AI Engineer World's Fair 2026

Forward Deployed Engineering 101

Kevin Bai· Member of Technical Staff, Anthropic17:48

Loading slides…

Read the talk

Forward Deployed Engineering 101

Forward deployed engineering connects complex platforms to business outcomes, but its economics depend on shared primitives and a clear boundary between customer work and product development.

From a talk by Kevin Bai

What does organized data do for the business?

What does a customer actually gain from buying a powerful platform? That is the practical question behind forward deployed engineering as a go-to-market motion. Kevin Bai approaches it from work at Anthropic’s applied AI team, where he introduces himself as a member of technical staff; before that, he was the first person on Rippling’s FDE team, and earlier worked at Palantir. Bai says Rippling’s FDE team grew to around 25 people in one year. His starting point is why Palantir adopted the function and when another business might need it.

Consider Palantir’s Foundry. It centralizes an organization’s data, creates an ontology, and provides a platform for building applications. Bai introduces the ontology through a warehouse example: instead of disconnected tables, the business gets a single source of truth for its warehouses. Data acquires business meaning—proper nouns that people can work with. This is an introductory analogy; the Ontology encompasses business objects, relationships, actions, and functions, rather than simply consolidated tables.

An industry buyer can still reasonably ask what this organization of data does for the business. The platform’s success depends on the customer’s ability to use it, and that creates an adoption cost beyond the purchase price: employees must learn the platform before they can build useful applications. The vendor has sold a capability, but realizing its value remains the customer’s problem.

FDE combines software and engineering services into a purchased outcome. Engineers learn the customer’s business and build a solution on Foundry. The customer is buying neither software alone nor simply someone’s time. A consumer packaged goods company, for example, might want more shelf placement or higher sales throughput. How its data is organized matters as an implementation detail supporting those results.

0:160:31
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:16 · section reference included

Match the product’s complexity to the buyer

Bai frames the adoption problem as a two-axis grid, using a Punnett-square analogy: what are you selling, and who is buying it? Complexity alone does not require FDE. What matters is whether the customer can absorb that complexity and turn the product into something useful.

OfferingBuyer or userHow complexity is handled
GitHub or DatadogCTOs, CIOs, software engineersTechnical users learn and operate the software
Rippling, Jira, or SlackNontechnical buyers and usersUsers configure the tool
A platform requiring custom developmentNontechnical business buyerFDEs build the customer’s solution

The second row reflects how Bai positions these products in this comparison: configurable tools, rather than platforms the customer must develop on. His criterion for FDE is the third case—a technically demanding offering that must reach a nontechnical buyer.

That mismatch helps explain Palantir’s market. Google, Meta, and AI labs have engineers who can build internal applications. A Fortune 500 oil-and-gas company may lack comparable software engineering depth; its pipelines move physical materials, not data. Selling it an application platform leaves a substantial gap between purchase and use.

FDE fills that gap by supplying engineers whom the customer does not need to recruit, hire, manage, or retain. They already understand the platform and work closely with the customer to discover and solve its problem. Bai compares this attention to a waiter at a fine-dining restaurant: understanding what the customer needs is part of delivering the service. Combined with the platform, that became Palantir’s way of selling to the global Fortune 500.

4:034:15
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

4:03 · section reference included

The commercial case for delivering outcomes

To illustrate the commercial appeal, Bai compares customer spending across what he describes as public SaaS companies in the Fortune 500. He calls the measure ACV, meaning average contract value here; it should not be confused with annual contract value, another common use of the abbreviation.

Company or groupAverage contract value in Bai’s account
Palantir$4 million, as of his last check
ServiceNow$1.2 million
Workday$600,000, recalled tentatively
All remaining public SaaS companiesNone above $500,000

These are Bai’s reported figures, without a specified measurement date, calculation, or independently established ranking. They illustrate his case for the model’s commercial success rather than establish a like-for-like financial benchmark. He also points to Palantir’s high valuation relative to a workforce of a few thousand people, without supplying a valuation figure.

6:316:39
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

6:31 · section reference included

Scale the design partnership, not the reinvention

The operating model resembles a startup design partnership. Early in a B2B company’s life, neither the product nor the customer’s purchase may be fully defined. The customer supplies context about its problem; the startup supplies time, technology, resources, and engineering to build a solution. That close collaboration helps the startup find product-market fit. FDE extends the same arrangement into enterprise engagements instead of treating it as something a company must leave behind after its earliest stage.

