes better than expected, sometimes worse, with no way to know which in advance. The problem is not the generation. It is that the system around it has no model of what performing well actually means.

Agustín Da Fieno Delucchi, co-founder of Trilogica Global and former director of Global Data and AI at Microsoft, calls the fix “context as a system.” Instead of reconstructing context from scratch and passing it along with every prompt, context becomes living infrastructure: a persistent, structured layer of operational data that the whole system infers from. It is not a replacement for better models; it is what makes better models actually land. And right now, that context-infrastructure side of the equation is significantly underdeveloped.

This reframes what localization measures against. Linguistic correctness, brand compliance, and regulatory adherence do not go away, but they are a floor, not a ceiling, and the industry has spent a long time treating the floor as the goal. What becomes possible is modeling content against structured market data before it ships, surfacing not just what is compliant but what is likely to resonate, which framing drives action and which reads as aggressive or carries an unintended connotation. The teams building that context infrastructure now accumulate an advantage that compounds; the ones waiting for the technology to mature will start from zero while competitors already operate from a base.

Key Insights

• Better generation has a ceiling. Optimizing models and prompts in isolation makes content more fluent and scalable, but the surrounding system still has no model of what “performing well” means, so market performance stays a matter of guesswork.
• Context should live in the architecture, not the prompt. When context is reconstructed from scratch at every step, each generation cycle starts blind. Treating it as persistent, structured operational data that the whole system infers from is what lets better models actually deliver.
• Correctness is the floor, not the goal. Linguistic, brand, and regulatory checks remain necessary, but they only confirm content is correct, not that it will resonate. The higher bar is choosing the content that matters for a specific market at a specific moment.
• Simulation before shipping is partly here, and the advantage compounds. The pieces already exist, LLMs, semantic and cultural inference, vector databases, sentiment and social signals, to flag cultural misalignment or register mismatch before publishing. What is still maturing is the feedback loop, and organizations that start building the infrastructure now accumulate market knowledge that sharpens over time.

More “Field Notes” Episodes

Explore more topics in our Field Notes series, where we break down complex localization concepts, ideas, and experiments for industry professionals. Check out our other discussions here.

Field Notes – Episode 15: Context as a System

Below is an automated transcript of this episode

Stephanie Harris-Yee: Hello, welcome to Field Notes. I’m Stephanie, and today we have a bit of a special episode with a special guest. So we have Agustín Da Fieno Delucchi, the co-founder of Trilogica Global and former director of Global Data and AI at Microsoft, and someone who’s been shaping global content systems for over 25 years.

So today he’s gonna walk us through something he likes to call context as a system. So really what it means to treat context as this living infrastructure rather than just a prompt, and then also what it changes for localization leaders. So Agustín, thanks so much for joining us.

Agustín Da Fieno Delucchi: Thank you so much for having me here, Stephanie.

Stephanie Harris-Yee: So to dive in then, we’ve spent these last couple years in the industry really obsessing over some things about which models to be using with AI, how to prompt better, how to fine-tune things, and then with this concept, you’ve been somewhat arguing that this is the wrong conversation.

So what is that wall that you’re seeing teams hit when they’re focusing just on those things?

Agustín Da Fieno Delucchi: Yeah, absolutely. But let me actually reframe it slightly. So I think that it’s not that the conversation about models and prompting is wrong. That work matters and it’s delivered real gains. I think that the issue is that it’s becoming insufficient on its own.

It’s that teams have gotten very good at generation. Faster, more fluent, more scalable. And that progress is real and worth keeping. But there is a ceiling you hit when you’re optimizing for generation in isolation from a much broader system. The output gets better, but yet organizations are still finding themselves surprised by how content performs across markets.

And not because the generation has failed, but because the system around it has no model of what performing well actually means. So a bit of, like, guesswork, if you like. A pattern I see repeatedly: teams roll out AI generating product content across, let’s say, 15 markets.

Linguistically, it passes every quality check with flying colors, no doubt about it. But content performance across those markets is essentially unpredictable. Sometimes better than expected, sometimes worse. And the system can’t tell you which of those in advance because it has basically no memory of what worked in those markets before, no structured model of who the audience is you’re writing for.

Each generation cycle starts from scratch. Better prompts helped at the margins. But what is actually needed is a more systemic approach, one where context isn’t reconstructed each time from scratch and passed along every single time, but something that lives in the architecture as a structure and, let’s say, persistent operational data that the whole system infers from.

And that’s not instead of the better models. We still need those. It is what makes better models actually land. The two have to work together, and right now, the context infrastructure side of that equation is significantly underdeveloped.

Stephanie Harris-Yee: So you’ve framed this whole concept as like the next step of progression after 40 years. We had linguistic control, we had process context, then content systems, data-driven context, now generational AI. So this would be that next shift to have that contextual ecosystem, shall we say.

So what do you see is a little bit different about this shift as compared to things in the past?

Agustín Da Fieno Delucchi: So every previous shift, I think, was about getting closer to the content itself. More control, more consistency, more scale. Translation memory gave us linguistic reuse. Style guides keep giving us process control. Content management is giving us structure.

Stephanie Harris-Yee: Structure.

Agustín Da Fieno Delucchi: But data-driven approaches give us feedback signals after the fact. Generative AI is giving us the ability to produce at a completely different scale. So each step is real progress, and we’re seeing it. But they were all still about the content artifact, the sentence, the segment, the asset.

And the bar we measure it against has always been essentially the same. Is this correct? Does it comply with the guidelines? Does it really pass the style check? I think those guidelines are still necessary, and I want to be clear about linguistic correctness, brand compliance, regulatory adherence.

