AI Careers in 2026: Roles, Salaries, and How to Get Hired

The starkest data point in technology in 2026 is that one labour market has detached from the others. Through 2024 and 2025, headcount at Meta, Amazon, and Google contracted while AI-specific roles at the same companies grew. By Q1 2026, AI engineering compensation was running 35-50% above equivalent-tenure software engineering, and the gap was widening, not closing. Anthropic alone has hired roughly 450 net employees since the start of 2024. OpenAI has crossed 3,000. Microsoft AI's rebadged hiring drive added 700 in twelve months. Ireland and the UK, once peripheral to AI hiring, are now serious centres. If you are deciding whether to pivot into AI work, the relevant question is not whether the jobs exist. They do. The questions are which role fits your background, where the salary actually goes when you negotiate, and which doors are still narrow enough to be worth banging on.

Table of contents

• Why AI hiring is up while the rest of tech is down
• The role taxonomy: who actually does what
• ML engineer / AI engineer / LLM engineer
• Research scientist (and why most people should not aim here)
• AI product manager
• AI governance, policy, and trust and safety
• Applied AI roles in non-tech companies
• Salary data: real numbers from real offers
• Hiring funnels at OpenAI, DeepMind, Anthropic, and Microsoft AI
• How to break in mid-career
• Frequently asked questions
• The bottom line

Why AI hiring is up while the rest of tech is down

The 2023-24 layoff cycle was the cleanest test of where each company actually placed its bets. Meta cut 21,000 roles. Amazon cut 27,000. Google cut roughly 12,000 across 2023 and 2024. Almost none of those cuts touched AI infrastructure, applied research, or the model-training organisations. The same period saw Anthropic grow from around 150 to roughly 600 employees, OpenAI's research org alone double, and DeepMind absorb the entire Google Brain team into a single unified structure. The signal was unambiguous: AI labour was being paid for out of a different budget line, and that line kept growing.

The capital flows justified the hiring. Anthropic raised at a $61B post-money in late 2025; OpenAI's secondaries traded at over $300B by the same quarter. Microsoft's $80B AI infrastructure spend over four years was announced in January 2025. Amazon's commitment to Anthropic alone reached $4B. Google's TPU buildout absorbed another $50B+ in capex. xAI raised at $50B in November 2025. When a single industry absorbs that volume of capital across that many balance sheets, the binding constraint becomes people, not money, and wages adjust accordingly.

The role mix changed. In 2022, "AI engineer" was barely a real title; most postings used "ML engineer" or "applied scientist". By late 2025, "AI engineer" was the fastest-growing job title on Indeed, LinkedIn, and Greenhouse. The new role's premise: someone who composes existing model APIs into production systems and owns prompts, retrieval, agents, and evals, but does not necessarily train models from scratch. It is a different job from ML engineering and it pays better in 2026. We dig into that comparison in our ML engineer vs data scientist vs AI engineer guide.

The UK and Ireland matter more than most US-based candidates realise. Anthropic opened its London office in late 2024 and has hired over 80 people there by mid-2026. DeepMind's London headquarters remained the centre of gravity for one of the four major frontier labs throughout the consolidation. Microsoft AI's UK office picked up much of the Inflection acquisition cohort. Stripe, Wise, and Revolut have built sizeable AI engineering teams in London and Dublin. Compensation in those metros runs at roughly 70-80% of San Francisco numbers, which, given the lower cost of living, often means similar or better real terms.

There is a caveat. Most of the salary peaks are concentrated in five metros: San Francisco, New York, London, Zurich, and Tel Aviv. Outside those, AI compensation drops fast. Real remote AI jobs exist (see our remote AI jobs guide) but they pay closer to remote SWE rates than to on-site frontier-lab rates. The premium is partly the work, partly the location.

The other caveat: this is a narrow market that hires aggressively at the top of the funnel. OpenAI receives roughly 10,000 applications per quarter for fewer than 200 open roles. DeepMind's research scientist openings receive over 40 applications per slot. The market is paying well, but the gates are still gates. A naive application strategy aimed only at the four major labs is the single most common reason capable candidates fail to land an AI job in 2026; we discuss the funnel in our job-hunt playbook.

