APRIL 10, 2026 · BRYAN LEE

Unlocking the AI Overhang

The case for forward deployment

Most large companies are sitting on exactly what AI systems need to become useful: proprietary datasets, institutional knowledge, and decision-making artifacts that have been refined over years. The right models and agents can turn this institutional knowledge into a compounding system that automates high-value workflows, outperforms general-purpose AI tools on company-specific tasks, and gets better with every decision.

At Applied Compute, we work with large enterprises to build and deploy agents inside production environments. That means sitting in customers’ offices, poring over unstructured data, and translating research into real production deployments.

What we’ve seen is that there's an immense overhang in enterprise AI capabilities.

A compelling demo and a reliable production system are very different things, and that gap is where most enterprise AI initiatives stall. At Applied Compute, we’ve trained models which outperform frontier systems on company specific tasks and economically valuable domains. We’ve also developed an agent platform that encodes institutional knowledge to replicate expert level judgement across many tasks in an enterprise. The signal is already inside the organizations: the hard part is finding and structuring it, then closing the loop to learn from it continuously.

The overhang is the delta between a model's capability and its utility in a specific workflow, and forward deployment plays a critical role in bridging that gap.

Bridging the gap: forward deployment

The forward deployed model Palantir pioneered a decade ago has substantially evolved in the AI era.

Today, enterprise customers want agents that automate entire workflows, which requires us to operate end-to-end across agent training, evaluation, deployment, and production support. In high-value processes, domain experts know what "good" looks like but haven't formalized it into a reward signal. Any existing data often captures outcomes but not the reasoning behind them. The margin in enterprise software has shifted from storing data to doing useful work on top of it. In practice, deployments are often orchestrations of data pipelines, multiple agents, humans in the loop, and purpose-built models that collectively outperform what any off-the-shelf LLM or even a single human could produce.

For our customers, our goal is to continually improve their business processes through workflow-specific agents. This is why we think of the forward deployed role a bit differently. At Applied Compute, we have two forward deployed roles: Forward Deployed Engineers (FDEs) and Applied Research Engineers (AREs).

• FDEs: Over the last few years, the unit of value for customers has moved from lines of code to teams of agents and the outcomes they drive. At Applied Compute, we’ve shaped the FDE role to reflect this. FDEs bring agents to production by constructing robust evaluation frameworks, creating rich environments in which agents operate, developing large scale context ingestion engines across a customer’s data corpus, and deploying our core platform which enables us to expand into large enterprise verticals.
• AREs: For many agentic workflows, customers also care about each incremental performance gain. AREs combine production-grade software skills with research ability. They work with FDEs to convert institutional knowledge into agent context, fine-tune models on proprietary data, optimize agent performance, and deploy systems which continuously improve against real-world outcomes.

In both roles, there are more similarities than differences: the core motion starts with understanding a customer's reality deeply enough to build systems that actually work inside it, and surfacing valuable feature sets from each engagement and integrating them back to the core platform.

Our mantras: what we look for in FDEs & AREs

While the FDEs and AREs have different areas of expertise, the two roles have high overlap in fundamental skillsets, potential customer impact, and core intuitions.

Through working with large enterprises and expanding our forward deployed team, we've identified skills that enable FDEs and AREs to thrive in their roles:

They have engineering foundations and a research mentality:

• FDEs and AREs have strong foundations as engineers focusing on developing production ready systems. They value strong coding ability and the “shake your computer until you understand” mentality, but they don’t over-index on perfection.
• At the same time, they run agent ablations and repeatedly iterate on robust, task-specific evaluations which require a principled research mentality and creative ingenuity to think of the best solution for a problem.
• They embody what we call “full stack AI engineering” -- bringing a model or agent from training, to evaluation, to production deployment.

They bring the frontier to production:

• In a given week for an engagement, AREs might fine-tune a model on enterprise data no frontier lab has ever seen, while FDEs deploy a large scale context ingestion engine for a customer to power a new agentic system, and both come together to present results to the customer's leadership at end of week.
• Both FDEs and AREs often apply techniques from research papers (from training algorithms to harness engineering) to customer data to deliver meaningful bottom-line impact.

They have high customer empathy:

• FDEs and AREs are not just "at the desk" roles. They involve sitting with customers, interviewing subject matter experts, collaborating with internal teams, and digging into the implementations of customer systems.
• They are also credible counterparts to customer stakeholders, able to negotiate technical scope, communicate project direction, and build rapport.
• They are as effective at interacting with customers as they are coding in a terminal or Python notebook.

They’re excited by large scale, unstructured problems:

• Enterprise AI engagements rarely come with clean specifications.
• In our experience, the best FDEs and AREs don't wait for clarity, they dig through the messiness to create it, but are also conscious of thorough scoping to focus on the highest order bits to deliver value the fastest.
• They thrive in the ambiguity of a new deployment, turning vague goals into concrete deliverables.

They demonstrate end-to-end ownership:

• FDEs and AREs own entire customer engagements, not individual tasks.
• This requires a strong bias toward action. We’ve found a good heuristic for strong members in this role is their first reaction when they feel blocked: they have an “internal locus of control” where they figure out something they can improve and creative avenues to pursue to expedite the process or set themselves up well for the future.

The road ahead: what we’re building towards

We’re in the early innings of a generational technology shift that can become a compounding advantage for companies. Our vision is a future where every enterprise owns its Specific Intelligence: agent workforces which continually learn from the enterprise’s unique datasets and business operations. Unlocking these capabilities is enabled by FDEs and AREs that have strong engineering ability, research rigor, and customer empathy.

The need for this work will only grow as AI becomes more deeply embedded across organizations. If this excites you, join us.