A forward deployed engineer is a software engineer who works inside your business instead of at a vendor's desk. You hand them a vague operational problem, the sort nobody has written a specification for. They build the fix into the systems your staff already use, and they stay accountable after go-live for whether people use it. Palantir describes forward deployed software engineers as engineers who work directly with customers to understand their greatest problems and build solutions against them. Anthropic describes the role as an engineer who embeds with strategic customers to drive AI adoption and ships applications inside the customer's own systems. The title came out of large enterprise contracts. The work under it fits a firm of ten as well as it fits an agency.
The test that separates the role from a demo
The Bureau of Labor Statistics defines a sales engineer as someone who sells business products or services, such as software or support, that require technical expertise. That occupation exists in the official handbook. Forward deployed engineer does not. That is part of why the title means several things depending on who prints it.
In a vendor meeting, ask who owns the result. A salesperson or a solutions engineer proves the software can work: demo tenant, sample data, answers to your questions. The contract starts there. The forward deployed engineer is the person who makes it work for this customer, on their data, after signature.
Kevin Bai of Anthropic's applied AI team gave a session titled Forward Deployed Engineering 101 at the AI Engineer World's Fair, which ran June 29 to July 2, 2026 in San Francisco, and the session's written summary defines the role as a customer-facing software engineer who combines software and engineering services into a purchased outcome. You are buying something that runs.
Exhibit 1
A conventional vendor and a forward deployed engineer split the same project three ways: who writes the spec, where the building happens, who owns the result after go-live
Where the title came from
Palantir calls itself the originator of the job, writing that it pioneered the position by embedding engineers directly with customers. The company's page gives no year. The date most often printed online comes from a16z, where Tom Hollands writes that Palantir gave its solutions and integration engineers the forward deployed engineer title in 2011 and called them Deltas. That is a venture partner's account in an essay, not a Palantir statement.
Palantir's job board runs the role under two team labels, Echo and Delta, and lists Deployment Strategist roles alongside Forward Deployed Engineer roles, with onsite visits to customer analysts named in the strategist posting.
Nabeel Qureshi, who worked at Palantir as a forward deployed engineer after joining in 2015, describes a two-track split: engineers who go to the customer and a product team that turns what they built into software. In his account, forward deployed engineers worked at the customer's offices three or four days a week, and his recollection covers the mid 2010s. Palantir's current posting compares the job to a startup chief technology role: small teams, little supervision, ownership of a project end to end.
Why the AI labs hire it now
Joe Schmidt wrote in a16z in June 2025 that enterprise AI products need more implementation work than earlier software did, because they need deep integrations and context. Buyers of AI want the result but need the vendor to set it up. He describes the forward deployed engineer as a professional services role, sometimes a rebranded implementation or solutions specialist. That is how the essay describes the role, and the title has no entry in the official occupation handbook.
The same essay says that model providers themselves hire forward deployed and solutions engineers, because models alone do not land inside a customer's systems, and that much of the difference between products is how the same underlying technology gets implemented. That is an investment thesis in an essay, with no sample and no method.
The job pages match it. OpenAI describes forward deployed engineers as owning discovery, scoping, design, build and production rollout, working alongside the customer's own engineers, and says the role is measured on adoption in production and on workflow impact it can point to, not on demos. Anthropic lists two kinds of work: speeding up adoption of products the customer already bought, and building new applications on its models.
What one week of the work looks like
The engineer sits with the people who already do the task and asks them to walk through it on their own screen. A screen shows the workaround and the second spreadsheet kept beside the official one.
Then a small build, done in public. Something touches real data early and goes in front of the staff who will use it before it goes in front of the owner. Their criticism is cheap then and expensive after go-live. The owner's demo answers other questions: what changed, what it costs to run each month, what breaks and who fixes it. After that comes live use, and the fixes come from what staff do with it.
Exhibit 2
One engagement week runs from watching the task, through a small build on real data, to a staff demo, an owner demo, and fixes taken from live use
OpenAI's list of what the role owns has five stages in it, and one word of those five is code. The rest is conversation, scoping and follow-through. The person doing this work cannot also be the person selling it.