The role taxonomy: who actually does what

The cleanest cut for a 2026 reader is research vs applied vs governance. Most candidates spend weeks confused because job titles are inconsistent across employers; once you can read a job description through that taxonomy, the choices become clearer.

Research roles extend what models can do. Outputs are papers, novel architectures, capability releases. Most research roles sit at frontier labs (OpenAI, Anthropic, DeepMind, Microsoft Research, Meta FAIR) or top university labs. The headcount is small. Per-lab, "research scientist" populations in the low hundreds is typical even at the largest labs. A typical week for a research scientist: two-thirds time on a single experimental track (training runs, eval design, ablations), one-third on the broader research community (paper review, internal seminars, paper writing). The cycle from idea to publishable result is rarely under three months.

Applied roles use existing models to ship product. This is where the bulk of new headcount sits. The titles vary: AI engineer, applied AI engineer, ML engineer (applied), staff engineer with an AI specialty. The work is similar across them: prompts, retrieval, agents, evals, latency, and cost. Every Series B startup with an AI product is hiring for this role, plus all the FAANG-scale companies, plus banks, consultancies, healthcare systems, and increasingly governments. A typical week for an applied AI engineer: feature work on a current product surface, an eval-debugging deep dive when something regresses, and one or two cross-functional reviews with PM or design. The cycle from idea to shipped feature is usually two to six weeks.

Governance roles sit between law, policy, and engineering. Trust and safety, AI risk, model governance, AI red-teaming, responsible AI. Three years ago these were rounding errors at most companies; in 2026 they are line-managed teams of 20 to 100 people at frontier labs and growing teams at every regulated industry. We cover this in detail in our AI governance jobs guide. A typical week for an AI policy lead at a regulated industry: drafting or reviewing internal policy, briefing one product team on requirements, attending a regulator-facing meeting (real or rehearsed), and reviewing one or two AI use-case applications.

There are also adjacents. AI product managers (a real career path now). AI designers, especially for agentic and conversational interfaces. AI developer relations, for companies whose API is the product. Forward-deployed engineers, a growing category at the labs themselves. Each has its own salary band and its own funnel.

The table below shows the cleanest version of who reports where, because the org-chart placement tells you who you will be evaluated by.

ML engineer / AI engineer / LLM engineer (and the differences)

The three titles overlap, but the focus differs. In 2026 the compensation order is AI engineer first, ML engineer second, LLM engineer treated as a sub-specialty of either. That is the reverse of 2022, when ML engineering was clearly the more prestigious title.

An AI engineer in 2026 typically owns: prompt design and prompt evals, retrieval-augmented generation pipelines, agent orchestration (often using frameworks built on top of provider APIs), model selection across providers, latency and cost tuning, output validation, and integration into existing product surfaces. They are evaluated on shipped features and on user-facing metrics. They do not usually train models. Median total compensation at a top-tier startup is in the $260K-$380K band. At the frontier labs, applied engineering total comp can clear $500K with equity.

An ML engineer still owns the training pipeline. Data preprocessing, distributed training, hyperparameter tuning, experiment tracking, model deployment, monitoring, drift detection. The title is closer to its 2020-22 meaning. Pay is in the $230K-$340K band at top startups, with a bias toward base salary because the role demands more multi-year domain expertise.

An LLM engineer is a sub-specialty. The most common framing is "ML engineer who works on transformer-based models specifically" or "AI engineer with deep architecture knowledge". The title is mostly a recruiting flag for candidates who know inference-time engineering tricks (KV cache management, speculative decoding, quantisation), but pay tracks whichever of the parent roles the work most resembles.

The skills that distinguish the three roles in interviews look like this:

The practical question for a candidate: which of the three should you target? If you have shipped product features and want to keep doing so, AI engineer pays best with the lowest barrier to entry. If you want to own training and have a strong stats or PhD-adjacent background, ML engineer is the cleaner fit. If you have a research background and want to go deep on architectures, target LLM engineer roles at companies that train their own models, which narrows the list to roughly a dozen employers worldwide. The path-specific skills are covered in our AI engineering jobs deep dive.

