knowledge machine

knowledge > tech

Why FDE? Why Now?

In an era when AI writes code, why are AI companies investing in sending engineers to customers? Fifteen questions explore how FDE differs from systems integration and how a business built on embedded engineers can turn a profit and scale.

11/13/2026#forward-deployed-engineering#enterprise-ai#ai-agents#product-development

Why is there so much discussion about Forward Deployed Engineers right now? What role do they play in AX (AI transformation) projects? I looked for answers to the questions I had.

  1. What is the history of FDE, and how did the role emerge?
  2. What is an FDE, and how does the role differ from SI, SA, and other embedded development roles?
  3. What skills does an FDE need?
  4. Which customers should an FDE project target, and what should it promise?
  5. What is the deployment gap in enterprise AX, and how can FDEs close it?
  6. Why should enterprise AX teams avoid implementing customer requirements at face value?
  7. FDE is an established role. Why is it attracting attention and investment now?
  8. Is the enterprise AX market involving FDEs large and profitable enough?
  9. What do big tech and frontier AI companies gain by acquiring FDE businesses?
  10. FDE deployment adds to AX project costs. How do companies employing FDEs handle this expense?
  11. In what sequence do FDE projects proceed, and what matters most at each stage?
  12. How are FDE-built systems operated after the contract or development period ends?
  13. Which products and customers suit FDE? Do companies with development teams still need it?
  14. Do FDEs build on existing products or develop separately and integrate later? How do they choose?
  15. FDE work seems difficult to scale. What scaling and platform approaches have actually been tried?

Q1. What is the history of FDE, and how did the role emerge?

The current FDE model developed as Palantir built software to support intelligence work. Founded in 2003, Palantir initially developed software for US intelligence agencies’ counterterrorism investigations and operations, then expanded into commercial markets.[1]1Palantir Technologies Inc., 2020 IPO prospectus, 2020, “Business—Overview.” Describes the company’s founding in 2003 and its initial work for US intelligence agencies. ↩ www.sec.gov

Bob McGrew, who led Palantir’s early product and engineering work, recalled how FDE emerged. In the mid-2000s, the development team did not fully understand how intelligence agencies actually worked. Because outside developers struggled to learn their customers’ workflows, the team first built prototypes and showed them to agency staff. Staff explained where the prototypes failed to fit their work and what capabilities were missing. The developers used that feedback to understand the workflows and revise the product.

Working with customers exposed another problem: a product that helped one customer did not necessarily work for the next. Workflows and required capabilities differed. Delivering an identical finished product to every customer could not accommodate those differences. Palantir developed an approach that put shared capabilities into a platform and implemented customer-specific requirements in the field. Engineers worked directly with customers and built capabilities the platform lacked. The product team identified field-developed capabilities that other customers also needed and incorporated them into the shared platform. Understanding customers’ work and directly building the software they needed became the FDE’s role.[2]2Bob McGrew, Y Combinator, The FDE Playbook for AI Startups with Bob McGrew, 2025-09-08, 03:19–07:55. Checked against the publicly available transcript. ↩ www.youtube.com

Within this structure, engineers developing the shared platform were called Devs, while those solving customer problems were called Deltas. Palantir’s official 2019 explanation distinguished their focus as “one capability for many customers” and “many capabilities to solve one customer’s problem.”[3]3Palantir, Dev versus Delta: Demystifying engineering roles at Palantir, 2019-04-08, “Wait, what’s a Delta?” ↩ blog.palantir.com For the same customer problem, a Dev improved a capability that many customers could use, while a Delta combined that capability with other tools to improve the customer’s actual work.

Other companies created similar teams to support complex customer workflows. Ramp, which helps companies manage spending, says it began forming its FDE team in fall 2023 to address larger customers’ complex requirements.[4]4Leo Mehr, Ramp Builders Blog, Forward Deployed Engineering, 2025-08-05, “FDE at Ramp.” ↩ engineering.ramp.com This approach predates the widespread adoption of generative AI. More recently, it has expanded into applying AI to customers’ day-to-day work.

Q2. What is an FDE, and how does the role differ from SI, SA, and other embedded development roles?

An FDE works with a customer to define business goals, directly implements the software needed to achieve them, and carries it through to production operations. The role combines defining the customer’s problem, writing code, and evaluating business outcomes.

First, FDEs continue to own implementation, deployment, and validation of business code after diagnosis and design. Their role does not end with recommending an architecture and handing development to another team. They connect customer data and existing systems, implement business rules, and fix problems found in use. Palantir’s FDEs also contribute necessary capabilities and bug fixes directly to the shared product, working with the product team on design and code review.[5]5Palantir, Dev versus Delta: Demystifying engineering roles at Palantir, 2019-04-08, product-code contributions in “Wait, what’s a Delta?” and “What does the day-to-day of your role look like?—Delta.” ↩ blog.palantir.com When a customer needs a capability the product lacks, the FDE participates in implementation rather than stopping after submitting a request to the product team.

Second, FDEs agree on business goals with customers and assess whether those goals have been achieved. Palantir gives the example of reducing a manufacturer’s defect rate. The FDE selects the data connections and capabilities needed to reach that goal, and performance is judged by improvement in the customer’s metrics.[6]6Palantir, Dev versus Delta: Demystifying engineering roles at Palantir, 2019-04-08, “Wait, what’s a Delta?” Manufacturing defect-rate example and performance evaluation against customer goals. ↩ blog.palantir.com Setting the goal first changes decisions about what to build. After delivering a feature, the FDE compares business metrics before and after adoption, identifies why results fall short, and makes changes.

Suppose a team builds a returns agent to reduce employees’ manual work. If the agent answers inquiries but employees still have to look up orders and register returns manually, the workload may barely change. The FDE identifies why people still need to intervene, adjusts the order-system integration or approval process, and measures the change in processing time. In this example, deployment begins the process of validating results, and subsequent improvements remain part of the project’s responsibility.

Compared with an SA (solutions architect) whose remit ends at design advice, or an SI (systems integration) engagement or contracted developer responsible for implementing a fixed specification, an FDE spans problem definition through operational results. SI teams and SAs that also own implementation and business outcomes overlap with FDE. Combining those two responsibilities characterizes the FDE approach discussed here.

Q3. What skills does an FDE need?

An FDE must break customer problems into development tasks, implement the necessary systems, and agree on scope with the customer. AI projects also require the ability to evaluate agent failures and adjust operations.

Engineering is the foundation. FDEs must connect data and external systems, implement business rules, and diagnose and fix production errors. Palantir describes the ability to understand a high-level business problem down to the code needed to solve it. It also emphasizes independently learning unfamiliar languages and tools, building plugins, and debugging platform problems.[7]7Palantir, Dev versus Delta: Demystifying engineering roles at Palantir, 2019-04-08, “What are some examples of skills that are important for your role?—Delta,” on problem decomposition, independent learning, and technical judgment. ↩ blog.palantir.com

FDEs must also learn the customer’s work and make problems concrete. “Returns are slow” calls for investigating order lookup, policy decisions, and approval queues separately. Findings must be explained as workflow changes to business users and implementation tasks to developers. Palantir identifies user empathy, an interest in business strategy, and independent judgment as attributes of successful FDEs.[8]8Palantir, Dev versus Delta: Demystifying engineering roles at Palantir, 2019-04-08, “What are some examples of skills that are important for your role?—Delta,” on user empathy, business strategy, and independent judgment. The returns analysis in the body is illustrative. ↩ blog.palantir.com This means providing reasons for which problems to solve first and building agreement with the customer.

