• Summary:
  • Francessca Vasquez, AWS’s VP of Frontier AI Engineering and Services, unpacks AWS plans to align human-focused Forward Deployed Engineering with its new AWS Context knowledge graph service informed by nine years of watching enterprise AI succeed and stall.

Early AI aspirations to deploy advanced AI tools to modernize the enterprise are bumping up against real gaps in the supporting role of what is increasingly being called “context” and human expertise. In 2026, context deserves quote marks because it is so widely used in different ways to basically characterize all of the challenges with quality data and mapping AI use cases to what humans actually meant when they curated a data set for a particular purpose.

AWS is hoping to bring a little order and sensibility to the various processes of figuring out what the original data was supposed to mean in the first place and how it may make sense for various new use cases, called AWS Context. It’s basically a service, infrastructure, and governed agentic services layer for informing agents when, why, and how they might repurpose existing data assets and services for new workflows.

In concert with all of this, AWS recently announced a $1 billion Forward Deployed Engineering (FDE) organization that combines human best practices distilled from helping enterprises build a semantic layer that supports new AI and agentic use cases. The thesis is that this new organization can help enterprise teams connect this semantic layer to enterprise data sources, enrich metadata, and use AI to publish a governed, versioned knowledge graph. This promises to provide the foundation over which agents can reason. And here is the biggie - this knowledge graph will somehow encode the domain expertise that lives in the customer’s code rather than the institutional knowledge that could rotate off.

I think it’s important to clarify more broadly that the terms FDE and knowledge graphs are more aspirational directions than technical terms. In 2021, I recall reading one professional services firm declare 2021 as the year of the knowledge graph with one practical example of a company basically joining two databases together with a new schema. And in 2026, pretty much all vendors with AI aspirations are looking to build a more scalable framing for front-line experts popularized by Palantir’s framing of the FDE in connecting an existing data estate for new use cases.

AWS efforts here, because of their outsize presence in cloud infrastructure more broadly and the $1 billion investment in the people tasked with extending this to new use cases, promise to move the salient edge of knowledge graphs and the FDE model forward more broadly.

And all this comes at a time when all of the hyperscalers and frontier labs are converging on Palantir’s twenty-year history in fleshing out the embedded engineering model. Microsoft’s $2.5 billion Frontier Company in early July, Google Cloud’s FDE hiring wave in May, Accenture’s FDE practices with Microsoft, ServiceNow and SAP, Deloitte in December, IFS's skin in the game, and Salesforce’s commitment to a thousand FDEs. The label has gone from Palantir quirk to industry default in under a year.

But then what are these new FDEs doing that’s actually new and different from each other and Palantir? Francessca Vasquez, AWS’s VP of Frontier AI Engineering and Services, helpfully unpacks the AWS version of this story:

We’ve been on this journey for several years, starting with our solution architects and our machine learning solution lab, where in the early days of AI and ML, we had teams that were out helping customers define use cases. We then, three years ago, established a group called the Gen AI Innovation Center that was solely focused on helping take those use cases into working prototypes, and then to production.

The new FDE organization is the third iteration of a services model AWS has been running and revising since before the current wave, with thousands of engagements behind it. Vasquez predicts the next wave will need to focus on context and composability:

A lot of the value of FDE in today’s world is going to happen with co-building with customers and their contextual intelligence, and that matters. So much of the value, to your point, of what will they leave with, what about the skills, what about the culture — a lot of that is being developed through various forms of knowledge graphing, semantic layers, ontologies, kind of pick your flavor. And our FDE outcome will be highly indexed on just being able to offer that contextual intelligence directly with customers.

Context about context

In 2026, context seems to be so widely used in so many different ways it tends to take a minute or two to figure out what people are pointing to. What until recently essentially meant the background or thinking about something within its periphery has come to take on a more technical bent of somehow grounding these agentic things to essentially make less stuff up and hopefully align it with what the humans are asking for (assuming they can all agree on this).

