The State of AI-Native Jobs

115,348 jobs were open at the 14,849 companies in our index when we froze the data for this report on 2026-07-02. 39,882 of them, 34.6%, sit at the 5,849 companies our classifier marks AI-native: labs, AI-first startups, and the infrastructure and applied-AI companies built around them. Every number in this report is computed from those live postings. Nothing here is a survey, and nobody self-reported anything.

The wider market makes that snapshot worth taking. Indeed's Hiring Lab measured AI-mentioning postings at 134% above their pre-pandemic baseline at the close of 2025, while total US postings sat just 6% above the same line. Hiring in general is flat. Hiring around AI is not, and the gap between those two lines is where this report lives.

We wrote it because the usual instruments read this market badly. Salary surveys lag it by a year. Title taxonomies miss jobs that didn't exist in 2024. And the loudest claims, that AI kills jobs, that every engineer now earns a million dollars, dissolve on contact with posting-level data. What follows is what 115,348 open roles actually say: who AI companies hire, what they pay, which skills carry a premium, and where the openings physically are.

One reading note before the numbers. Our index holds live postings only, so it photographs the market's present shape rather than its growth curve. Where a section needs a growth claim we cite external trackers and say so. Everything else is ours, and you can check the method at the end.

Start with the number that surprises almost everyone. At AI-native companies, AI/ML engineering accounts for 5.4% of open roles: 2,139 postings. Sales and go-to-market alone is 14.5%, nearly three times bigger. Add support and marketing and the customer-facing block reaches 9,692 postings. The companies building the models are, by posting volume, sales organizations with a research department attached.

This is not a data quirk. It matches what the labs say about themselves. OpenAI grew its enterprise sales team from fewer than 10 people to 500 in two years, and the Financial Times reported in March 2026 that its planned doubling to 8,000 employees is weighted toward go-to-market more than research. Ben Horowitz put the reason plainly: "Right now with OpenAI and Anthropic, everybody wants to buy AI. They're already predisposed to buy." Someone still has to run the deal, the pilot, the security review, and the renewal.

Software engineering is still the biggest single family at 21.4% of AI-native postings. But the gap between engineering-as-posted and engineering-as-imagined matters for anyone planning a move into this market. For every open AI/ML engineering seat there are roughly four customer-facing ones, and those roles increasingly demand real technical depth. We come back to that when we get to the forward-deployed engineer.

I think this is the most useful reframe in the whole dataset. The AI hiring boom is not a research hiring boom. It is a deployment and distribution hiring boom, and the surface area for getting into it is far wider than a model-training resume.

On June 30, 2026, AWS announced a $1 billion investment in forward-deployed engineers: thousands of engineers embedded inside customer organizations to ship agentic AI systems. Two days later we counted what our own index says about the role. There are 5,528 open forward-deployed and solutions-engineering postings across our companies, 3,077 of them at AI-native firms. At those firms the family now makes up 7.7% of all open roles, against 3.2% everywhere else, and it has overtaken AI/ML engineering (2,139 postings) outright.

The role is older than the hype. Palantir formalized it in the early 2010s: an engineer who writes production code inside the customer's environment and owns the outcome, not the demo. What changed is that every AI company selling to enterprises hit the same wall Palantir built that role to climb. Enterprises do not buy model capability. They buy working systems inside their own messy data, and someone has to go build those systems where they live. OpenAI stood up its forward-deployed team in 2024; Anthropic's Applied AI group runs the same play; AWS just industrialized it.

Pay has two distinct tiers, and conflating them produces most of the bad takes. Across the 1,076 FDE postings in our index that disclose a band, the median lands at $180k, with the 90th percentile at $256k. That agrees with Bloomberry's independent analysis of 1,000 FDE postings (median $173,816) from a sample a fifth the size of ours. The frontier-lab ladder is a different animal: Perspective AI's report on 1,200 FDE data points puts mid-level lab FDEs at $385k total comp and principals above $1.2M, with equity making up 60 to 70% of the package. Market-rate FDE work pays like strong solutions engineering. Lab FDE work pays like frontier research.

Who is hiring tells its own story. Databricks leads our index with 281 open forward-deployed and solutions roles, ahead of Palantir, OpenAI, Mistral, ElevenLabs, and LangChain. Note the spread: a data platform, a defense-adjacent veteran, two labs, a voice AI company, and an agent-framework startup. When a role shows up across that many business models at once, it has stopped being a fashion and started being infrastructure.

One honest caveat from our own data. Within 2026 the FDE share of monthly postings has held roughly flat, so the explosive-growth framing comes from the outside world (the AWS announcement, BCG naming it an emerging role of the AI era), not from our curve. What our data shows is scale: the role is already one of the largest technical families in AI hiring, and it was barely a category three years ago.