Product judgment is needed to choose a useful scope within limited time and decide how much customer-specific work to allow. Ramp combines questioning requirements to reduce scope with building capabilities reusable across customers.[9]9Leo Mehr, Ramp Builders Blog, Forward Deployed Engineering, 2025-08-05, “FDE Principles,” including “Always be scoping” and the trade-off between reusable capabilities and minimum implementation scope. The return-window configuration in the body is illustrative. ↩ engineering.ramp.com For example, a team might start with common, routine returns while exposing different customers’ return windows as settings. The decision must account for both the immediate deadline and reuse for the next customer.

AI projects add the ability to evaluate model responses and actions. Sierra describes preserving errors from actual conversations as tests and checking whether they recur after changes.[10]10Clay Bavor·Bret Taylor, Sierra, The Agent Development Life Cycle, 2024-06-03, “Quality assurance: Continuous, structured human feedback” and “Testing: Regression tests for conversations.” The returns-agent test cases and failure diagnosis in the body are illustrative. ↩ sierra.ai A returns agent needs tests for routine requests, expired return windows, failed order lookups, and cases requiring employee approval. The FDE must distinguish missing order data, incorrectly applied policies, and judgments the model struggles with, then adjust integrations, rules, and human involvement accordingly.

Q4. Which customers should an FDE project target, and what should it promise?

Selling an FDE project starts with assessing both the value of the problem and the customer’s ability to execute. Taken together, the sources suggest an ideal customer profile (ICP): a company whose potential improvement is large enough to cover implementation and operating costs, that has a budget for the work, and whose solution can use a shared product. Field development costs are difficult to recover when the benefit is small or every customer requires a completely different system. Even a large contract can consume much of its revenue in customer-specific development and support if dedicated staff must continually build and maintain new capabilities.[11]11Marc Andrusko, Andreessen Horowitz, The Palantirization of everything, 2026-01-16, “Not every problem is a ‘Palantir-grade’ problem” and “A More Realistic Framework: When Does ‘Palantirization’ Make Sense?” The customer-selection criteria in the body synthesize its discussion of problem value, customer-level economics, and shared workflows. ↩ a16z.com

Business participation and internal approval matter alongside budget. Bob McGrew, who led Palantir’s early product development, emphasizes executive sponsorship because it enables approval for data access and workflow changes.[12]12Bob McGrew, Y Combinator, The FDE Playbook for AI Startups with Bob McGrew, 2025-09-08, later part of “Why AI Companies Adopt It,” on organizational approval and executive sponsorship. Checked against the publicly available transcript. ↩ www.youtube.com A business owner must define the goal, and IT and security staff must participate in data access and system integration. Companies that cannot define a problem, or will not provide data and participate in reviewing the work, should be excluded from the target customer pool for an FDE implementation offering.

When those conditions are met, sales staff and FDEs jointly examine the current workflow’s costs, delays, and errors, then agree on metrics and goals. The FDE examines data and system constraints to distinguish what the existing product supports from what requires additional development. Ramp involves FDEs from sales through adoption and operations to refine requirements and scope early and carry customer commitments through to implementation.[13]13Leo Mehr, Ramp Builders Blog, Forward Deployed Engineering, 2025-08-05, “Team operating model: map to the customer lifecycle” and “FDE Principles—Always be scoping.” ↩ engineering.ramp.com

A proposal should specify the initial scope, timeline, success criteria, and required customer participation. The commitment covers implementing the agreed workflow, evaluating whether it meets the goal, and improving it. It should not guarantee unvalidated results or unlimited implementation of every request.

Pricing can reflect the customer’s expected savings or additional revenue as well as the effort involved. McGrew describes pricing against the value of the problem and expanding contracts as the provider takes on more important work.[14]14Bob McGrew, Y Combinator, The FDE Playbook for AI Startups with Bob McGrew, 2025-09-08, later part of “Why AI Companies Adopt It,” on problem value, pricing, and risk, and “What Success Metrics Look Like,” from 36:17. Checked against the publicly available transcript. ↩ www.youtube.com Instead of bundling every requirement into a large initial contract, the parties negotiate the next scope and price based on early validation. Pricing based on customer value and billing based on actual outcomes are different models. The former uses expected savings or additional revenue in price negotiations; the latter calculates invoices using contractually agreed performance metrics and formulas.

Q5. What is the deployment gap in enterprise AX, and how can FDEs close it?

Colin Jarvis, OpenAI’s Head of Forward Deployed Engineering, describes the “deployment gap” as the difference between what a model appears capable of doing and what people can reliably use in actual work.[15]15Colin Jarvis, OpenAI, What FDE means at OpenAI, 2026-10-22, introduction defining “the deployment gap” between model capabilities and usable systems. ↩ openai.com Closing it in enterprise AX involves connecting models to customer data, tools, permissions, and workflows so they can be used consistently in daily operations.[16]16OpenAI, OpenAI launches the OpenAI Deployment Company to help businesses build around intelligence, 2026-05-11, “Building for where frontier AI is headed.” Describes connecting models to customer data, tools, controls, and workflows for reliable daily use. ↩ openai.com

For example, a returns agent that understands a request but fails to check the purchase date, or approves an order outside the return window, is difficult to trust with real work. Sierra, which builds customer-service AI agents and an operating platform, describes giving agents both business goals and mandatory rules: for example, “help customers with returns” alongside “returns are only allowed within 30 days of purchase.” When an error appears in a real customer conversation, the conversation becomes a test used to check whether the problem recurs after changes.[17]17Clay Bavor·Bret Taylor, Sierra, The Agent Development Life Cycle, 2024-06-03, introduction; “Development: Declarative goals and guardrails”; “Quality assurance: Continuous, structured human feedback”; and “Testing: Regression tests for conversations.” The 30-day return condition is an example from the source. ↩ sierra.ai

FDEs connect these capabilities to customer systems and workflows, then directly fix issues revealed during pilot operations. PagerDuty, a software company supporting incident detection and response, spells this work out in its job description. Its FDEs connect incident-response agents to customers’ monitoring tools and deployment systems. During pilots, they determine whether agents have the necessary access, can retrieve logs and metrics showing system conditions, or lack integrations with external tools, then write code to resolve the problems. Integration tools and capabilities needed by multiple customers enter the shared product after product-team review and evaluation.[18]18PagerDuty, Forward Deployed Engineer (SDEV) - London, England, United Kingdom, Undated; accessed 2026-11-14. Company and role introduction and “Key Responsibilities.” Describes the advertised responsibilities, not outcomes from a particular customer project. ↩ careers.pagerduty.com

Continuing the returns example as a hypothetical implementation: if the goal is to reduce manual work, the FDE follows the process from return guidance through order lookup and registration to identify why employees keep intervening. If order lookup is blocked, the FDE provides the required permissions and integration. If return policy is being applied incorrectly, the FDE changes the information and rules used for decisions. Afterward, the team measures whether fewer cases need rework and whether processing takes less time.

Problems that additional development cannot solve require operational changes. If a model still cannot reliably judge exceptional return reasons despite having sufficient order information, the FDE implements escalation to employees and narrows the automation scope. Workflows where automation increases the review burden are excluded. These decisions guide changes to implementation scope and operating procedures so the customer can achieve the intended results.

Q6. Why should enterprise AX teams avoid implementing customer requirements at face value?

