Forward Deployed Engineer Becomes a Hot AI Job as Firms Build Frontline Teams

The role has drawn major corporate investment. Anthropic said it will invest $100 million to train 10,000 FDE engineers by the end of 2027. OpenAI set up a deployment-focused company with $4 billion in investment. AWS allocated $1 billion to form an FDE division and send thousands of engineers to customers. In China, Kimi announced it would work with several IT service providers to build an FDE team, while Tencent Cloud launched what it calls the industry's first FDE engineer certification and began recruiting FDE partners. Independent developers are also taking orders to help enterprises adopt AI.

The job is not new in kind. Deployment engineers, implementation consultants, solution engineers and after-sales engineers have long done similar work, sometimes treated as dirty or laborious tasks. In the AI era, the role has gained prominence because enterprises often do not know where AI can be used, so FDEs first help customers decide what to build. In the classic FDE model, two roles often appear: Echo, who understands the business, communicates requirements and designs solutions, and Delta, who writes code, connects data and builds the solution. Backend developers build the standard platform. The difference is that a developer creates one capability for many customers, while a Delta solves many problems for one customer. Some Chinese FDE job postings now label Echo or Delta tracks, a change that was not seen two months earlier.

Much frontline work revolves around Ontology. In simple terms, Ontology is a business map that software and AI can call: what employees, departments and orders exist, how they relate, and what actions each can perform. FDEs gather information scattered across systems, processes and employee experience into this map, then build applications and solve problems on top of it. When the model moved to China, it produced variations.

01.AI, led by Kai-Fu Lee, is a typical example. The company has shifted fully to enterprise AI. Its path is to find the company's No. 1 leader, send frontline deployment engineers onsite, sort out the enterprise's Ontology, and eventually turn the work into products and a platform. In an early project, it sent five FDEs onsite plus a five-person backend team. Lee said the product is productized and subscription-based, but 'it is not a plug-and-play product.' Building a company's Ontology takes one to three months, and its database, roles and processes must be ready before the customer can get full results.

Layla works as an FDE at a SaaS company, though her title is still product manager. She began entering customer projects around March or April 2025. Her clients are listed companies with more than 10,000 employees, in chain stores and manufacturing, and she stays onsite for one or two months at a time. At first, no one called the work FDE. After the model became popular, she looked into it and found it was the same thing. Her process is to enter the site and understand the business, identify what the customer really wants from roles in procurement, marketing, finance and IT; propose a solution and build a demo for quick validation; hand validated work to the company's existing delivery team; and abstract reusable capabilities back into the company's product. She believes FDE observations and experience should ultimately settle into standard products.

Sonic, an enterprise FDE, works on audit AI projects such as finance and tax audits and advertising material review. His team began studying FDE in 2025. He says one of the most critical steps is sorting out Ontology. If an enterprise wants an AI assistant for organization management, it must first clarify how many departments it has and how they relate, as well as attributes such as a person's rank and salary. Only when these are abstracted into a structure an agent can understand will the assistant answer accurately rather than guess. The step is often difficult because FDEs encounter incomplete historical data from upstream and downstream systems and old systems without open APIs. Sonic says the step cannot be skipped: 'If the quality is high enough, it will be very effective. Subsequent delivery quality mainly depends on the quality of the Ontology.' For Layla and Sonic, being onsite is also about aligning business teams on what AI can and cannot do.

Cheng Tianshu of Silicon Valley company Baseten said its FDEs generally do not stay onsite. Baseten is an AI inference cloud company. Customers deploy models on its platform, and Baseten makes those models run faster, more stably and more cheaply. Its customers are almost all engineers with specific needs, such as switching from OpenAI's API to an open-source model, choosing which model to use, how to deploy it, and how to reach expected speed and cost. FDEs meet customers, learn technical requirements and build a prototype to validate the plan. Problems may involve cloud infrastructure, model inference optimization, training or product experience, handled by FDEs with different specialties. Most communication can be done online because Baseten offers a standardized developer product that covers one clear link in the customer's business. Customization is limited, and the product has low coupling with the customer's own business logic.

Outside companies, a group of independent FDEs is growing quickly. They do not belong to a company with a product; they find clients, take orders and help enterprises adopt AI. AI has changed the economics: custom systems once took a team months to build, but now code is nearly free and one person can deliver. Lawted, who calls himself the No. 1 FDE online, calls this group 'local FDEs.' Big-company FDEs mainly promote their own products, while local FDEs serve local bosses, can use OpenAI or Gemini, have no sales quotas, and follow the principle: 'You want something, I give you something.'