For every junior opening at an AI-native company there are 20.8 senior openings. Not four, not six: 20.8. Junior and intern roles together make up 4.6% of the AI-native postings that carry a level. The rest of our index runs top-heavy too (12.9 seniors per junior), but AI-native companies have pushed the ladder shape further than anyone.

Our snapshot lands inside a pile of converging external evidence. Stanford's Digital Economy Lab, working from ADP payroll records covering 25 million workers, measured a 16% relative employment decline for 22-to-25-year-olds in AI-exposed occupations since late 2022, with young software developers down nearly 20% from peak. SignalFire found new grads collapsed to 7% of Big Tech hires, half the 2019 share. The St. Louis Fed clocked recent-grad unemployment at 4.59% against 3.25% in 2019, concentrated precisely on credentialed young workers.

Here is the twist our data adds. The junior jobs that survive are not junior jobs in any traditional sense. Among AI-native junior and intern postings, 40% ask for "ownership" or end-to-end responsibility (against 29.8% elsewhere), 26.3% mention system design or architecture, and 20.3% put the junior in front of customers or stakeholders. PwC's 2026 Jobs Barometer called this pattern seniorization: AI-exposed entry-level work grew, but it now demands what used to be senior skills. Our postings confirm it line by line. The entry-level job didn't vanish so much as mutate into a small senior job at a junior price.

The labs themselves quietly hedge the strategy. Anthropic runs a Fellows Program that pays $3,850 a week and converts more than 40% of its cohort to full-time hires, which makes it one of the highest-yield junior pipelines ever disclosed in tech, operating at exactly the moment its industry stopped posting junior roles. Whatever companies say about not needing juniors, the ones with the most money are still growing their own.

Where do the 1,011 labeled AI-native junior and intern openings actually sit? Hardware and robotics leads, then software engineering, and forward-deployed work carries 149 of them. If you are early-career and aiming at this market, those are the doors that still open.

20,870 of the postings in our index disclose a salary band, and we compute everything in this section from those bands: the midpoint of each posted range, with hourly and implausible figures rejected. The headline: the median posted package at an AI-native company is $186k, against $165k at everyone else. A 13% premium for doing the same broad job at a company whose product is the model rather than around it.

The premium is real but it is not evenly spread, and where it concentrates is the finding. At mid level, AI-native companies pay 21% over the rest of the market ($173k vs $143k). By staff and principal the gap compresses to single digits. That is the opposite shape from the skill premium Levels.fyi measures, which widens with seniority. Our read: senior comp is anchored by a market-wide rate card, but AI-native companies are in an open bidding war for the mid-level engineers who do the volume of shipping. If you are three to six years in, this market is bidding for you specifically.

By function, the premium hides in unexpected places. AI/ML engineering at an AI-native company posts a $233k median, but the largest relative premium in our data belongs to customer support and success at 35%. Support at an AI company increasingly means debugging model behavior with technical customers, and it is priced accordingly. Forward-deployed work, meanwhile, pays almost identically inside and outside the AI-native world: it is a market-rate craft wherever you practice it.

For calibration against the outside world: Lightcast measures a 28% posting-level premium for AI skills across 1.3 billion postings, and PwC measures a 62% worker-level wage premium globally. Those numbers differ from ours because they answer different questions (a skill premium and a worker premium, versus our company-type premium). All three point the same direction. And yes, the $100M+ Meta packages are real but confined to a few dozen researchers; they are compensation news, not compensation data.

Read the descriptions instead of the titles and a new stack appears. Agent and agentic-systems language now shows up in 20% of engineering postings in our index (5,806 roles), and those postings pay a 10% higher median than engineering roles that skip it. LLM experience appears in 18.7%. Then the specialist layer: inference and serving (7.2% of engineering postings, and the biggest pay bump in our data at +15%), evals (6.1%, +9%), fine-tuning (4%, +13%), retrieval (4.9%).

The fastest mover is the newest. MCP, the protocol that lets agents call tools, went from 1.2% of engineering postings in January 2026 to 4% by late spring in our index. A tripling inside five months, and one of the few growth claims our data can make cleanly since it is a share, not a count. Stanford's AI Index measured the same explosion one layer up: agentic AI skill mentions grew 280% in a year across the wider US market.

Evals deserves its own paragraph because employers keep telling us, through their postings, that measurement is the bottleneck. Anthropic listed a Research Engineer for model evaluations at $320,000 to $485,000 this May, research-scientist money for the job of proving whether the model works. In our index the evals premium is more modest at +9%, because the skill is diffusing into ordinary engineering roles rather than forming a separate priesthood. Either way the direction is set. When output gets cheap, verification gets expensive.