Customers understand the difficulties they experience at work, but identifying the exact software that will solve them is harder. Requirements mix real problems with the customer’s proposed solution. A request to “build this dashboard,” for example, describes a screen but may omit who will use it, what decisions they will make, and what action follows. A development team can satisfy the screen specification and still leave the underlying business problem unresolved.

Vinoo Ganesh, who created a program at Palantir for engineers to participate in field work, describes a logistics customer with 47 pages of dashboard requirements. An engineer visited the site and asked what the employee would do after reviewing that information on Monday morning. The actual workflow emerged: identify delayed trucks and contact dispatchers to arrange follow-up action. A Slack alert about delays was enough to trigger that action. Ganesh recalls that the team implemented the alert in four hours and resolved the immediate dispatch problem. They examined the assumption that a dashboard was necessary and scoped the implementation around what people needed to do after receiving the information.[19]19Vinoo Ganesh, AI Engineer, How Forward Deployed Engineering is done at Kepler, AI Engineer World’s Fair 2026, “Find the decision behind the dashboard,” 05:00–08:27. The requirements length and implementation time are the speaker’s recollections. ↩ ai.engineer

The same reasoning applies to deciding what to automate first. Sunny Rekhi, FDE CTO at customer-service AI agent company Decagon, explains that the customer’s initial request may differ from the workflow offering the greatest benefit. Decagon analyzes historical support conversations and draws on experience deploying agents across companies to recommend initial automation targets. Alongside the requests customers consider important, it examines which tasks recur in actual conversations and where automation can have the greatest effect.[20]20Sunny Rekhi, AI Engineer, How Forward Deployed Engineering is done at Decagon, AI Engineer World’s Fair 2026, “Use support history to choose what to automate,” 13:44–15:02. ↩ github-main-preview.ai.engineer

When FDEs receive requirements, they first follow the workflow that prompted the request. They examine what information users gather, where they wait, and whom they ask to do what after receiving a result. Investigating undocumented exceptions and manual work can uncover additional requirements or show that some requested features are unnecessary. FDEs share these findings with customers and agree again on the problem and implementation scope. After implementation, they measure changes in processing time and manual work to go beyond delivering requested features.

Q7. FDE is an established role. Why is it attracting attention and investment now?

As AI agents become capable of taking on parts of human work, companies are attempting automation across a broader range of tasks. Buying a model does not automate a workflow. Someone must define the work, connect company data and systems, and develop and test the implementation. Demand is growing for FDEs who do this together with customers.

Bob McGrew, who led Palantir’s early product development, explains that AI agent businesses often need to discover with customers what problems their products should solve. When replacing existing software, its capabilities and usage provide a starting point. When handing previously human work to AI for the first time, teams must first decide what to automate and where people should intervene. FDEs make the product’s capabilities and usage concrete by examining customer workflows and implementing solutions directly.[21]21Bob McGrew, Y Combinator, The FDE Playbook for AI Startups with Bob McGrew, 2025-09-08, “Why AI Companies Adopt It,” from 23:04. Checked against the publicly available transcript. ↩ www.youtube.com

Consider returns: even if AI writes return instructions, completing the workflow requires looking up the order, checking eligibility, and recording the result. Exceptional requests may need to reach an employee. FDEs review this process with the people doing the work and develop the necessary capabilities. If employees still have to check and enter the same information afterward, identifying and fixing the cause is necessary to reduce the workload.

AI’s ability to lower customer-specific development costs also supports the investment case for FDE. Traditional SaaS recovers the development cost of a shared product across many customers. Its economic advantage comes from not rebuilding the same capabilities as the customer base grows. A capability built for one customer cannot easily spread its development cost across others, and it also creates separate modification and maintenance work. Product companies have therefore limited customer-specific development.

AI helps reduce the time and cost of this custom development and integration work. Joe Schmidt of venture capital firm a16z identifies a new possibility in the AI era: using AI to make the implementation work of applying AI to customer workflows more efficient and automated. Examples include mapping data fields and moving data between different systems.[22]22Joe Schmidt, Andreessen Horowitz, Trading Margin for Moat: Why the Forward Deployed Engineer Is the Hottest Job in Startups, 2025-06-04, “The implementation necessity.” Presents the investment thesis that AI can make AI implementation work more efficient and automated. The body explains the economic implications of lower customer-specific development costs. ↩ a16z.com If customer-specific implementation requires fewer people, the same contract can generate more profit, and solutions become viable for customers previously priced out by development costs. Growing demand for AI adoption, together with expectations of lower implementation costs, supports investment in the FDE approach.

Some companies see a shortage of implementation staff as a constraint on adoption of their technology. In June 2026, AWS announced a $1 billion investment in a dedicated FDE organization and support for partners building customer systems, explaining that demand for operational AI systems exceeded what its own organization could handle.[23]23Chandra Pinapala·Erik Farr·Karthik Sonti, AWS Partner Network Blog, Introducing Forward Deployed Engineering for Partners: Winning the Future of Enterprise AI, 2026-06-30, introduction and “Why Partners, Why Now.” ↩ aws.amazon.com As more customers seek to use AI, AWS aims to expand the workforce that understands their work and builds their systems.

In October 2026, Anthropic announced a $100 million investment in Claude Frontier Academy to train engineers at customers and partners, targeting 10,000 people by the end of 2027. It calls the role Frontier Deployed Engineer in this program and states that its aim is to address the shortage of skilled people needed for enterprise AI adoption.[24]24Anthropic, Anthropic invests $100 million to train 10,000 engineers and tackle the enterprise AI talent gap, 2026-10-02, introduction and “The first program trains 10,000 Frontier Deployed Engineers.” ↩ www.anthropic.com

Q8. Is the enterprise AX market involving FDEs large and profitable enough?

Enterprise spending on AI and demand for implementation are substantial, and some companies achieve high profitability while using FDEs.

Menlo Ventures estimated enterprise AI spending at an annualized $177 billion as of September 2026.[25]25Tim Tully·Joff Redfern·Deedy Das·Derek Xiao, Menlo Ventures, 2026: The State of Enterprise AI, 2026, “Enterprise AI Reaches Scale” and “Methodology”; accessed 2026-11-14. Market size is an annualized revenue estimate combining surveys, disclosures, and external sources. Historical figures were restated to match the classification framework. ↩ menlovc.com This extrapolates spending at the time of the survey to a full year and includes AI applications, models, and infrastructure. It shows the size of enterprise AI budgets. Accenture provides an example of actual service revenue: it reported $2.7 billion in generative and agentic AI revenue and $5.9 billion in generative AI bookings for fiscal 2025.[26]26Accenture, Reinventing what’s possible: FY2025 Annual Report, 2025, p. 5, “AI Leadership—Yesterday, Today and Tomorrow.” Revenue and bookings cover advanced AI, excluding data, traditional AI, and AI used in service delivery. ↩ investor.accenture.com This is enterprise AI demand turning into contracts and revenue. Palantir provides an established example of profitability: its company-wide gross margin was 82% and operating margin approximately 32% in 2025. Those figures cover the whole company, including products, subscriptions, and services.[27]27Palantir Technologies Inc., 2025 Form 10-K, Filed in 2026 for fiscal 2025, Item 7, “Results of Operations” and “Cost of Revenue and Gross Profit.” Gross and operating margins are company-wide GAAP figures using rounded percentages from the filing. ↩ www.sec.gov

These sources support a substantial business opportunity in enterprise AI, while the profitability of an individual FDE business depends on its customers and implementation model. Marc Andrusko of venture capital firm a16z points out that even large contracts can create a growing maintenance burden when each customer receives an independent system.[28]28Marc Andrusko, Andreessen Horowitz, The Palantirization of everything, 2026-01-16, “The services trap is real” and “A More Realistic Framework: When Does ‘Palantirization’ Make Sense?” ↩ a16z.com If staffing and support costs grow at the same rate as revenue, greater scale does little to improve margins. Without reusable products and capabilities, the development and support organization grows alongside the customer base.