Zaniel, a Stanford PhD student, began taking enterprise AI projects alone in April. In more than five months, he has completed dozens of projects and delivered all of them by himself. His clients include banks, law firms, coal chemical companies, postpartum care centers, domestic agencies and construction companies. After doing many scenarios, his focus shifted from developing AI systems and tools for customers to teaching enterprise employees how to use AI. His first client was a bank where several employees exported data from a data system every day and compiled reports for leaders. He built an automatic generator that produced reports when data was dragged in and a button was clicked. It quickly fell out of use because report rules changed daily and exported table formats were not identical, so the hard-coded program broke within days. He later found that no new software was needed. He taught employees to give de-identified data to a general AI for aggregation, organize high-frequency processes into Skills, and connect them to existing data systems. Employees then actually used it. Zaniel says that in white-collar work, most problems can be solved by a general Agent plus Skills and connections to existing systems, without developing other new software.

Another independent FDE, Ruanmeng Zishen, described a similar situation. She worked for a foreign trade company as an FDE. After discussing with the boss how to increase employees' AI usage, the conclusion was to buy WorkBuddy. She trained employees to use WorkBuddy and connected it to the group's existing internal systems. These local FDEs found that in many cases no new product or system was needed, and their business models shifted toward consulting. Zaniel has two cooperation models with enterprises: project-based and advisory. The advisory model is becoming more common. As an advisor, he teaches employees how to use AI tools, helps enterprises think about how business should keep up when new tools appear, and builds small needs along the way as part of the advisory service. He calls himself an external CAIO, or chief AI officer. Lawted also feels his business is becoming like a consulting company. He runs an FDE community called HA7CH, organizing offline meetups and hackathons to match enterprises that want AI transformation with independent FDEs who sign orders. Enterprises submit needs, he selects one or two participants, and the participants enter the factory for 48 hours to map workflows and build a demo. If the boss is impressed, they sign a contract on the spot, and the participant handles delivery. After several rounds, Lawted increasingly sees FDE as very consulting-like. Enterprises often only know they want to use AI but cannot say where the problem is, so FDEs must enter the site, diagnose and then build a solution. He believes consultants may be better suited than programmers to move into FDE work, and only need to add vibe coding skills.

In FDE work, coding and development take little time, while communication takes most. Practitioners say communication and requirements gathering account for more than 70 percent of their time, with development about 30 percent. Friction with customer employees and difficulty obtaining API permissions often occur in this process. While connecting WorkBuddy to internal systems, Ruanmeng Zishen repeatedly talked with procurement, IT, ERP product managers and external engineers. Some systems had no ready-made APIs, so she had to push the enterprise to develop them. After more than a month, many APIs were still not open, and she said she would continue to argue with IT the next day. She does not think employees deliberately obstruct the work. Everyone has their own tasks, and AI transformation may not be a high priority for them. Failure to get permissions is sometimes a process issue. AI also brings subtler resistance: past software deployments rarely directly affected someone's position, but AI can change a team's original work. One enterprise wanted an AI system covering dozens of business modules. Zaniel estimated he could complete it in three to four months, but the client required him to work with the IT team as a consultant. Despite signing a confidentiality agreement, the client was unwilling to open its codebase, yet often asked him to troubleshoot remotely. Zaniel also suggested that some modules use a general Agent rather than full self-development. A leader replied: 'You are using WorkBuddy, then what will they do?' Zaniel does not simply attribute this to a client being uncooperative. Clients may have confidentiality requirements, existing supplier relationships or concerns from different departments. 'For an enterprise, the cost of changing a way of working is often much greater than changing a tool,' he said.

Another frequently discussed question is whether FDE is outsourcing. At Baseten, the answer is no. Cheng Tianshu said its FDEs have the same hiring standards and interview process as regular engineers, receive fixed salaries and equity, and do not rely on commissions. Their services are not charged separately. Even if a prototype is built, the customer does not pay if it ultimately does not use Baseten's cloud platform. For Baseten, FDEs serve its own product: they help customers use it and bring field feedback back to improve the product. Layla and Sonic also say that feeding experience back into products and reusing it is an important difference between FDEs and outsourcing. Layla said that in the past, outsourcing was distinguished by the fact that it did not build its own products or accumulate its own capabilities. 'If we also do not accumulate our capabilities, do not build our own products, and start every project from zero, others can say it is outsourcing and we cannot really refute it.' Sonic said traditional outsourcing cannot do FDE work. 'If you do not believe it, try outsourcing.' As FDEs, they sediment reusable modules from projects by scenario into an evaluation and iteration platform, then build a Skill Hub, according to the report.