What faded is just as instructive. Prompt engineering appears in 3.2% of engineering postings as a mentioned skill, but as a job title it has all but disappeared, which the next-but-one section counts precisely. The market stopped paying for clever prompting on its own and started paying for the machinery around it: context, tools, retrieval, evaluation.

Group our AI-native postings by funding stage and you can watch a company change shape in real time. At seed, 41.4% of open roles are software engineering; sales and GTM is 8.5%. By growth stage, engineering has fallen to 17.9% of postings while GTM has doubled its share to 15.5%. Forward-deployed roles barely exist at seed (4.1%) and more than double by the time companies have enterprise customers (9.7%). Companies are born as engineering organizations and hire themselves into distribution organizations.

Overlay the outside evidence and three distinct company shapes emerge. The first is the lean AI-native startup: Lovable reached $400M in annualized revenue with 146 employees, and Forbes counts the cohort's revenue per employee at $2M to $4M against a $300k SaaS average. These companies skip sales teams entirely and stay engineering-dense for as long as product-led growth carries them. Wabi's Eugenia Kuyda runs the thesis at its extreme: 10 to 15 people, no juniors, ever.

The second shape is the frontier-lab commercializer, and it hires in the opposite direction. OpenAI's planned 4,500-to-8,000 doubling is GTM-heavy. Anthropic went from roughly 400 people in 2023 to about 5,000 by mid-2026 while targeting $20B+ in revenue. These organizations pay the equity-heavy ladders from section 3 and run the quiet junior pipelines from section 4. The third shape is the enterprise FDE hub, the AWS and Palantir model: thousands of engineers who live inside customer accounts. Our stage data suggests every venture-backed AI company migrates from shape one toward shape two or three the moment enterprise revenue arrives.

One number worth sitting with: growth-stage AI-native companies in our index average 34.3 open roles each, across 401 companies. The lean-team story and the hiring-machine story are both true. They are just different companies, at different stages, and the stage data says which one a candidate is actually walking into.

The most repeated fact about AI-industry work arrangements is that the labs killed remote. It is true and misleading at the same time. True: OpenAI lists just 4% of its 721 open roles as fully remote, Google DeepMind badge-tracks a three-day office week, and Sam Altman called remote work a mistake. Misleading: across all 27,322 AI-native postings in our index with a stated work mode, 29.7% are fully remote, slightly more than the 25.2% at everyone else. The famous labs are the outlier, not the market.

What the market actually did is split the work by what it is. Human-data roles, the AI trainers and annotation specialists, run 80.2% remote in our index because the work is judgment applied through a browser. Evals-titled roles run 26.3% remote. Forward-deployed roles are 70.2% hybrid or on-site for the obvious reason that the customer's building is the job. Remote didn't retreat evenly. It retreated from the roles where physical presence is the product and pooled in the roles where it isn't.

Geography concentrated the on-site half hard. San Francisco carries 4,305 AI-native postings in our index, about as many as the next two metros combined; New York and London follow, with Paris and Bengaluru holding the largest non-Anglosphere pools. If your plan involves a badge and a desk, the map matters again in a way it hasn't since 2019.

Job titles are where a new economy shows its handwriting, so we counted them across all 115,348 postings. "Member of technical staff", the deliberately flat title the labs adopted, appears 433 times at a $275k median, led by Perplexity. Research engineer runs 321 postings at $275k, Anthropic first among hirers. Agent-titled engineering roles number 1,046. Forward-deployed appears in 860 titles. Applied AI, the labs' word for engineers who make models useful, carries 312.

Now the obituary. "Prompt engineer", the title that launched a thousand thinkpieces and six-figure salary screenshots in 2023, appears in exactly 4 of our 115,348 postings. Its heir apparent, the context engineer, musters 2. The skill did not die; prompt-craft moved inside agent, evals, and product roles. The standalone job died. Karpathy's line that context engineering is the real discipline got the direction right, but even that has produced titles slower than tweets.

One more title family deserves a flag: human-data roles (AI trainers, annotation specialists, RLHF operations) show 196 postings in our index, and that badly undercounts a market that mostly lives off job boards. Outlier alone claims 700,000+ onboarded experts and over $500M paid out; Mercor's specialist rates run $100 to $200+ an hour. It is the biggest hiring story that never shows up in postings data, which is exactly why we flag it here.

A note on the "evals engineer" you keep reading about: as a title it is still rare (147 postings, and most title matches are old-school defense test-and-evaluation roles). Evals is a skill spreading through descriptions, not yet a titled profession. Watch this table next year.