Profitability in this market depends on winning large contracts for important customer workflows and reducing the development and support repeated across contracts. Reusing order lookup and approval capabilities for the next customer avoids paying again to develop what already exists. A company that builds an entirely new system for each customer must keep adding corresponding development staff as revenue grows. Differences in product reuse and staffing explain differences in profitability among companies using FDEs.

Q9. What do big tech and frontier AI companies gain by acquiring FDE businesses?

OpenAI is a prominent example of acquiring a company employing FDEs. In May 2026, it announced the launch of its Deployment Company for enterprise AI implementation and an agreement to acquire AI implementation company Tomoro. It said this would bring approximately 150 FDEs and deployment specialists, with more than $4 billion in initial funding to expand the business.[29]29OpenAI, OpenAI launches the OpenAI Deployment Company to help businesses build around intelligence, 2026-05-11, introduction, “Accelerating deployment from day one,” and “Scaling deployment across the economy.” ↩ openai.com In July, it announced another acquisition agreement, this time for Northslope, which embeds in customer organizations to build AI for critical workflows. The announcement described Tomoro as recently acquired and explained the goal of expanding the specialist workforce to hundreds of people and strengthening implementation expertise across industries.[30]30OpenAI Deployment Company, The OpenAI Deployment Company to acquire Northslope, 2026-07-08, introduction and “What’s next.” Describes Tomoro as recently acquired and the Northslope agreement as subject to closing conditions. ↩ deploy.co

An acquisition brings both FDE staff and the team’s accumulated implementation experience. Enterprise projects require more than code connecting a model: they involve obtaining data-access approval, coordinating workflows with business users, and overcoming legacy-system constraints. Acquiring a team that has already delivered projects together reduces the time needed to hire individuals and establish ways of working.

OpenAI also aims to build a structure for finding implementation customers and using deployment experience in product development. Investment and consulting partners have relationships with companies they invest in or advise. Those relationships and their experience changing business operations can help identify AI use cases and support adoption. OpenAI says it will combine this partner experience with its technology and have implementation teams work closely with its research and product teams.[31]31OpenAI, OpenAI launches the OpenAI Deployment Company to help businesses build around intelligence, 2026-05-11, “Building for where frontier AI is headed” and “Scaling deployment across the economy.” ↩ openai.com

Suppose an FDE building an agent discovers that it frequently misunderstands customer requests. The FDE shares the failed requests and expected outputs with research and product teams. If the same problem appears across customers, this evidence informs decisions about improving shared capabilities or the model. When research and product teams provide an improvement, the FDE applies it to customer systems and reruns previously failed requests to validate the result. Problems found in the field inform product improvements, which are then tested in the field again.

Commercially, this organization helps expand AI use from demonstrations into daily work. A demo may handle only a handful of return inquiries; in a production returns system, every incoming inquiry uses the model and product. If the workload reduction persists, customers have a reason to continue using it or extend it to other workflows. Suppliers gain an opportunity for sustained usage and revenue. Acquiring an FDE business is a way to assemble the team for this operational transition quickly.

Q10. FDE deployment adds to AX project costs. How do companies employing FDEs handle this expense?

When FDEs continue supporting customers after implementation, labor costs continue too. Palantir includes compensation for operations, maintenance, and professional-services staff, along with field-support expenses, in cost of revenue.[32]32Palantir Technologies Inc., 2025 Form 10-K, Filed in 2026 for fiscal 2025, Item 7, “Cost of Revenue.” Includes compensation for operations, maintenance, and professional-services staff and field-support costs in cost of revenue. ↩ www.sec.gov Companies recover these costs through separate service fees or product contracts, and reduce customer-specific effort by turning repeated development into product capabilities.

Palantir prices staffing and maintenance separately. Its 2024 UK public-procurement G-Cloud 14 price list charges £150,000 per person per quarter for implementation and engineering services. Workflow-specific maintenance is priced at £6,000, £14,000, or £20,000 per month depending on scale. Software licenses and these service items are sold separately.[33]33Palantir Technologies UK, Limited, Palantir Platform: Foundry & AIP – Pricing Document, 2024, G-Cloud 14 Framework, “Implementation and Support Services,” items 17–20. Lists implementation and engineering at £150,000 per person per quarter and workflow-specific maintenance at £6,000, £14,000, or £20,000 per month. Published public-procurement prices for 2024. ↩ assets.applytosupply.digitalmarketplace.service.gov.uk These are published customer billing rates. The structure charges corresponding service fees when more staff are assigned or support is extended.

Rapidflare, which makes AI software for the electronics industry, includes ongoing FDE support in its product contracts. It states that every customer receives an FDE from initial adoption throughout the relationship, without a separate professional-services agreement. When new requirements arise after launch, the assigned FDE helps modify configurations, data, integrations, and workflows.[34]34Rapidflare, Forward Deployed Engineers for Your AI Rollout, Undated; accessed 2026-11-14. Introduction, “Forward Deployed Engineering is not scoped around a project,” and “Frequently asked questions.” States that FDE is included in every customer contract, continues beyond initial implementation, and requires no separate professional-services agreement. ↩ www.rapidflare.ai Here, FDE support is part of the service sold with the product. The supplier covers continuing staffing costs from product-contract revenue.

Companies can also reduce the work required. Decagon moved tasks previously requiring customer-specific code into product configuration features and reported an 80% reduction in custom engineering work per agent.[35]35Katherine Xiao, Decagon, Introducing Agent Development: How Decagon is redefining forward deployment, 2026-06-22, “Forward deployment as a bottleneck” and “How we’re building it: Agent Development.” The 80% figure is the company-reported reduction in custom engineering work per agent. ↩ decagon.ai Instead of having engineers change code for every similar request, implementation staff use product capabilities to handle it. FDE support continues while engineers perform less repetitive development. Separate service billing and support included in product contracts determine how costs are charged; productization reduces the labor those payments must cover.

Q11. In what sequence do FDE projects proceed, and what matters most at each stage?

Teams define the problem and success criteria, implement a narrow workflow end to end, and expand while establishing validation and operating practices. OpenAI describes diagnosing and prioritizing high-value workflows, followed by design, development, testing, and deployment.[36]36OpenAI, OpenAI launches the OpenAI Deployment Company to help businesses build around intelligence, 2026-05-11, project process described in “Building for where frontier AI is headed.” ↩ openai.com Decagon similarly agrees on metrics, support channels, and target outcomes during initial discussions, then chooses a scope that can demonstrate value early.[37]37Sunny Rekhi, AI Engineer, How Forward Deployed Engineering is done at Decagon, AI Engineer World’s Fair 2026, 09:11–10:21 and 12:41–15:02. ↩ github-main-preview.ai.engineer Applied to the returns example, this approach looks as follows. The process below is hypothetical and intended to illustrate the approach.

First, the team reviews actual return inquiries alongside employees’ handling process. It measures time spent looking up orders and checking eligibility and identifies repeated inquiries. This evidence supports a choice of goals, such as reducing manual work or customer waiting time. Initial processing times and repeated-inquiry counts provide the baseline for assessing results after deployment. Development focuses on work that improves those metrics.