Those don’t go away. They still need to be there, but they are a floor, not a ceiling. So for a long time we’ve been treating the floor as the goal. Think about, you know, a major retailer manages seasonal campaigns across let’s say 20 markets. Today, even with sophisticated tooling, the approach is essentially produce content, localize it, ship it, and wait to see what happens.

The feedback loop is reactive. You learn something went wrong when the market tells you it did. What becomes possible now is to move well beyond correctness. The model that runs campaigns against structured market data before it ships. And it shows and surfaces not just what’s technically compliant, but what’s actually going to resonate. Which words carry the most weight for this audience, which framing drives action, and which one gets ignored. And that is even before a single asset goes out.

And that’s a fundamentally higher bar. We are not just getting better at producing content. We are starting to get serious about choosing the content that matters for a specific market at a specific moment. And that is a new territory for the industry at that level of detail. Of course, with all the assets we have today, there are more generic indicators and there is responsibility about it, but the opportunity to unleash all this extra stuff is big.

Stephanie Harris-Yee: So this concept of kind of simulation before shipping even, so predicting how content will land in the market before it goes out, what does that actually look like in practice? Is that something that’s real today, or is it something that we’re heading towards?

Agustín Da Fieno Delucchi: Yes, and it’s partially real and partially where we are heading. And I think it’s important that we are honest about that distinction. We can today build structured representations of audiences and markets. Not just demographic data. We’re talking about behavioral patterns, linguistic preference, emotional registers, what resonated and what’s fallen flat. We can track that.

The technologies making this possible right now are things most teams already have access to. Large language models, semantic analysis, cultural inference, vector databases for storing and retrieving market and latest knowledge. Sentiment analysis, you can do that on social listening tools that we all use that feed real behavioral signal into the system. And also structured data schemas that formalize what we know about a given market into something that we can actually interpret.

What’s less common is the architectural discipline to connect those pieces into a coherent system rather than using them as isolated tools. And I think that’s where the main difference is. We can run content against those models before deploying and then surface those risks. Cultural misalignment, semantic drift, register mismatch, all that.

But if you’re launching, for example, a financial services campaign in Brazil and Japan simultaneously, let’s say, a structured content system can flag before anything is published that the directness of the call to action is likely to read aggressive in one market, and the visual metaphor anchoring the copy carries an unintended connotation in the other.

And that’s not guesswork. It’s the system really interpreting a model of those audiences that has been built and refined over time. And today that kind of signal often comes from a human reviewer after the fact. And it comes if it comes at all. I think what is still developing is the feedback loop. This ability to continuously refine the model, to retrofit based on the actual performance so the system gets better at predicting outcomes over time.

And that’s the full version of it, is closing the loop and finalizing the circle. And we are not uniformly there yet. But here’s what I would say to someone skeptical: the organizations that start building that context infrastructure now are accumulating something that really compounds. Every market signal, every outcome, every data point and refinement goes into a system that gets sharper. And not just by updating translation memory, not just by updating terminology, by updating real market signals.

The ones waiting for the technology to mature before they start building will be starting from zero when their competitors are already operating from a significant base.

Stephanie Harris-Yee: So let’s bring it back to, say, like the localization manager, right? So you talk about these sort of experts who are now moving from that reviewing the outputs to then designing the system, from the end of the line to the instrument panel, as it were.

For a localization lead listening right now, what does this actually mean for how they would spend their week six months from now?

Agustín Da Fieno Delucchi: Right, it’s a great question. So the two things that I have very practically. The first is going back to the drawing board and how context actually travels through your pipeline, through your workflow. Now, not how it’s supposed to travel, because we know that we all have ideal situations, but how it actually happens can be different. So map it. That’s important. Where is it that the context is created? Where is it handed off? Where exactly does it get dropped?

Most teams, when they do this honestly, find that context is being reconstructed from scratch, if at all, at multiple points during the pipeline because there is no persistent place where it lives. That map is your starting point. It tells you where the architecture needs attention.

The second, I think, is like look at your current system and ask honestly whether they can tell you or they can anticipate performance. Not measuring it after the fact or separately, but modeling it before the content goes out and is created. Can your system tell you really before deployment that a given piece of content is likely to underperform in a specific market? If the answer is no, that’s the gap you’re designing towards.

So six months from now, a localization lead doing this work isn’t spending their weekend reviewing outputs. They are asking, “What does the system know about this market? What is it telling us before we ship?” That’s a fundamentally different professional orientation, and I think a significantly more valuable one, where you are looking at the signals that the system is producing, understanding why, and then modeling the system to actually do that. And that requires a lot of also linguistic effort in fine-tuning, but most importantly, it’s about understanding your markets, understanding where you’re going, mapping that into your brand, into your content goals and all that.

Stephanie Harris-Yee: We’re about out of time, and we’ve just scratched the surface here. So is there somewhere online where someone could go to either connect with you or to learn more about these concepts?

Agustín Da Fieno Delucchi: Yes, absolutely. I think the best place to start is LinkedIn. I publish regularly on this work, including pieces of the framework we’ve been talking about today.

And if you want to go deeper into what Trilogica Global is building in this space, the website is trilogicaglobal.com, and there is context there on our approach to global content systems and where this work is heading. We are also developing more formal published materials, articles, additional talk sessions, so following either of those channels is perhaps the best way to connect.

Stephanie Harris-Yee: So yeah, we’ll include those links in our episode pages. So yeah, feel free. Go ahead and connect with him there. Okay. Agustín, thank you so much for joining us.

Agustín Da Fieno Delucchi: Thank you so much, Stephanie.

Share this post