Remote-vs-on-site availability differs across the three. AI engineering roles have the highest remote share (roughly 35% of openings in 2026). ML engineering hovers at 20% remote because training infrastructure access is centralised. LLM engineering is overwhelmingly on-site or hybrid because the work happens close to the GPU clusters, which are geographically concentrated. If location flexibility matters to you, AI engineering is the path with the most options.

Research scientist (and why most people should not aim here)

Research scientist remains the role most candidates daydream about and the one with the worst hire-rate. A PhD is strongly preferred at OpenAI, DeepMind, and Anthropic, though "research engineer" tracks at all three accept strong publications without a PhD. The hard truth: total per-year hiring across the four largest labs combined is in the low hundreds, and the published-papers requirement is real. If you do not already have NeurIPS, ICML, or ICLR papers, the path takes years, not months.

The salary data, where you can find it, is exceptional. Research scientist offers at OpenAI in 2025 ranged from $370K base + $1M-$3M four-year equity for new PhDs to $500K base + $5M+ equity for senior people. Anthropic and DeepMind cluster slightly below those numbers. The catch: at frontier labs, the equity is privately held and only liquid via secondary markets, which are episodic. Research at Microsoft or Meta carries lower equity but liquid stock.

The academia-to-industry path itself has changed. Five years ago, a strong PhD candidate would do a postdoc and then go on the academic market. By 2026, the academic market for AI/ML is brutally compressed: the same candidate with the same paper portfolio earns 4-7x more in industry, and roughly two-thirds of new AI/ML PhDs go straight to industry without a postdoc. The signal that distinguishes top industry research candidates: a single first-author paper at a top venue (NeurIPS, ICML, ICLR) where the work shows experimental rigor and a defensible methodological contribution. Volume matters less than quality plus a clear research direction the lab can imagine the candidate pursuing for two-plus years.

For a reader without a PhD already in progress, research scientist should not be the goal. Research engineer is more accessible and pays comparably. Applied AI engineering pays at least 70% as well, has 20-50x the open-role volume, and rewards the kind of evidence (shipped projects, OSS contributions) that you can build in months. We expand on the path question in our job-hunt playbook.

AI product manager

AI PM emerged as a distinct role in 2024. It is now a real career path. The work differs from generic PM in three ways: you must understand model evaluations well enough to read a model card; you must reason about hallucination, latency, and cost as first-class product trade-offs; and you must own prompts and prompt regression as features, not as engineering details.

Compensation in 2026 sits roughly 15-25% above non-AI senior PM, with the largest premium at the frontier labs. OpenAI's product roles trade at staff-PM levels; Anthropic's product team is small but pays at parity to engineering. Outside frontier labs, AI PM at Series B/C startups pays $220K-$320K total comp.

The most common path in is "senior PM at a non-AI company who has shipped two or three AI features as a side specialism". The next is "ML engineer who pivots to product", which is rarer but often produces stronger PMs because the candidate already has the technical credibility. A pure outside hire without either background is rare; if that is you, expect to start as an associate AI PM or PM at a smaller company first.

What kills AI PM candidates in interviews, in our hiring-manager interviews, is a uniform pattern: they cannot articulate the hallucination-vs-helpfulness trade-off in real product terms. Asked "how would you decide between a model that is right 95% of the time but slow and a model that is right 92% of the time but fast?", candidates who treat it as an abstract question fail. Candidates who turn it into a question about the user's tolerance for incorrect output in this specific product context, the cost of a single wrong answer, and the existence of a recovery path, succeed. The signal the interviewer wants is "has thought about model behaviour as a product variable, not just a technical one".

AI governance, policy, and trust and safety

This is the highest-growth role family on a percentage basis, going from a few hundred dedicated roles in 2023 to over 5,000 globally in early 2026. The work spans three sub-specialties: policy (drafting internal AI use policies, responding to external regulation), technical governance (model evaluations for safety, red-teaming, bias audits), and trust and safety operations (handling abuse, takedown requests, content moderation in AI products).