Next, the team connects the required information and permissions. It identifies where order data resides, how customers are identified, and which return policy applies. Permission to look up an order is considered separately from permission to register a return or issue a refund. If refunds above a threshold require employee approval, that condition enters the implementation scope. Tasks lacking the necessary data or permissions are excluded from the first release.

Development starts narrowly, for example with “routine return inquiries received by email.” Within that scope, implementation runs from understanding the inquiry through order lookup, eligibility checks, registration, and notification of the result. Stopping at plausible response text can leave the rest of the employee’s work intact. Business users try actual cases, and the team corrects missing conditions and unnecessary manual steps.

Validation covers normal cases as well as expired return windows, failed lookups, and requests requiring human judgment. The team checks when automation should stop and what information should reach an employee, then begins limited operations. Sierra describes turning errors found in real conversations into tests to check that a modified agent does not repeat them.[38]38Clay Bavor·Bret Taylor, Sierra, The Agent Development Life Cycle, 2024-06-03, “Quality assurance: Continuous, structured human feedback” and “Testing: Regression tests for conversations.” The returns workflow in the body is illustrative. ↩ sierra.ai In the returns example, adding a new rule triggers a rerun of existing inquiry tests, catching changes that break previously successful requests before deployment.

After operations begin, the team remeasures the agreed processing-time and employee-intervention metrics. It expands workflows with demonstrated benefits and assigns ownership for configuration changes, error response, and cost monitoring. If real use reveals more exceptions or a heavier review burden than expected, the FDE revisits earlier stages and adjusts the automation scope. Each stage provides evidence for the next decision; a first deployment does not freeze the goals or implementation scope.

Q12. How are FDE-built systems operated after the contract or development period ends?

Options include continuing product subscriptions and operational support, having FDEs continue improvements, and handing configuration and daily operations to the customer. These approaches can be combined.

Palantir provides operations and maintenance with both cloud-hosted and customer-hosted software. Professional services include user support, interface configuration, training, and continuing support for ontologies and data models that define the structure of business data. Professional-services contracts may have the same duration as product subscriptions or a different one.[39]39Palantir Technologies Inc., 2025 Form 10-K, Filed in 2026 for fiscal 2025, Item 7, “Palantir Cloud,” “On-Premises Software,” and “Professional Services.” States that professional-services terms may differ from subscription terms. ↩ www.sec.gov Under this arrangement, ending the initial implementation project does not end software operations and support.

At Ramp, the same FDE organization participates in sales, adoption, and long-term support. Its operating principle is continued responsibility for helping customers obtain value after onboarding.[40]40Leo Mehr, Ramp Builders Blog, Forward Deployed Engineering, 2025-08-05, “Team operating model: map to the customer lifecycle.” ↩ engineering.ramp.com The same organization follows how customers use the initial capabilities and what problems emerge later, carrying the workflow context learned during adoption into long-term support.

Decagon designs for customers to understand their agents’ structure and operate them themselves over time.[41]41Katherine Xiao, Decagon, Introducing Agent Development: How Decagon is redefining forward deployment, 2026-06-22, “Forward deployment as systems design.” ↩ decagon.ai Training customers to take over operations is part of implementation. Its official implementation guide includes training customer-service staff on monitoring tools and the Decagon management interface before deployment.[42]42Ryan Smith, Decagon, Complete AI customer support setup: Tools, integrations, and launch strategy, 2026-01-10, “Week 5: Convergence and preparation” and “Post-launch: Optimization and expansion.” Covers prelaunch training on monitoring tools and management interfaces, and postlaunch monitoring and workflow improvements. ↩ decagon.ai

Engineers, agent configuration staff, and customer education teams divide the handover work. Engineers handle complex integrations and product development, configuration staff handle setup and testing, and customer education and Decagon University support customers’ own operation.[43]43Sunny Rekhi, AI Engineer, How Forward Deployed Engineering is done at Decagon, AI Engineer World’s Fair 2026, “Make engineering work available through self-service,” 11:17–12:40, and slide 10, “Pushing work down the funnel.” Describes engineering, configuration, and customer-education responsibilities and support for customer operation through Decagon University. ↩ github-main-preview.ai.engineer In the returns example, the team exposes return windows and approval conditions as settings and teaches customer staff to modify them. Staff also learn to run tests after changes and review actual return-processing results. Transferring both capabilities and operational knowledge moves changes previously handled by the original FDE into the customer’s daily operations.

Q13. Which products and customers suit FDE? Do companies with development teams still need it?

FDE suits products that require substantial work to apply, customers who struggle to do that work alone, and implementations whose lessons can be reused elsewhere. Kevin Bai, an engineer on Anthropic’s applied AI team, describes selling complex platforms to nontechnical departments that cannot readily use them on their own as a representative FDE use case. Customer-specific solutions are built on a shared platform.[44]44Kevin Bai, AI Engineer, Forward Deployed Engineering 101, AI Engineer World’s Fair 2026, speaker introduction from 00:16, 04:03–06:30, and 09:29–13:09. ↩ github-main-preview.ai.engineer A product usable through configuration alone needs less field development; one requiring multiple systems and business rules to be connected needs engineers to link it to the customer’s work.

Employing developers does not mean having every capability required for successful AX. An internal team may know the company’s data structures and business rules without much experience of a specific model’s limitations, agent evaluation, or a supplier platform’s constraints. Conversely, an FDE familiar with the product and model cannot know the customer’s exceptions and approval structures from the outset. Whether the two teams can fill each other’s knowledge gaps informs the staffing decision.

In the example of jointly building a returns system, the internal team explains order-data semantics, permissions, and refund-approval conditions, and handles changes to internal systems. The FDE implements those conditions in the product and teaches the team how to find and evaluate agent failures. As the internal team learns to make changes and perform evaluations, it can handle more work without external help.

Anthropic’s training program illustrates this collaboration. It targets engineers with software-development fundamentals and LLM experience. Participants who pass initial training and a practical assessment work on their own company projects for 12 weeks with support from Anthropic engineers.[45]45Anthropic, Anthropic invests $100 million to train 10,000 engineers and tackle the enterprise AI talent gap, 2026-10-02, “The first program trains 10,000 Frontier Deployed Engineers.” Describes engineers working on their own organizations’ projects with support from Anthropic engineers. ↩ www.anthropic.com The aim is to add AI implementation experience to existing development skills so customer organizations can build systems themselves.

If an internal team has both the skills and the capacity to define problems, connect data, evaluate systems, and operate them, a permanently assigned external FDE is unnecessary. A team lacking only model-evaluation experience can develop that expertise through joint work at that stage. Companies with developers use FDEs to supplement product and AI implementation experience and help their own teams become capable of operating the systems.

Q14. Do FDEs build on existing products or develop separately and integrate later? How do they choose?