Why it stayed an enterprise arrangement
Embedding an engineer at a customer's office for most of the week only pays on a large contract, and a firm of ten cannot buy it on those terms. The setup work on older software was thin enough that a vendor could ship self-serve onboarding and let the customer work out the rest.
The Census Bureau measured national AI use at 19.8% of businesses as of the survey period ending May 3, 2026, with overall use between 17% and 20% from December 2025 through May 2026. Use rose among firms with at least 20 employees over that window and did not change significantly among firms with fewer than 20. In the survey period covering September 21 to October 4, 2026, 23.9% of firms with four or fewer employees said they used AI in at least one business function in the previous two weeks, up from under 20% for that size class in the December 2025 to May 2026 window. Census changed the question in November 2025, from AI use in producing goods or services to AI use in any business function, so figures on either side of that change are not the same measurement.
The U.S. Chamber of Commerce reported in October 2026 that 66% of small businesses use AI, based on a survey of 3,732 firms with fewer than 250 employees fielded between July 23 and August 6, 2026. Carry the size caveat with that number, since the band is far wider than the firm most readers run. The same report, covering firms with fewer than 250 employees, finds that about two thirds of small businesses rely mostly or entirely on AI tools built by other companies, and only 8% are building their own. Among small businesses that pay for AI, the Chamber reports median monthly AI spend of $250.
Small firms buy AI. They rarely build it, and the buyer has no engineering team to absorb the setup. Bai sets two conditions before a company needs the role at all: a technically complicated product sold to a buyer who cannot absorb the complexity, and a shared platform for the engineers to build on. General assistants and off-the-shelf integration tooling are that platform. The deployment work did not disappear. It moved to whoever the buyer hires to do it.
What the role looks like scaled to ten people
At this size the engagement is one workflow at a time, and choosing it is a separate job from building anything. Our guide to finding the bottleneck that is actually holding the business back covers how to choose, and so does listing the everyday tasks AI can take over in your own firm. Which workflow goes first decides whether the engineering is worth anything.
Three rules carry the rest. The staff keep the screens they know, so the work goes into the booking system, the practice software, the job board and the invoice tool you already pay for. A named person approves anything irreversible, because a machine that sends, refunds and deletes on its own will eventually do all three at once. And you measure the change before and after on the same terms, counted the same way in both periods, by someone who writes the method down before the build starts.
Exhibit 3
A first small-business deployment holds one workflow, existing tools, an approval step and a measurement, and leaves out a platform rebuild, a new system of record and unattended irreversible actions
The regulated version is the same work with paperwork attached. Federal guidance defines a business associate as someone outside a covered entity's workforce engaged to carry out functions or services that involve protected health information, and it requires a written business associate agreement before that information is disclosed. The guidance names a third-party AI chatbot on a provider's patient portal as an example, and the obligation runs downstream to any subcontractor that touches the data. A firm becomes a business associate when its work involves protected health information for a covered entity. What HIPAA requires of a medical practice using ChatGPT sets out the detail.
The limits, and what to ask anyone who offers it
This is a services engagement. A person's days are the product, so the price behaves like a person's days. Schmidt's description of it as professional services work is the useful frame: expect it to be quoted, scoped and invoiced that way.
Custom work has to be left behind in a form the business can run without the builder. Ask before the engagement starts who maintains it in month seven, what happens when the model behind it changes, and what documentation exists at handover. An implementation only its author can touch is a dependency you have rented.
On the call, ask four questions. Which workflow, named. Measured how, by whom, counted from which system in both periods. Who owns the accounts, the data and the keys afterwards. What happens when it breaks, and who is on call for it. A vendor who calls every engagement forward deployed is selling a sales call with a technical name on it. Where the data goes is a decision of its own, and which parts of your work can go in ChatGPT, and which need AI that stays in-house sets out that split. Paying for this work or doing it yourself is a separate judgment.
Nihul Consulting does this implementation work for firms of three to twenty people, described on our AI enablement service page.