A Forward Deployed Engineer (FDE) is the software engineer who works inside the client's operation, not at a distance. They build, integrate and put the solution into production in the environment where it will run, alongside the team that will use it, and they measure their own work by the client's result, not by the delivery of the code.
The role is no longer a niche. According to Indeed data published by Business Insider, job postings for the role grew about 729% between April 2025 and April 2026, and Anthropic, OpenAI, Palantir and Stripe are among the companies hiring. Our reading of why: AI models became good enough, and the bottleneck moved to putting them to work inside an actual company, with its own data, systems and exceptions.
Where the model comes from
The term comes from Palantir. In a 2019 post on the company's blog, it describes its two largest engineering roles. The product developer builds the platform, which will serve many clients. The Forward Deployed Software Engineer, known internally as Delta, is part of a team dedicated to a single client: they deploy and adapt the platform to solve a critical problem for that business, and they measure success by the impact on the client's goal.
Palantir itself sums up the difference in two phrases: the developer thinks in terms of "one capability, many customers"; the Delta, in terms of "one customer, many capabilities". That second phrase explains why the model came back so strongly with generative AI. The language model is general. What decides whether it works in a company is specific to that company: where the data is, which systems need to be read and written, which exceptions only the people who operate it know, where the decision needs a person. That ground cannot be covered from a distance.
What a Forward Deployed engineer does
- Gets into the process before the code. Shadows the people who do the work today, maps the exceptions and finds out where the data really is, which is rarely where the org chart says.
- Integrates with the real systems. ERP, CRM, service channels, databases and internal APIs, reading and writing where the process requires it, with the minimum permission needed.
- Builds on a platform, not from scratch. Reuses what is common to every agent (execution, observability, governance) and writes only what is specific to that client.
- Takes it to production alongside the team that will operate it. The people who will work with the agent learn during go-live, not in a handoff at the end of the project.
- Measures by the business number. The definition of done is the process indicator moving, not the scope being delivered.
- Stays after go-live. Tunes the behavior with real use, tracks quality and cost, and takes the next case to production.
Forward Deployed is not consulting, a fixed-scope project or support
The three best-known models of outside help end before the point where AI usually fails. The difference lies in where responsibility for the result sits.
| Model | What it delivers | Where it ends |
|---|---|---|
| Traditional consulting | Assessment, recommendation and roadmap. | At the report. Another team implements. |
| Fixed-scope software project | Code that follows a specification. | When the scope is delivered. Any change becomes a contract amendment. |
| Vendor support | Help using the product as it is. | When the ticket is resolved. |
| Forward Deployed | A solution running in production, measured by the result. | It does not end at go-live: it continues with the operation. |
That does not make the other models wrong. A well-built roadmap has value, and there are cases where a fixed scope is exactly right. The point is that an AI agent in production almost never fits in a specification written beforehand: the right behavior emerges in contact with real data and with the people who operate, and someone needs to be there when that happens.
When the model makes sense, and when it does not
It makes sense when
- the process depends on data spread across more than one system;
- there are many exceptions, and only the people who operate know them;
- the case requires governance from the start: human approval, an audit trail, the LGPD (Brazil's data protection law);
- the internal team knows the business but has not yet put an agent into production;
- the goal is an operation running, not a demo for leadership.
It does not make sense when
- an off-the-shelf tool solves it, like a writing assistant or a meeting summarizer;
- the company already has a mature AI engineering team and only needs infrastructure;
- the case has no owner in the business. Without an owner, no deployment model saves the project.
What to ask anyone offering Forward Deployed
The name became fashionable, and fashion attracts new labels for old services. Six questions separate the genuine model from consulting under another name:
- Who on your team will work inside our operation, and until when?
- What stays with our company at the end: code, configuration, documentation, access?
- Which business number will measure success, and who tracks that number after go-live?
- Which platform do you build on, and what happens if we want to switch AI models?
- How do human approval, the audit trail and LGPD compliance work?
- Who on our team will operate the agent, and when are those people trained?
If the answer to the first question is "nobody, but we have a great platform", what is being offered is a license. It may be the right choice, but it is not Forward Deployed.
How Runflow applies the model
At Runflow, Forward Deployed work has two roles. The AI Deployment Strategist maps the processes, prioritizes the cases by business value and tracks the return of each agent. The AI Solution Engineer builds the agents on the Runflow platform, integrates them with ERP, CRM, internal APIs and databases, configures governance and observability and puts the operation into production. A Client Manager looks after the account from the first contact onward.
The platform solves what is common to every agent, which is to build, operate and measure it. The engineers solve what is unique to your operation. The first agent goes into production in weeks, not months, and the team stays on after go-live. See how the team works, or start with a free assessment of the process you have today.