Using shared product capabilities is the default, but customer-specific development can supply missing functionality. Whether it later enters the shared product depends on its usefulness to other customers. Kevin Bai, an engineer on Anthropic’s applied AI team, distinguishes combining existing platform components into customer solutions, retaining one-customer behaviors in customer-specific implementations, and moving widely needed capabilities into the platform.[46]46Kevin Bai, AI Engineer, Forward Deployed Engineering 101, AI Engineer World’s Fair 2026, “Scale the design partnership, not the reinvention,” 07:19–09:28, and “Let field work inform the platform,” from 15:58. ↩ github-main-preview.ai.engineer

When the existing product falls short, development speed and long-term reuse both matter. Palantir describes choosing between adding a customer’s capability to the shared platform and providing a faster custom implementation as part of the FDE’s role. FDEs sometimes modify shared product code directly, with product-team design and code review, while coordinating large capabilities with the product roadmap.[47]47Palantir, Dev versus Delta: Demystifying engineering roles at Palantir, 2019-04-08, platform-versus-custom implementation decisions in “What are some examples of skills that are important for your role?—Delta,” and product-team collaboration in “Wait, what’s a Delta?” ↩ blog.palantir.com

The decision depends on when the customer needs the capability and how broadly it can be reused. A shared product capability can take longer because it must accommodate multiple customers and its relationship with existing functionality. Customer-specific implementations help validate immediate problems quickly but create a separate code-maintenance responsibility.

When the required functionality exists, teams can combine it and adjust its configuration. Expressing data needs, business rules, and execution order through settings reduces new coding. A particular customer’s system integration or unique business rules may require code connected through product APIs or extensions. Requirements unique to that customer remain in separate code.

If multiple customers need a capability and the roadmap can accommodate it, FDEs and product teams can build it as a shared feature from the outset. Settings or extensions handle differing customer conditions. If the existing product cannot readily support it and its business value is uncertain, separate code or a prototype can validate it first. Once its effect and reuse potential are established, the team redesigns it as a shared capability or keeps it customer-specific; ineffective implementations are discontinued.

Bringing customer-specific code into a product requires redesign to accommodate multiple customers. Bob McGrew, who led Palantir’s early product development, explains that directly adding one customer’s implementation can overfit the product to that customer. At Palantir, field-tested capabilities went to the product team, and FDEs serving other customers with similar needs joined the design process. Together they distinguished shared requirements from customer-specific differences.[48]48Bob McGrew, Y Combinator, The FDE Playbook for AI Startups with Bob McGrew, 2025-09-08, product-team responsibilities in the later part of “Consulting or Real Software?” and “The Birth of Palantir’s Ontology,” 17:54–23:03. Checked against the publicly available transcript. ↩ www.youtube.com

Continuing the illustrative returns example: existing order lookup and approval workflows can be reused, with separate code connecting the customer’s order system. An unproven new returns-decision capability can first be implemented within a limited scope. If it proves useful across customers, it becomes a shared capability, with varying return windows and approval conditions retained as settings. Reusable capabilities and designs accumulate in the product; field-written code does not all need to be incorporated unchanged.

Q15. FDE work seems difficult to scale. What scaling and platform approaches have actually been tried?

If staffing must grow in proportion to customers, hiring and training constrain business growth. Platformization turns repeatedly needed capabilities into a shared product and accommodates customer differences through configuration or limited additional development. Kevin Bai, an engineer on Anthropic’s applied AI team, explains that FDEs should build customer solutions on shared data models and basic capabilities.[49]49Kevin Bai, AI Engineer, Forward Deployed Engineering 101, AI Engineer World’s Fair 2026, 07:19–09:28 and 13:10–14:51. ↩ github-main-preview.ai.engineer Actual attempts extend reuse beyond feature development to deployment, evaluation, and operational knowledge.

Decagon’s CRM integration illustrates turning repeated customer-specific development into product functionality. For a customer-service agent to use customer information, it must connect to the customer relationship management (CRM) system storing it. Initially, Decagon engineers wrote this integration code for each customer. Sunny Rekhi, Decagon’s FDE CTO, says that around the 25th custom integration, the team built an integration feature that customers or agent implementation staff could configure themselves.[50]50Sunny Rekhi, AI Engineer, How Forward Deployed Engineering is done at Decagon, AI Engineer World’s Fair 2026, 11:17–12:40. ↩ github-main-preview.ai.engineer Subsequent customers using it configure the integration in the product instead of requesting new code from an engineer. Capabilities developed from earlier customers’ requirements become available to later customers.

Changing business rules followed a similar path. Katherine Xiao, Decagon’s Director of Implementations, gives the example of changing a promotional policy. Previously, an engineer spent a full development sprint writing custom code; now an agent implementation specialist designs, debugs, and tests the rules in natural language in an afternoon. She reports that moving repeated development into product features reduced custom engineering work per agent by 80%.[51]51Katherine Xiao, Decagon, Introducing Agent Development: How Decagon is redefining forward deployment, 2026-06-22, “How we’re building it: Agent Development.” The original measurement is “custom engineering work per agent.” ↩ decagon.ai Consider an illustrative rule such as “apply the discount only to customers who purchased during the campaign.” Work that previously required changing code can now be expressed and tested as text within the product. This reduces repeated customer-specific code changes and expands what implementation specialists can handle themselves.

Software deployment and updates can also become shared tooling. Palantir’s Apollo deploys and manages software across clouds, on-premises servers, and disconnected networks. It defines constraints such as deployment windows and software dependencies, then coordinates changes accordingly.[52]52Palantir, Apollo documentation: Introduction, Undated; accessed 2026-11-14. Introduction and “The Apollo solution.” ↩ www.palantir.com Even when customers allow updates at different times, a shared system can check those conditions, reducing engineers’ work of checking schedules and deploying separately in each environment. Customer differences are expressed as rules the system can process.

AI agent evaluation and change management are reusable too. Sierra groups versions of the model, prompt, and knowledge resources into a single deployment unit and turns errors in actual conversations into repeatable tests.[53]53Clay Bavor·Bret Taylor, Sierra, The Agent Development Life Cycle, 2024-06-03, “Release: Immutable agent snapshots” and “Testing: Regression tests for conversations.” ↩ sierra.ai This makes it easier to identify the exact combination behind a problematic agent and roll back. Previously collected cases can also be rerun before a new deployment. Teams avoid rebuilding validation procedures from scratch as personnel change or the scope expands.

AWS’s partner FDE program preserves implementation tools and decision rationales for future projects. It creates reusable assets such as data models defining business concepts, evaluation frameworks, and MCP servers connecting AI to external systems, and records why an architecture was selected and which evaluation criteria apply. Partners retain these assets for independent reuse in later implementations.[54]54Chandra Pinapala·Erik Farr·Karthik Sonti, AWS Partner Network Blog, Introducing Forward Deployed Engineering for Partners: Winning the Future of Enterprise AI, 2026-06-30, “What Makes This Different: The Delivery Harness” and “How it Works.” ↩ aws.amazon.com A code-only handover can force the next team to rediscover why the system was built that way. Recording applicability conditions and reasoning makes it easier to decide what to reuse and what to change for a new customer.

Some efforts connect field problems to improvements in the model itself. Colin Jarvis, OpenAI’s Head of Forward Deployed Engineering, describes testing different slide representations while building a slide-generation tool for a Japanese sales team. The team selected a representation suited to the model, created examples, and passed them to the post-training team. He recalls improved slide-generation results in a later model version.[55]55Colin Jarvis, South Park Commons, The Future of Forward Deployed Engineering | OpenAI, Ramp, Nominal, Dataland, 2026-05-28, 29:37–32:56. Checked against the publicly available transcript. ↩ www.youtube.com Post-training improves a pretrained model’s capabilities through additional training. Here, the FDE shared field-discovered methods and examples with researchers to improve model functionality.

These examples moved customer-specific repeated work into products at different levels. Integrations and policy changes became configuration features; deployment and validation became shared tools; knowledge useful across projects became design documentation and evaluation frameworks. Their approach to scale assigns products and operating tools the code and procedures FDEs would otherwise repeat for every customer.

