← Back to Archive Field Note June 2026 中文
Field Note · Computing Business History

Same job,
eight costumes in sixty years

The FDE isn't a new role. From the IBM systems engineer to the forward deployed engineer, the "last-mile" labor of bending general-purpose technology to one company's specific business has never disappeared. It just keeps changing workstations and changing names.

Format
Essay / Industry History
Author
Odin Wang
Published
2026 · 06
Verified
Adversarial 22/3 + targeted
— The Argument in Brief

Integration labor never disappears. It only migrates between four workstations, resurfacing under a new job title every decade. The FDE is just the 2024 costume.

01

Four workstations

Last-mile integration labor either gets bundled free into the vendor's price, outsourced to channels and systems integrators, baked into the product, or pushed onto the buyer's own IT. What decides where it lives is product maturity and gross margin.

02

Renamed, revived

Roughly every decade a new model declares it will kill this labor (boxed software, SaaS's "No Software", low-code, AI agents). A few years later it returns under a fresh job title. The job title is the fossil record.

03

The SaaS mirror

Salesforce promised "no software, no consultants," then bled 63% of customers a year and was forced to invent Customer Success. The FDE is the Customer Success of the AI era: it won't vanish, it'll be renamed and routed to partners.

§I
The Machine

First, the relocation machine

Whenever a new technology first goes to market, the product is half-finished. It runs in a demo, but drop it into a real company's dirty data, legacy systems, and undocumented workflows and it falls over. Getting it from "runs" to "actually usable" takes a person who understands both the business and the technology, sitting next to the customer. That is integration labor.

That labor's cost is always there. The only question is whose ledger it lands on. There are four ways to book it, and these four workstations are the entire key to the next sixty years.

Bundled free into the vendor's price

The vendor pays for it and throws the engineer in as a freebie. The customer never sees the labor bill; it's hidden inside the price of the hardware or software.

Trigger: margin fat enough to carry people

Outsourced to channels and SIs

Handed to a ring of partners: value-added resellers, systems integrators, implementation consultants. The vendor just builds the product; the integration grunt-work goes to the channel.

Trigger: product standardized, vendor can't touch every deal

Baked into the product itself

The labor gets understood and written into product features: config options, wizards, templates, today called Agents. The workstation everyone wants to reach, and the hardest to stay parked in.

Trigger: patterns settled enough that code replaces people

Pushed onto the buyer's own IT

The customer staffs it themselves: the corporate MIS department, today's "citizen developers." The labor doesn't vanish; it becomes a fuzzy line item on the buyer's books.

Trigger: customer forced to do it in-house

The pattern: the earlier the technology, the more the labor crowds into ; the more mature, the more it sinks toward ; and when it can't sink, it bounces back and forth between and . The chart below is the path that labor walked over sixty years. Watch the line lunge twice toward "③ baked into the product," and twice get bounced right back.

Fig · 01 Sixty years of integration labor on the move 1960 — 2024
① In the price ② To the channel ③ Into product ④ Onto buyer 1960s 1970s 1980s 1990s 1999 2006 2010s 2024 1969 unbundling ↓ Systems Engineer VAR Boxed softwareescapes to internal IT Systems Integrator Customer Success Solutions Architect Low-code FDE
Same integration labor, where it lands The 2024 costume Dashed ③ = the workstation everyone wants, no one stays in
§II
Sixty Years, Eight Costumes

A chronicle: same job, eight relocations

Put this machine on a timeline and it runs eight full laps. At every stop the integration labor switches workstations and changes its name, but the work itself stays the same.

1960s
In the price

IBM threw the engineer in for free

In the mainframe era IBM didn't sell machines, it sold the whole capability of getting one running inside your business. IBM, Univac, Burroughs and RCA all bundled software, systems engineering and training into the hardware price with nothing priced separately. The person doing the work was the Systems Engineer (SE), paired one-to-one with a salesperson. IBM even pulled people fresh off the Apollo 11 programming team to deploy on-site. Before 1969, these services were "provided free of charge at the discretion of the IBM Branch office"[IBM history]. Free, hidden in the hardware price. That is the prehistoric form of the FDE.

5–10%The machine's direct cost was only 5–10% of the sale price; the rest was recovered through services and software, which were treated as a "marketing cost" and given away.[oral history, 3-0]
1970s
To the channel

Minicomputers kicked the labor out to the channel

In 1969 an antitrust threat forced IBM to "unbundle" and price systems engineering separately (regulation was a key trigger, not the only one), and the SE's low-level programming work flowed to cheaper third parties. Then DEC and Data General minicomputers dropped the price floor; vendors couldn't afford on-site engineers per customer, so they handed integration to OEMs and VARs. When Novell shipped its LAN in 1983, installing one was beyond ordinary users, and dealers became "network integrators" overnight.