Last year's context was all about narrowing the filter for LLM processing with more precise retrieval-augmented generation (RAG). Now the focus is expanding to widening this filter informed by subject matter experts. Vasquez elaborates:

There definitely is a new sort of direction around context. And to my perspective, and what we see with customers, our ability to help our customers have deeper semantic context — and what fuels that is we have to also help customers get their data in a state that they can make use. So I think it’s a little bit more than just scripting and tools. It’s making it very relevant in a customer’s environment, where each of our customers has different constraints around data sovereignty controls. And that is where I think we will add a different capability around value for them, in their environments, with our accounts.

The FDEs across vendors are likely incentivized to advance their particular take on best practices that align with their sales model. And AWS in particular has a lot of models on offer that can complement what the frontier labs are doing. Vasquez explains:

Anthropic and OpenAI are great partners of ours, by the way, and so one of the natural benefits I think that we have is we tend to be model agnostic, and we can meet customers where they are. So in some cases we’re driving FDE capability that could be on the OpenAI models with Bedrock, and or on Anthropic’s. And so that’s like one good thing — we have a great partnership with them, and I foresee our ongoing direction for FDE in close partnership with the additional model providers.

Elsewhere she extends the same generosity to tooling, listing AWS’s own spec-driven Kiro alongside “third-party model provider coding tools like Codex and Anthropic’s Claude Code.” More broadly, this is part of the growing recognition that it’s not about locking customers to a particular model, but rather helping enterprises develop their own internal capabilities for applying new AI models to new and more helpfully curated data models.

And unlike even the best traditional professional services engagement, the ultimate goal of the AWS version is a kind of graduation to enterprise self-sufficiency informed by successful forward momentum. Vasquez says:

We don’t think it behooves customers to have embedded FDEs to operate on their behalf for years. We want to be able to get customers self-sufficient.

She went further when I asked what success looks like a decade out:

I always tell my team, a good outcome is if we stand up a business to help, and then in a couple years our customers then say we’re so good at it, or the product and innovation is so seamless and frictionless and great to use, that then they don’t need our help. And to me, that’s always going to be a continuum flywheel — we will build better products and services, our customers will get value, they’ll upskill in the process, and by the way, it may lower the cost structure as well, and that’s an okay thing for customers too.

It should be noted that virtually every new data stack vendor is coming up with their own framing for the same goal. Snowflake announced Horizon Context and Cortex Sense, Microsoft pushed Fabric IQ as a semantic ontology layer, Pinecone shipped Nexus, and Redis produced a context platform. And across all of these, it will require some specialized expertise, whether that’s each vendor’s FDEs or a company’s own local experts helping to hand-carry these new context layers into the enterprise, because the fully automated version is not quite ready.

For their part, AWS is framing this work internally as AI 45 for iteratively building on smaller successes that inform bigger successes as part of a virtuous loop at different scales such as fast ideation in 45 minutes, validation in 45 hours and sprints in 45 days. Vasquez suggests the real value is the prospect for cultural change by working through these larger and larger loops as a team:

One of the unintended benefits we also see — and it’s not the technology, it’s what happens after you do one of these sprints — our customers’ culture around thinking and operating like a frontier AI type company, even though they may be in an industry that they’ve been in for a hundred years or fifty years. It starts that cycle of cultural change, and that is something that is very hard to do in this day and age.

My take

This is all well and good, and clearly the bar for new consulting services models and supporting data models that pay a bit more attention is moving in the right direction. But long-term adoption of AI more broadly feels likely to require a bit more progress to address pushback on out-of-control tokenomics costs and the human frustration of AI being imposed on them.

Vasquez says the value of these more immersive FDE models lies in surfacing the things nobody writes down:

The FDEs from the partners like AWS have to be entrenched in the customer’s business process, and so they don’t get a chance to just observe from the outside; they have to be in the flow.