The backgrounds that work are wider than most candidates assume. Lawyers with technology experience can move into AI policy. ML engineers with an interest in evaluation can move into technical governance. Audit and risk professionals from finance can move into model risk. The frontier labs are all aggressively hiring. Anthropic's policy team grew from under 10 in early 2024 to over 60 in 2026.

The EU AI Act is the single most consequential regulatory driver of governance hiring in 2025-26. The Act's high-risk-system obligations created a near-instant demand for AI governance leads at any company selling AI-powered products into the EU. Banks, insurers, healthcare systems, and large software vendors all built or expanded model risk management teams in response. Job postings tagged with "EU AI Act" on LinkedIn went from under 100 in Q1 2024 to over 4,000 in Q1 2026. Candidates with both ML literacy and a working knowledge of the Act's risk categories are scarce; pay reflects this scarcity.

Salary data is harder to find publicly because the role is newer and recruiters hide the bands. Approximate ranges in 2026: junior AI policy at major lab $180K-$230K total; senior AI policy $260K-$400K; AI governance lead at a Fortune 500 (banking, healthcare, insurance) $250K-$350K. The deep dive is in our AI governance jobs guide.

Applied AI roles in non-tech companies

The non-tech market is what most career-changers should target first. JP Morgan posted over 2,000 AI-tagged roles in 2025. Goldman has a dedicated AI engineering org. The big consultancies (McKinsey, BCG, Bain, Accenture, Deloitte) all have AI implementation practices that have been hiring at the rate of hundreds per quarter through 2025 and 2026. Healthcare systems, government contractors, and insurance carriers have followed.

The upside: less interview-loop brutality, often higher base salary, frequently better stability, and the ability to make a real impact because most non-tech companies are AI-immature. The downside: lower equity, slower technical career growth, and the work is more about deploying off-the-shelf models than pushing any frontier.

The role title to look for is sometimes "AI engineer", sometimes "applied AI engineer", sometimes "AI translator" (a McKinsey-introduced title for someone who bridges business problems and AI/ML technical teams). Salary in this market: $180K-$280K total comp at the senior IC level in major US metros, lower elsewhere. The interview difficulty is roughly half that of a frontier lab.

The European market for applied AI in non-tech is the most underrated opportunity in 2026. Lloyds, Barclays, HSBC, BNP Paribas, and Deutsche Bank are all hiring AI engineers at scale, often into newly-created internal AI platform teams. The Big Four (Deloitte, PwC, EY, KPMG) collectively expanded AI consulting headcount by an estimated 30,000 globally between 2023 and 2026, and a substantial share of that hiring is mid-career laterals from technical roles. Salaries are 30-40% lower than the equivalent US bank, but the visa and relocation friction is much lower for non-US candidates and the work-life balance is markedly better.

For someone breaking in mid-career, this is the rational target. We discuss the matching pattern in our AI for business hub and in automating your workflow with AI, both of which are the kinds of articles your interviewer will expect you to have engaged with.

Salary data: real numbers from real offers

The numbers below are 2026 medians from Levels.fyi, Pave, and Anthropic / OpenAI / DeepMind self-reported ranges. They are total compensation, including base, target bonus, and four-year equity grant divided by four. They are US-centric; UK and EU figures are typically 60-80% of US, with London at the top of that band.

Two things to keep in mind when reading those numbers. First, equity at private labs is the largest line item and the most volatile. A four-year grant calibrated at $61B Anthropic post-money pays out very differently if the company raises at $200B in 2027 versus a flat round. Most candidates underweight this. Second, the "non-tech F500" column is what most career-changers will actually see, because the frontier-lab rejection rate is brutal. Anchoring on the top column and being disappointed by the bottom column is a common emotional trap.

Regional adjustment matters more than candidates expect. The same AI engineer role at the same company can have very different total comp depending on the market. The table below shows approximate adjustments off the SF Bay Area benchmark.