Everything above describes the demand side. The supply side has its own AI story, and it is an arms race. Greenhouse's CEO says the average recruiter now receives about 400% more applications than a few years ago, most of them AI-polished into sameness. On the other side of the table, 70% of hiring managers trust AI to screen faster, while 8% of job seekers call AI hiring fair. Both sides automated, and the signal drowned.

The countermeasures are physical. Google, Cisco, and McKinsey brought back in-person interviews specifically to counter AI-assisted cheating, and interview-assistant tools claiming 10 million users explain why. This is Goodhart's law eating the resume: once a signal can be manufactured at zero cost, employers stop reading it and go looking for signals that can't be faked at scale.

Put this section next to section 4 and the market's message to candidates becomes coherent. Companies flooded with synthetic-looking applications respond by hiring people who are already proven, which is precisely what a 21-to-1 senior-to-junior ratio looks like from the inside. The scarce commodity in this market is not the application. It is verifiable evidence of capability: shipped work, measurable outcomes, and skills like evals and agent-building that show up in what you have built rather than what you claim.

The data suggests a sharper playbook than "learn AI". First, aim where volume meets premium. Agent-adjacent engineering is 20% of engineering postings and pays 10% over baseline; inference and serving is smaller but carries the fattest premium at +15%. Pairing one high-volume family (software engineering, forward-deployed) with one scarce skill (evals, inference, fine-tuning) beats optimizing either alone, because the first gets you interviews and the second wins them.

Second, take the forward-deployed path seriously even if you think of yourself as a product engineer. It is the largest technical family AI-native companies hire for after core software engineering, it pays a $180k median with a fast ladder into the lab tiers, and it teaches the one skill this report keeps finding scarce: making models work inside real organizations. It is also, quietly, one of the few families still posting junior roles (149 in our index).

Third, if you are early-career, stop reading the ratio as a verdict. 1,011 junior and intern roles are open at AI-native companies right now, the lab fellowships convert at rates worth chasing (Anthropic's above 40%), and expert human-data work pays professional rates while you build a portfolio. The junior jobs that exist expect ownership and customer contact on day one. That is unfair and it is also the arbitrage: demonstrate those two things and you are competing against very few people.

Last, an opinion rather than a statistic. What I would tell a candidate this month: pick one system you can build end to end with an agent framework, add an eval harness that proves it works, and put the numbers in front of hiring managers. In a market drowning in AI-generated applications, a measured artifact is worth a hundred tailored resumes.

A report like this should make claims that can be wrong. Here are ours, checkable against this same index in twelve months. One: forward-deployed and solutions roles overtake AI/ML engineering across the whole index, not just at AI-native companies (today: 5,528 vs 3,735). Two: MCP mentions pass retrieval mentions in engineering postings (today 2.7% vs 4.9%, and closing fast). Three: the junior share of AI-native postings stays below 5% unless fellowship-style programs scale by an order of magnitude. Four: the mid-level premium compresses as the rest of the market reprices, the way senior comp already has.

The wider forecasts frame the stakes. BCG expects 50 to 55% of US jobs reshaped within three years and names forward-deployed engineers among the roles the reshaping creates. PwC's Barometer shows the two-track market widening: jobs AI professionalizes are growing twice as fast, with 42% faster wage growth, than jobs it commoditizes. And the first wave of AI-layoff reversals (employers quietly rehiring after automation under-delivered) hints that the reallocation story will stay messier than the press releases.

We will re-run every number in this report as the index refreshes, and the page you are reading updates in place. If one of our four predictions dies, it dies in public, which is the way data journalism should work.

Scope: 115,348 live postings at 14,849 companies, scraped continuously from applicant-tracking systems and careers pages, snapshotted 2026-07-02. A company counts as AI-native when our classifier flags it so, based on what it builds and sells; the flag covers 5,849 companies from frontier labs to applied-AI startups, and includes a small set of AI-forward incumbents (Databricks, Salesforce, NVIDIA) whose hiring is dominated by AI products. We disclose this because "AI-native" has no industry-standard definition, and ours is deliberately broad at the commercial end.

Compensation: computed from the 20,870 postings that disclose a salary band, taking each band's midpoint, rejecting hourly, weekly, and implausible figures. Posted bands are US-weighted because US pay-transparency laws drive disclosure. Seniority and work-mode shares are computed over postings that carry those labels (roughly half and two-thirds of the index respectively) with counts shown throughout.

Trends: our index holds active postings only, so absolute counts cannot support growth claims; when the report says something grew, the claim is either a within-2026 share of monthly postings or an externally cited figure. All external statistics link to their sources inline and in the list below. Numbers refresh weekly with the index; prose is updated by humans. If you cite this report, cite it as "Landed, The State of AI-Native Jobs, 2026" and link this page, and if you want a cut of the data we didn't publish, email us. We answer.