Notes

  1. 1.Palantir Technologies Inc., 2020 IPO prospectus, 2020, “Business—Overview.” Describes the company’s founding in 2003 and its initial work for US intelligence agencies. ↩www.sec.gov
  2. 2.Bob McGrew, Y Combinator, The FDE Playbook for AI Startups with Bob McGrew, 2025-09-08, 03:19–07:55. Checked against the publicly available transcript. ↩www.youtube.com
  3. 3.Palantir, Dev versus Delta: Demystifying engineering roles at Palantir, 2019-04-08, “Wait, what’s a Delta?” ↩blog.palantir.com
  4. 4.Leo Mehr, Ramp Builders Blog, Forward Deployed Engineering, 2025-08-05, “FDE at Ramp.” ↩engineering.ramp.com
  5. 5.Palantir, Dev versus Delta: Demystifying engineering roles at Palantir, 2019-04-08, product-code contributions in “Wait, what’s a Delta?” and “What does the day-to-day of your role look like?—Delta.” ↩blog.palantir.com
  6. 6.Palantir, Dev versus Delta: Demystifying engineering roles at Palantir, 2019-04-08, “Wait, what’s a Delta?” Manufacturing defect-rate example and performance evaluation against customer goals. ↩blog.palantir.com
  7. 7.Palantir, Dev versus Delta: Demystifying engineering roles at Palantir, 2019-04-08, “What are some examples of skills that are important for your role?—Delta,” on problem decomposition, independent learning, and technical judgment. ↩blog.palantir.com
  8. 8.Palantir, Dev versus Delta: Demystifying engineering roles at Palantir, 2019-04-08, “What are some examples of skills that are important for your role?—Delta,” on user empathy, business strategy, and independent judgment. The returns analysis in the body is illustrative. ↩blog.palantir.com
  9. 9.Leo Mehr, Ramp Builders Blog, Forward Deployed Engineering, 2025-08-05, “FDE Principles,” including “Always be scoping” and the trade-off between reusable capabilities and minimum implementation scope. The return-window configuration in the body is illustrative. ↩engineering.ramp.com
  10. 10.Clay Bavor·Bret Taylor, Sierra, The Agent Development Life Cycle, 2024-06-03, “Quality assurance: Continuous, structured human feedback” and “Testing: Regression tests for conversations.” The returns-agent test cases and failure diagnosis in the body are illustrative. ↩sierra.ai
  11. 11.Marc Andrusko, Andreessen Horowitz, The Palantirization of everything, 2026-01-16, “Not every problem is a ‘Palantir-grade’ problem” and “A More Realistic Framework: When Does ‘Palantirization’ Make Sense?” The customer-selection criteria in the body synthesize its discussion of problem value, customer-level economics, and shared workflows. ↩a16z.com
  12. 12.Bob McGrew, Y Combinator, The FDE Playbook for AI Startups with Bob McGrew, 2025-09-08, later part of “Why AI Companies Adopt It,” on organizational approval and executive sponsorship. Checked against the publicly available transcript. ↩www.youtube.com
  13. 13.Leo Mehr, Ramp Builders Blog, Forward Deployed Engineering, 2025-08-05, “Team operating model: map to the customer lifecycle” and “FDE Principles—Always be scoping.” ↩engineering.ramp.com
  14. 14.Bob McGrew, Y Combinator, The FDE Playbook for AI Startups with Bob McGrew, 2025-09-08, later part of “Why AI Companies Adopt It,” on problem value, pricing, and risk, and “What Success Metrics Look Like,” from 36:17. Checked against the publicly available transcript. ↩www.youtube.com
  15. 15.Colin Jarvis, OpenAI, What FDE means at OpenAI, 2026-10-22, introduction defining “the deployment gap” between model capabilities and usable systems. ↩openai.com
  16. 16.OpenAI, OpenAI launches the OpenAI Deployment Company to help businesses build around intelligence, 2026-05-11, “Building for where frontier AI is headed.” Describes connecting models to customer data, tools, controls, and workflows for reliable daily use. ↩openai.com
  17. 17.Clay Bavor·Bret Taylor, Sierra, The Agent Development Life Cycle, 2024-06-03, introduction; “Development: Declarative goals and guardrails”; “Quality assurance: Continuous, structured human feedback”; and “Testing: Regression tests for conversations.” The 30-day return condition is an example from the source. ↩sierra.ai
  18. 18.PagerDuty, Forward Deployed Engineer (SDEV) - London, England, United Kingdom, Undated; accessed 2026-11-14. Company and role introduction and “Key Responsibilities.” Describes the advertised responsibilities, not outcomes from a particular customer project. ↩careers.pagerduty.com
  19. 19.Vinoo Ganesh, AI Engineer, How Forward Deployed Engineering is done at Kepler, AI Engineer World’s Fair 2026, “Find the decision behind the dashboard,” 05:00–08:27. The requirements length and implementation time are the speaker’s recollections. ↩ai.engineer
  20. 20.Sunny Rekhi, AI Engineer, How Forward Deployed Engineering is done at Decagon, AI Engineer World’s Fair 2026, “Use support history to choose what to automate,” 13:44–15:02. ↩github-main-preview.ai.engineer
  21. 21.Bob McGrew, Y Combinator, The FDE Playbook for AI Startups with Bob McGrew, 2025-09-08, “Why AI Companies Adopt It,” from 23:04. Checked against the publicly available transcript. ↩www.youtube.com
  22. 22.Joe Schmidt, Andreessen Horowitz, Trading Margin for Moat: Why the Forward Deployed Engineer Is the Hottest Job in Startups, 2025-06-04, “The implementation necessity.” Presents the investment thesis that AI can make AI implementation work more efficient and automated. The body explains the economic implications of lower customer-specific development costs. ↩a16z.com
  23. 23.Chandra Pinapala·Erik Farr·Karthik Sonti, AWS Partner Network Blog, Introducing Forward Deployed Engineering for Partners: Winning the Future of Enterprise AI, 2026-06-30, introduction and “Why Partners, Why Now.” ↩aws.amazon.com
  24. 24.Anthropic, Anthropic invests $100 million to train 10,000 engineers and tackle the enterprise AI talent gap, 2026-10-02, introduction and “The first program trains 10,000 Frontier Deployed Engineers.” ↩www.anthropic.com
  25. 25.Tim Tully·Joff Redfern·Deedy Das·Derek Xiao, Menlo Ventures, 2026: The State of Enterprise AI, 2026, “Enterprise AI Reaches Scale” and “Methodology”; accessed 2026-11-14. Market size is an annualized revenue estimate combining surveys, disclosures, and external sources. Historical figures were restated to match the classification framework. ↩menlovc.com
  26. 26.Accenture, Reinventing what’s possible: FY2025 Annual Report, 2025, p. 5, “AI Leadership—Yesterday, Today and Tomorrow.” Revenue and bookings cover advanced AI, excluding data, traditional AI, and AI used in service delivery. ↩investor.accenture.com
  27. 27.Palantir Technologies Inc., 2025 Form 10-K, Filed in 2026 for fiscal 2025, Item 7, “Results of Operations” and “Cost of Revenue and Gross Profit.” Gross and operating margins are company-wide GAAP figures using rounded percentages from the filing. ↩www.sec.gov
  28. 28.Marc Andrusko, Andreessen Horowitz, The Palantirization of everything, 2026-01-16, “The services trap is real” and “A More Realistic Framework: When Does ‘Palantirization’ Make Sense?” ↩a16z.com