The obvious objection is maintenance. If every customer receives a separate application built from scratch, the company accumulates bespoke systems that engineers must learn and support. Bai makes the burden tangible with the prospect of learning 55 repositories. In his distinction, that is a development shop: a legitimate business, but not the platform-based FDE model he is describing.

Shared primitives make enterprise design partnerships maintainable. FDEs assemble existing platform capabilities into customer applications, workflows, and solutions. The customer-specific work sits on top of an already supported foundation. Without that foundation, each engagement repeats the same engineering work, maintenance costs consume the business’s profit and loss, and engineers inherit an increasingly unattractive collection of systems.

7:197:30
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

7:19 · section reference included

Two gates before building an FDE function

Applying the model starts with necessity, not enthusiasm for a fashionable role. Bai proposes two gates:

  1. Does the business need this go-to-market motion? Identify a situation where a technically complicated offering must be sold to a nontechnical buyer. If that mismatch is absent, FDE is probably the wrong fit. Developer relations and developer engagement can serve technical buyers; a sales-led motion can serve more traditional SaaS.
  2. Is there a platform, or a commitment to build one? Engineers need shared primitives beneath their customer work. The prospect of engineers directly generating revenue does not remove the cost of supporting what they build.

Even a robust platform leaves a substantial maintenance burden. The platform is a prerequisite for managing that burden, not a promise that it disappears.

AI makes the question more widespread. Speaking in a 2026 context, Bai contrasts the present with Palantir’s early market years, which he tentatively places around 2004 or 2005. Code and sophisticated custom software have become easier to produce, including agents for insurance, legal work, and other domains. His hypothesis is that software’s business model is changing: as platforms become agentic, they become more customizable, and more customers face a product whose possibilities they do not understand or know how to implement.

That creates a familiar adoption risk. If the vendor leaves success entirely to the customer’s implementation ability, moving upmarket or expanding into new horizontal or vertical markets becomes harder. Easier software construction does not automatically give customers the context or engineering capacity to realize its value. The subsequent audience questions turn to how an FDE function should operate; Bai excludes his current work from that discussion.

9:299:42
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

9:29 · section reference included

How much should a primitive already do?

The first implementation question is how atomic the shared primitives should be. For a platform involving data models, Bai suggests starting by avoiding the need to define a data model from scratch. Beyond that, the right level of reuse depends on the use case and customer base.

  • Substantial application foundations: In some industries, Bai suggests an application might be 60% prebuilt and 40% customized. The split illustrates a possible arrangement, not a prescribed target.
  • Granular building blocks: Other industries and use cases require fine-grained configuration and tooling, so a nearly finished application would constrain the work too much.

AWS provides the analogy for broad primitives. An engineering team could buy server racks, bring them online, and maintain them, but a platform takes that work off its hands. Within AWS, DynamoDB supplies a database capability the team does not need to invent itself. AWS serves a very broad customer base, which explains the usefulness of general building blocks. A narrower customer base may justify primitives that encode much more of the application.

13:1013:23
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

13:10 · section reference included

Share knowledge and clarify the working relationship

An audience question about collaboration initially draws an answer about having multiple FDEs on one project. Bai encourages that pattern because customer work should not depend on a single person holding all the context. If that engineer goes on vacation, the rest of the team still needs to understand and support the solution.

The questioner then clarifies that the engineers come from different companies—for example, AWS and Palantir working on the same project. After distinguishing a competitive bake-off from a partnership, Bai confirms that cross-company collaboration exists. His analogy is a contractor relationship: establish who occupies the contractor role. The answer supplies a way to think about the relationship, rather than a detailed process for managing friction between vendors.

14:5215:05
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

14:52 · section reference included

Let field work inform the platform

Where should a requested change live: in the platform or in the customer’s implementation? Bai draws the boundary around generalizability.

ChangeLong-term home
Bespoke behavior unique to one customerThat customer’s implementation
A capability useful beyond one customerThe shared platform

An early FDE program may begin with relatively few primitives. Field work then becomes a way to discover what else the product should support. FDEs scout ahead: they encounter customer problems, identify capabilities that generalize, and help reveal additional products and services the business can build. Generalization is a long-term direction, not a requirement to put every customer request into the platform immediately.

The final hiring question brings the function back to the person doing the work. Bai’s definition is a “customer-facing software engineer.” The candidate must meet the team’s software engineering hiring bar, and must also be someone the company trusts in front of a customer. Both requirements matter: the role has to understand the customer well enough to discover the right problem and have the engineering ability to build its solution.

15:5816:10
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

15:58 · section reference included

Resources

From the talk

Transcript download
Full transcript

Loading transcript…