The conversation in AI has shifted. In 2023, everyone was asking "Can we use AI?" In 2025, the question became "Why isn't our AI working?" In 2026, the question is simpler: "Who is going to build this?"
The answer, increasingly, is a Forward Deployed Engineer.
What Is a Forward Deployed Engineer?
The term comes from Palantir, which pioneered the model: instead of building a product and hoping customers figure out how to use it, you send an engineer to the customer. They work inside the client organization, own the technical integration end-to-end, and don't leave until the system is in production.
The key word is embedded. Not a consultant who writes a deck and flies home. Not a solutions engineer who does a demo and hands you a doc. An engineer who sits in your Slack, attends your standups, and writes code in your repo.
OpenAI saw the same pattern and launched the OpenAI Deployment Company in May 2026 — a standalone entity whose entire job is to embed engineers at Fortune 500 companies. Anthropic partnered with Deloitte in a $1.5 billion joint venture structured around the same model. The signal is clear.
Why the Model Works
1. The deployment gap is an engineering problem
Most AI pilots fail not because the model isn't good enough, but because nobody knows how to connect the model to the systems that need it. Data is in the wrong format. APIs are undocumented. Security reviews take months. Edge cases nobody anticipated start showing up on day two.
These are engineering problems. They require engineers — not strategists, not project managers, not another workshop.
An FDE brings the engineering capacity to solve them. They don't wait for a ticket. They find the problem, fix the problem, and ship.
2. Context is worth more than capability
The dirty secret of enterprise AI is that the model is the easy part. The hard part is understanding the customer's data, their workflows, their constraints, and their definition of "good enough."
A generic offshore team, no matter how technically capable, starts every engagement with a deficit of context. An FDE, embedded in the organization, builds that context daily. They know why the data is messy. They know which stakeholder will kill the project if the latency is above 2 seconds. They know the system that generates the training data has a bug that's been there for 18 months.
Context like that is worth more than raw engineering hours.
3. Accountability collapses distance
When something breaks in production at 2am, the question is always: who owns this? In a traditional services engagement, the answer is murky — the client team owns it but doesn't understand it; the vendor team understands it but doesn't own it.
The FDE model collapses that distance. The FDE is on your on-call rotation. They built the system. They know the runbooks. The accountability is clear.
What a Good FDE Function Looks Like
Not all FDE engagements are created equal. Here's what separates the ones that ship from the ones that stall.
Senior engineers only. FDE work requires judgment, not just execution. You're often making architectural decisions under uncertainty, negotiating with skeptical stakeholders, and working without complete information. Junior engineers can't do this independently.
Fixed scope, not time-and-materials. Open-ended engagements drift. The best FDE engagements start with a scoping exercise that produces a concrete definition of done: what system, what performance, what deadline. Scope can change — but changes should be explicit, not drift.
Eval-first. Every system an FDE builds should ship with an automated evaluation suite. If you can't measure whether the system is working, you can't trust it in production, and you can't detect when it stops working. This is non-negotiable.
Knowledge transfer as a deliverable. The goal of an FDE engagement is to make the FDE unnecessary. Runbooks, architecture docs, training sessions, annotated code — all of it should be included in the engagement scope. FDEs who create dependency aren't doing their job.
The Market in 2026
FDE job postings grew 800% between January and September 2025. Mid-level FDEs are clearing $300K–$450K total comp. OpenAI has 30+ open FDE roles. Palantir has 50+. There are 224 open FDE roles across 39 AI companies as of this writing.
The supply of qualified FDEs is nowhere near the demand. Which is why the agency model exists: you get access to an FDE without competing with OpenAI's recruiting machine or paying for a full-time hire before you know if the model fits.
When to Hire an FDE vs. Build In-House
The agency model is not always the right answer. Here's the honest decision tree.
Hire an FDE agency when:
- You have a defined AI system to ship and a deadline that matters
- Your internal team has strong product and domain knowledge but limited AI production experience
- You want the system delivered and handed off — not an ongoing dependency
- Your AI project is scoped (a specific pipeline, agent, or eval system) rather than open-ended
Build in-house when:
- AI is a core, permanent differentiator in your product — meaning the work never ends and the institutional knowledge is your moat
- You have 12+ months of runway to hire and ramp a senior AI engineer
- Your AI requirements are deeply entangled with proprietary systems that require months of context to understand
Most companies that come to us fall into the first category: they know what they need to build, they don't have the engineering capacity to build it, and they can't wait 9 months for a hiring process to complete. The engagement model — a fixed scope, a fixed timeline, full ownership, then handoff — fits that situation exactly.