  29. 29.OpenAI, OpenAI launches the OpenAI Deployment Company to help businesses build around intelligence, 2026-05-11, introduction, “Accelerating deployment from day one,” and “Scaling deployment across the economy.” ↩openai.com
  30. 30.OpenAI Deployment Company, The OpenAI Deployment Company to acquire Northslope, 2026-07-08, introduction and “What’s next.” Describes Tomoro as recently acquired and the Northslope agreement as subject to closing conditions. ↩deploy.co
  31. 31.OpenAI, OpenAI launches the OpenAI Deployment Company to help businesses build around intelligence, 2026-05-11, “Building for where frontier AI is headed” and “Scaling deployment across the economy.” ↩openai.com
  32. 32.Palantir Technologies Inc., 2025 Form 10-K, Filed in 2026 for fiscal 2025, Item 7, “Cost of Revenue.” Includes compensation for operations, maintenance, and professional-services staff and field-support costs in cost of revenue. ↩www.sec.gov
  33. 33.Palantir Technologies UK, Limited, Palantir Platform: Foundry & AIP – Pricing Document, 2024, G-Cloud 14 Framework, “Implementation and Support Services,” items 17–20. Lists implementation and engineering at £150,000 per person per quarter and workflow-specific maintenance at £6,000, £14,000, or £20,000 per month. Published public-procurement prices for 2024. ↩assets.applytosupply.digitalmarketplace.service.gov.uk
  34. 34.Rapidflare, Forward Deployed Engineers for Your AI Rollout, Undated; accessed 2026-11-14. Introduction, “Forward Deployed Engineering is not scoped around a project,” and “Frequently asked questions.” States that FDE is included in every customer contract, continues beyond initial implementation, and requires no separate professional-services agreement. ↩www.rapidflare.ai
  35. 35.Katherine Xiao, Decagon, Introducing Agent Development: How Decagon is redefining forward deployment, 2026-06-22, “Forward deployment as a bottleneck” and “How we’re building it: Agent Development.” The 80% figure is the company-reported reduction in custom engineering work per agent. ↩decagon.ai
  36. 36.OpenAI, OpenAI launches the OpenAI Deployment Company to help businesses build around intelligence, 2026-05-11, project process described in “Building for where frontier AI is headed.” ↩openai.com
  37. 37.Sunny Rekhi, AI Engineer, How Forward Deployed Engineering is done at Decagon, AI Engineer World’s Fair 2026, 09:11–10:21 and 12:41–15:02. ↩github-main-preview.ai.engineer
  38. 38.Clay Bavor·Bret Taylor, Sierra, The Agent Development Life Cycle, 2024-06-03, “Quality assurance: Continuous, structured human feedback” and “Testing: Regression tests for conversations.” The returns workflow in the body is illustrative. ↩sierra.ai
  39. 39.Palantir Technologies Inc., 2025 Form 10-K, Filed in 2026 for fiscal 2025, Item 7, “Palantir Cloud,” “On-Premises Software,” and “Professional Services.” States that professional-services terms may differ from subscription terms. ↩www.sec.gov
  40. 40.Leo Mehr, Ramp Builders Blog, Forward Deployed Engineering, 2025-08-05, “Team operating model: map to the customer lifecycle.” ↩engineering.ramp.com
  41. 41.Katherine Xiao, Decagon, Introducing Agent Development: How Decagon is redefining forward deployment, 2026-06-22, “Forward deployment as systems design.” ↩decagon.ai
  42. 42.Ryan Smith, Decagon, Complete AI customer support setup: Tools, integrations, and launch strategy, 2026-01-10, “Week 5: Convergence and preparation” and “Post-launch: Optimization and expansion.” Covers prelaunch training on monitoring tools and management interfaces, and postlaunch monitoring and workflow improvements. ↩decagon.ai
  43. 43.Sunny Rekhi, AI Engineer, How Forward Deployed Engineering is done at Decagon, AI Engineer World’s Fair 2026, “Make engineering work available through self-service,” 11:17–12:40, and slide 10, “Pushing work down the funnel.” Describes engineering, configuration, and customer-education responsibilities and support for customer operation through Decagon University. ↩github-main-preview.ai.engineer
  44. 44.Kevin Bai, AI Engineer, Forward Deployed Engineering 101, AI Engineer World’s Fair 2026, speaker introduction from 00:16, 04:03–06:30, and 09:29–13:09. ↩github-main-preview.ai.engineer
  45. 45.Anthropic, Anthropic invests $100 million to train 10,000 engineers and tackle the enterprise AI talent gap, 2026-10-02, “The first program trains 10,000 Frontier Deployed Engineers.” Describes engineers working on their own organizations’ projects with support from Anthropic engineers. ↩www.anthropic.com
  46. 46.Kevin Bai, AI Engineer, Forward Deployed Engineering 101, AI Engineer World’s Fair 2026, “Scale the design partnership, not the reinvention,” 07:19–09:28, and “Let field work inform the platform,” from 15:58. ↩github-main-preview.ai.engineer
  47. 47.Palantir, Dev versus Delta: Demystifying engineering roles at Palantir, 2019-04-08, platform-versus-custom implementation decisions in “What are some examples of skills that are important for your role?—Delta,” and product-team collaboration in “Wait, what’s a Delta?” ↩blog.palantir.com
  48. 48.Bob McGrew, Y Combinator, The FDE Playbook for AI Startups with Bob McGrew, 2025-09-08, product-team responsibilities in the later part of “Consulting or Real Software?” and “The Birth of Palantir’s Ontology,” 17:54–23:03. Checked against the publicly available transcript. ↩www.youtube.com
  49. 49.Kevin Bai, AI Engineer, Forward Deployed Engineering 101, AI Engineer World’s Fair 2026, 07:19–09:28 and 13:10–14:51. ↩github-main-preview.ai.engineer
  50. 50.Sunny Rekhi, AI Engineer, How Forward Deployed Engineering is done at Decagon, AI Engineer World’s Fair 2026, 11:17–12:40. ↩github-main-preview.ai.engineer
  51. 51.Katherine Xiao, Decagon, Introducing Agent Development: How Decagon is redefining forward deployment, 2026-06-22, “How we’re building it: Agent Development.” The original measurement is “custom engineering work per agent.” ↩decagon.ai
  52. 52.Palantir, Apollo documentation: Introduction, Undated; accessed 2026-11-14. Introduction and “The Apollo solution.” ↩www.palantir.com
  53. 53.Clay Bavor·Bret Taylor, Sierra, The Agent Development Life Cycle, 2024-06-03, “Release: Immutable agent snapshots” and “Testing: Regression tests for conversations.” ↩sierra.ai
  54. 54.Chandra Pinapala·Erik Farr·Karthik Sonti, AWS Partner Network Blog, Introducing Forward Deployed Engineering for Partners: Winning the Future of Enterprise AI, 2026-06-30, “What Makes This Different: The Delivery Harness” and “How it Works.” ↩aws.amazon.com
  55. 55.Colin Jarvis, South Park Commons, The Future of Forward Deployed Engineering | OpenAI, Ramp, Nominal, Dataland, 2026-05-28, 29:37–32:56. Checked against the publicly available transcript. ↩www.youtube.com
Written by Jonghyuk Max KimSend emailCopy linkShare on X
← Back to all postsPreviousNextRandom