Negotiation in 2026 has shifted toward sign-on bonuses (because base bands are tightly compressed) and toward refresh equity grants (because four-year cliffs were creating retention problems). If you are negotiating an offer, ask explicitly about the refresh policy. The actual answer to that question is often the most consequential part of the package.

The year-over-year trend matters too. Total comp at the major labs grew 18-25% on average between 2024 and 2026 across IC levels. That growth has decelerated in early 2026 as the market matures, and we expect future growth to track senior-engineering-elsewhere rates of 3-6% per year rather than the explosive growth of the previous cycle. If you are joining now and your equity package vests over four years, model the comp under the assumption of flat-to-modest valuation growth, not a doubling.

Hiring funnels at OpenAI, DeepMind, Anthropic, and Microsoft AI

The four major labs have visibly different funnels, and you should not interview-prep the same way for each.

OpenAI runs a relatively short loop: a recruiter screen, a hiring-manager screen, a take-home or live coding, and an on-site of three to five interviews. Total elapsed time is often under three weeks for engineering roles. Take-home weight is high. The bar on coding speed and clarity is the hardest in the industry, but the loop's short length means a yes is fast. Our OpenAI jobs guide walks through the loop and the take-homes you should expect. What OpenAI values most distinctively: speed of execution and willingness to push on under-specified problems.

Anthropic runs the most rigorous process of the four: a recruiter screen, two technical phone interviews, a take-home (often substantial, multi-day for senior roles), an on-site with five to seven interviewers, and a hiring committee that is famously hard to pass. Total elapsed time is six to ten weeks. Compensation is competitive; the trade-off is the time investment. What Anthropic values most distinctively: care, calibration, and the candidate's stated reasons for caring about safety being substantive rather than performative.

DeepMind emphasises research depth even for engineering roles. A research engineer loop will involve two technical interviews, a research discussion (you walk through a paper, sometimes one of yours, sometimes one of theirs), and an on-site of five interviews. Process length: five to nine weeks. The PhD-vs-no-PhD question is real here, more so than at OpenAI or Anthropic. We cover this in our DeepMind careers guide. What DeepMind values most distinctively: scientific taste, the ability to read a paper and identify what is novel and what is not.

Microsoft AI runs a more traditional Microsoft engineering loop with an AI-specific overlay. Standard four-to-five-interview on-site, a system-design round (often AI-system-design specifically: "design a real-time agent for X"), and at least one round on responsible AI / governance even for ICs. Process length: four to seven weeks. The compensation is lower than the other three but the work-life balance and stability are visibly better. What Microsoft AI values most distinctively: the ability to operate in a large, cross-team organisation; the loop screens hard for collaboration signals.

How to break in mid-career

If you are reading this with five to fifteen years of professional experience already, you are in the strongest position the market offers. Mid-career switchers are how most AI teams have actually been built since 2023. Four concrete paths.

Path 1: lateral from software engineering. Strongest signal you can produce in months: ship two to three AI features at your current company, then write detailed engineering posts about each. Recruiters at the labs scan public engineering blogs heavily. A staff-level SWE who has shipped a real RAG system and written about the eval methodology will get phone screens at every lab on a six-month timeline. Take our complete AI learning roadmap for the technical foundations to sequence.

Path 2: lateral from data science or ML. The pivot is from model-training mindset to product-shipping mindset. The skill gap is rarely technical (you already know the stats) but cultural (data scientists frequently overshoot on rigor for a feature that needs to ship in two weeks). Public evidence that you can ship matters: contribute to an open-source agent framework, build a small product, blog through it.

Path 3: full pivot from a different domain. Hardest path but increasingly viable. The successful versions tend to specialise vertically. A lawyer who pivoted into legal-AI engineering. A doctor who pivoted into healthcare-AI product. A finance professional who pivoted into AI risk. The vertical specialism is the wedge: the candidate becomes the obvious hire for one specific niche, instead of a generic AI candidate competing against thousands.