5% → 50%A dealer's services revenue (consulting, install, maintenance) grew from 5% to 20%, 30%, even 50% of the business. Boxes stopped paying; integration did.[industry history]
1980s
Product + Buyer

Boxed software: first lunge at "③," and the labor fled

Lotus 1-2-3, WordPerfect and dBASE compressed software into a retail box, "shrinkware." This was workstation ③ in the extreme: work that used to need a programmer, pressed into a mass-replicable commodity with near-zero marginal cost. But the cheaper the box, the more orphaned the integration work around it, so it fled two ways: one to the channel (reseller Pomeroy's services were 12% of revenue but over half its profit), one pushed onto the buyer's own internal IT, a brand-new destination.

70%Gartner: 70% of a PC's total cost of ownership is labor (install, networking, training, firefighting). The purchase price barely counts.[Gartner, cite as a range]
1990s
To the channel

Client/server grew workstation ② into a giant

Client/server was heterogeneous, multi-vendor, networked. An IEEE paper called it "a sisyphean task", with total cost of ownership 3 to 6 times a mainframe's. Complexity exploded, and integration labor swelled into a giant industry: systems integration was about $14.9B in 1990 across 1,600 dedicated firms, reaching roughly $87.8B by 2000. Andersen Consulting renamed itself Accenture in 2001; EDS signed a ten-year, $3.2B Xerox outsourcing deal.

40%+IBM's non-maintenance services were under $6B in 1991; a decade later IT services were 40%+ of an $86B top line, its single largest source, with 4 of every 10 employees in Global Services. A hardware company eaten alive by integration labor.[IBM official history]
1999
Back to the price

SaaS's "No Software": the first formal vow to kill implementation

Salesforce launched in 1999 with the No Software campaign, the word "software" in a red circle with a line through it. Benioff's own framing of the problem: "without software or hardware to install, and pricy consultants to hire"[official]. "No consultants" was written into the promise. That is the exact script of today's AI agent: "you don't manage it, it just runs." And the company that promised turnkey nearly died of churn, then was forced to staff up a small army it deliberately named Customer Success, not customer support.

63%In 2005 Salesforce's churn was ~8% a month, ~63% a year, nearly every customer leaving within a year. "No Software" didn't kill implementation labor; it just renamed it Customer Success.[Inc. + CS Association, two sources]
2006+
In the price

The AWS Solutions Architect: the FDE's direct ancestor

Once AWS commoditized infrastructure, all three workstations came online at once. AWS states in plain text that the Solutions Architect is "a free resource" (workstation ①), owning the customer's overall technical relationship, acting as trusted advisor, working "as an extension of our customers"[AWS official]. Alongside sit paid Professional Services (② in-house) and the APN partner ecosystem (② outsourced). This is exactly Palantir's FDE "embedding with the customer," just without the "forward deployed" branding. A caveat: SA and FDE are not the same role, the SA leans pre-sales and design, the FDE comes later and writes production code, they're variants of the same labor at different points in the sale.

200+AWS has 200+ services, so many that customers don't know where to start, and "this is where AWS Solutions Architects can help." General capability, configured into a specific solution. Same last mile.[AWS official]
2010s
Productized again

Low-code: second lunge at "③," and the same irony

Low-code platforms (OutSystems, Mendix, Microsoft Power Platform) pitched it bluntly: "eliminate ad-hoc development," "watch your backlog disappear," "no additional headcount." Another run at workstation ③, productizing the integration labor into a tool. And the same irony followed: low-code spawned an entire implementation-partner ecosystem, complete with certified "low-code functional consultants," and "low-code implementation services" became a category with its own analyst rankings. The labor didn't vanish; it relocated to "certified low-code consultant."

2024
The latest costume

The FDE: sixty years on, back to workstation ①

The modern FDE traces to Palantir's Forward Deployed Software Engineer (codename Delta), embedding to configure the existing Foundry/Gotham platforms, structurally identical to the 1960s IBM SE. Palantir even demonstrates the ③ feedback loop: a field configuration that turns out useful for others gets generalized back into the base platform. The 2024–2026 receipts: OpenAI's FDE team was founded only in January 2025 and grew from 2 people to 52 by year-end; Anthropic is hiring FDEs right now, with one French-region posting listing a base of €205,000–220,000 (the only employer-published FDE band, the seven-figure numbers floating around are content-marketing blogs, ignore them). The spark was a16z's essay "Trading Margin for Moat," which lands squarely on the "fat margin" variable.