The FDE Engagement Lifecycle
A well-run FDE engagement has three phases, each with clear deliverables.
Phase 1: Scoping (1–2 weeks). The FDE runs a rapid technical discovery. They map your existing systems, understand your data, identify integration dependencies, and define the technical success criteria for the engagement. The output is a concrete spec: what system will be built, what performance it needs to hit, what the handoff looks like.
Phase 2: Delivery (6–20 weeks depending on scope). The FDE embeds in your team. They own the engineering delivery — architecture decisions, code, integration work, eval framework design — and surface blockers proactively. Weekly reviews keep stakeholders aligned without creating overhead.
Phase 3: Handoff (1–2 weeks). The FDE writes runbooks and architecture documentation, trains your internal team, and sets up monitoring so your team can detect regressions. The engagement doesn't end until your team can own the system independently. That's the definition of a successful FDE engagement.
This lifecycle directly addresses the most common AI pilot failure mode: the handoff from "it works" to "someone owns it" that most project teams never complete.
What to Look For in an FDE Partner
Not every AI engineering firm that calls itself an FDE agency operates this way. Here's what separates serious FDE engagements from dressed-up staff augmentation.
Defined exit criteria. A legitimate FDE engagement specifies what "done" looks like before work begins. If a firm is happy to run time-and-materials until you decide to stop, they're not doing FDE work.
Eval-first delivery. Every production AI system needs automated evaluation. If the firm doesn't mention evals in the first conversation, they've never shipped a system that needs to be trusted long-term.
Practitioner-level depth. FDE work requires engineers who can make architectural decisions under uncertainty — not junior engineers following playbooks. Ask about the specific systems they've shipped, the failure modes they've encountered, and how they've handled stakeholder resistance to AI adoption.
Realistic timelines. Enterprise AI deployment is hard. Security reviews take months. Legacy integrations take longer than the AI work. A firm that promises a production RAG system in two weeks without understanding your existing systems is telling you what you want to hear.
Common FDE Engagement Mistakes
The engagement model is powerful, but companies frequently undermine it.
Treating the FDE like a vendor. The model only works if the FDE has real access: to your Slack, your code, your stakeholders, your data. Companies that try to mediate all communication or restrict system access get slower delivery and weaker outcomes. Embedded means embedded.
Changing scope mid-engagement without explicit renegotiation. Scope creep is the most common cause of FDE engagement failure. If a stakeholder adds requirements in week six, that needs to be treated as a new scope decision — not absorbed silently. Good FDE engagements have a clear change control process.
Skipping the handoff phase. The last two weeks of an FDE engagement often feel unnecessary when you're staring at a working system. They're not. Runbooks, architecture docs, and training sessions are what transform an FDE delivery from "we own a system we don't understand" into a system your team can evolve confidently.
Not involving end users early. The FDE builds what they're told to build. If the requirements don't include the workflows and constraints that end users actually have, the system works technically and gets low adoption. Involve end users in the scoping phase, not just the launch demo.
Frequently Asked Questions
How is an FDE engagement different from hiring a contractor? A contractor takes requirements and executes them. An FDE takes ownership of an outcome — they scope the problem, make architectural decisions, build the system, and own the handoff. The accountability model is fundamentally different.
What does it cost to run an FDE engagement? Typical engagements run $50K–$500K depending on duration, team size, and complexity. An 8-week solo FDE engagement for a well-scoped RAG pipeline is typically in the $80K–$130K range. Compare that to $400K–$600K for a full-time senior AI engineer hire, and the economics of the agency model become clear.
Can an FDE engagement work remotely? Yes, with important caveats. Remote FDE engagements require real-time Slack access, daily async standups, and weekly video check-ins at minimum. The "embedded" part of FDE isn't primarily about physical presence — it's about being deeply integrated into the team's communication and decision-making.
What systems can an FDE build? The most common engagements are RAG pipelines, AI agent systems, LLM evaluation frameworks, and on-prem/sovereign AI deployments. Less common but well within scope: fine-tuning pipelines, ML infrastructure, embedding systems, and production monitoring for AI systems.
How do I know if my project is right for an FDE engagement? If you can describe a specific AI system you need in production within a defined timeframe, you're a candidate. If you're looking for open-ended AI strategy consulting, you need a different kind of engagement.
For most companies, an FDE engagement is the fastest path from "we have AI ambitions" to "we have an AI system in production." If you're trying to close that gap, talk to us. We scope every engagement in 2 days.