How AI careers moved closer to outcomes

Each wave moved technical work nearer to real-world results.

2012–2016

Data scientist

Find patterns and build predictive models

Model

2017–2020

ML engineer

Productionize and scale models

Model

2020–2022

AI engineer

Build AI-powered applications

Application

2023

Prompt engineer

Direct generative models effectively

Application

2024–2025

GenAI / LLM engineer

Build RAG, copilots, and multimodal systems

Application

2025–2026

Agentic AI engineer

Build AI that plans, uses tools, and acts

Agent

2026

Forward deployed engineerRapid growth

Make AI work inside real customer environments

Outcome

Market signals

55→10,000

Job postings mentioning GenAI skills (nearly 10,000)

Jan 2021 → May 2025

643→5,330+729%

Forward deployed engineer job postings

Apr 2025 → Apr 2026

The direction is clear

ModelApplicationAgentOutcome

Roles overlap. Dates show periods of market prominence, not when each role was invented.

Source

Lightcast GenAI job-market analysis FDE job-posting growth LinkedIn 2020 Emerging Jobs Report

When AI writes the code, what is the data scientist’s job?

The shift is already happening.

25%→75%

of Google’s new code generated by AI

2024 → 2026

57%

of data scientist postings seek people working beyond core data science

73%

of surveyed open-source developers already used AI coding tools

2024

What changed?

AI increasingly produces

Data scientists increasingly own

SQL, pandas, and transformations

Problem framing and data meaning

Models, charts, and prototypes

Evaluation and trustworthy decisions

Code, APIs, and documentation

Integration, adoption, and impact

Less time producing code. More responsibility for making AI work.

The fundamentals still matter: they’re how you review and direct what AI produces.

Source

Google AI-generated code Data scientist posting analysis GitHub Octoverse

The model worked. The solution failed.

A familiar story.

95% accuracy

model tested

Deployed

in production

Users avoid it

low usage

Workflow unchanged

old process stays

0

Business impact

Would you call this project successful?

Production is not the finish line. Impact is.

What went wrong?

Wrong workflow fit Low user trust No adoption ownership

FDE mindset

Model quality×Workflow fit×User adoption=Real-world value

The model did exactly what it was designed to do. Nobody owned what happened after deployment.

Where the FDE sits changes everything

Two roles, side by side. Two centers of gravity.

Product team

Data scientist

Reusable capability

Customer mission

FDE

Operational outcome

Product-based data scientist

Forward deployed engineer

Embedded with

The product / data team

The customer’s mission team

Builds

Reusable capabilities

Adaptations for one environment

Feedback

Through product channels

Directly from users and constraints

Optimizes

The product or model

The customer’s operational outcome

Same technical foundation. Different proximity and mandate.

Boundaries vary by company. The distinction is the role’s center of gravity.

You are the FDE. What do you do first?

A restaurant owner says

“Build an AI voice agent to answer calls and take food orders.”

What happens today?

Customer calls

Staff pauses their work

Staff takes the order

Order enters the kitchen

Interview the owner, phone staff, and kitchen team. Look for:

When and why are calls missed? Which menu requests confuse staff? How are prices, changes, and pickup times confirmed? How does the order reach the kitchen? When must a human take over?

Your first move?

The requested technology is only the starting hypothesis. The workflow reveals the real solution.

Where data science works inside a voice agent

Click any step to zoom in.