2 → 52OpenAI's FDE team scaled from 2 to 52 in 2025, per the team lead's own public figures.[official / first-party]
§III
The Fossil Record

The fossil record: line up the job titles

Lay the sixty years of job titles in a row and you see the strata of one job. The sales engineer is the clearest seam: the role has drifted under new names since IBM coined it in the 1960s. As one veteran put it, "the name kept changing, but the nature of the position itself remained essentially the same."

1960s
Systems Engineer
1970s
VAR / network integrator
1990s
Systems Integration consultant
2005
Customer Success Manager
2006+
Solutions Architect
2010s
Low-code consultant
2024
Forward Deployed Engineer

The FDE is, for all intents and purposes, the old solutions engineer / technical consultant in a fresh uniform. The tasks haven't changed: embed with clients, fix their broken systems, keep contracts alive through hand-holding. Same fundamental work, different wrapper. What changed was the story.

Tom White · "Sexy Titles, Unsexy Work" · commentator

Every time someone declares this labor dead, it returns a few years later under a new job title. The job title is the fingerprint of the decade.

§IV
The Mirror

The SaaS mirror: the best predictor of the FDE's fate

SaaS deserves its own look, because the resemblance to today is uncanny. A model that vowed to kill implementation labor ended up merely renaming it Customer Success, then growing an implementation ecosystem six times its own size.

SaaS (1999)
AI Agent (2024)
The promise
No Software — no install, no consultants
You don't manage it, the agent just runs
The reality
Customers can't get value, 63% annual churn
95% of enterprise AI pilots show zero P&L impact [MIT NANDA]
The role it forced
Customer Success + implementation partner ecosystem
Forward Deployed Engineer + GSI partners
Who takes the standardized part
Accenture / Deloitte doing Salesforce implementations
Accenture × OpenAI, Cognizant × Anthropic
Follow the mirror: the FDE is the Customer Success of the AI era. It won't disappear; it'll be renamed and routed to a partner ecosystem. The GSIs are already signing on, which is integration labor migrating from workstation ① to ②, the same move as the SE being unbundled and handed to professional-services firms decades ago.
§V
The Honest Counterargument

The honest counterargument: "this time is different" deserves a hearing

You can't only collect evidence that flatters the thesis. The strongest counterargument has weight, so here it is, with each camp's best advocate.

Permanent-role camp · the FDE stays
This time really is different
Aaron Levie · Box CEO
Deploying an AI agent is far more complex than deploying traditional software; the professional-services opportunity is as big as the cloud-migration wave was. The FDE is a structural role, not a stopgap.
Black Matter VC
Every model leap just moves the bottleneck downstream; when generation gets cheap, reviewing the output becomes the expensive part. Disband the FDEs and you don't get an autonomous AI, you get a pile of confident, unowned, wrong output.
Recession camp · supports this essay
The old script will replay
Ethan Mollick · Wharton
"You'll know the AI labs really believe in ASI the day they disband the FDE teams they just built." Standing up a human consulting arm exposes that they don't quite trust the model to finish the job.
f-prime / Flybridge
If 30–40%+ of deployments need heavy FDE effort, the problem isn't GTM, it's product design; the FDE should be "scaffolding, not architecture," time-boxed to 3–6 months. If the work doesn't productize, the economics collapse.

My read: the recession camp is more convincing, Mollick's line is nearly decisive, and the SaaS mirror has already played out the ending once. But one timing point matters: for the next 12 months both camps see the same picture, labs expanding FDEs, services revenue compounding. The real split is whether the labs start disbanding these teams in 2027. The FDE will recede, but don't pull the date too close, this is a 2027 question, not a 2026 one.

§VI
A Ruler

The close: a ruler to test the claim

Connect the sixty years and the FDE's position is clear: it's what happens when an immature technology (large models) early in its life, pushed by fat margins and a long last mile, sucks integration labor back inside the vendor (workstation ①), the same move as IBM bundling the SE into its price. History also lets you predict: as models mature and patterns get baked into the product (③), and as GSIs take over standardized deployment (②), the FDE role will shrink back.

So the next time someone waxes lyrical about the FDE's "unique delivery philosophy," measure it with one question.

Your last mile, three years from now: does it settle into the product, or does it forever need a pile of people?

A ruler · The last-mile question

Those who can answer are genuinely building the slide into workstation ③. Those who can't, or who argue the opposite (that integration in the AI era can never be productized), either truly believe this time is different (then bring a Black Matter–grade argument), or they've dressed up "the product isn't mature yet" as "we have a unique delivery philosophy." This shadow has followed computing for sixty years. Every time it says it's leaving, it has only changed clothes.