Path 4: lateral from product or design. Underused but very viable. Senior product managers and senior designers who pick up enough technical depth to ship a real AI feature themselves are scarce in the market. The transition typically takes nine to fifteen months and works best when paired with a domain specialism: a senior fintech PM who has shipped two AI features and writes publicly about their approach to model selection has a different application story from a generic PM. Compensation usually does not drop, and at startups can rise meaningfully because the candidate's product-and-technical hybrid is exactly what early AI teams need.

For all four paths, the binding constraint is evidence, not credentials. The labs hire people whose work they have already seen.

Frequently asked questions

Do I need a PhD to work in AI in 2026?

No, with one exception. Research scientist tracks at the frontier labs (OpenAI, Anthropic, DeepMind) effectively require a PhD or equivalent published research. Every other role family, including research engineering, AI engineering, ML engineering, AI product, and AI governance, hires without a PhD. Most of the highest-paid AI engineers in 2026 have undergraduate degrees only. The market reads "evidence of work" over "credentials", and a strong shipped portfolio plus contributions to open-source projects beats a generic master's every time.

Can I get an AI job without a CS degree?

Yes, and this is increasingly common. Self-taught engineers, bootcamp graduates, and switchers from physics, mathematics, statistics, and engineering disciplines all hire well into AI engineering roles. The path: build two or three substantial public projects, contribute to an open-source AI framework, and target Series A-C startups first (their bar on credentials is lower than FAANG-scale). The frontier labs do hire non-CS candidates, but the bar on demonstrated technical depth is high.

Are AI bootcamps worth it?

The honest answer is "sometimes, and not for the reasons people advertise". A bootcamp gives you sequencing and accountability. The actual learning still has to happen by building. If you would not finish a self-paced curriculum on your own, a structured program with deadlines and a peer group can get you to a hireable level in six to twelve months. The certificate itself carries little weight with hiring managers; the projects you build during the program carry most of the weight.

What is the most accessible AI role for a career switcher?

Applied AI engineer at a non-tech Fortune 500 company, in 2026. The compensation is real ($180K-$280K total in major US metros), the interview difficulty is moderate, the volume of openings is high, and most of these companies are AI-immature so a candidate with practical shipping experience is highly valued. The trade-off is lower equity and slower frontier-skill development compared to a startup or lab role.

How long does it take to break in?

From a related technical background (SWE, data, ML), a serious six-to-twelve-month focused effort is usually enough. From a non-technical background, expect 18-24 months minimum, and the path benefits from picking a vertical specialism rather than competing as a generalist. The single most predictive variable across our reader survey is "hours per week spent shipping public projects", not "hours spent on courses".

Do I need to know how transformers work in detail?

For research and frontier-lab applied work, yes, including attention mechanisms, KV caching, and tokenisation in detail. For applied AI engineering at most companies, working knowledge is enough: you should be able to explain why context windows matter, what an embedding is, how a tokeniser maps text to model input, and the cost-latency-quality trade-offs across model tiers. Deep architecture knowledge becomes useful at the staff-engineer level and above.

What if I am 40 or older?

The data is reassuring. Across hires we have tracked at major labs in 2025-26, the median age has trended up, not down, since 2022. Mid-career and late-career switchers bring shipping discipline and product judgement that recent graduates often lack. The accessibility ranking by company tier in 2026 puts non-tech F500 first, then enterprise software companies, then high-growth startups, then frontier labs. Age signals are weakest in the first two tiers and strongest in startups, where the cultural skew toward youth is real but eroding.

Should I learn one big framework or many small ones?

Depth in one production stack beats breadth across many demos. By 2026 the consolidation around a few stacks (a major model provider, a vector database, an evaluation framework, an agent runtime) means that going deep on one full pipeline produces an interview story that ships. Breadth across ten tutorial-grade demos produces no interview story at all.

Do AI labs sponsor visas in 2026?

Yes, and aggressively, but the picture differs by lab and country. OpenAI and Anthropic both sponsor H-1B and O-1 visas for US roles; DeepMind sponsors UK Skilled Worker visas in London. Microsoft AI uses standard Microsoft global mobility, which is well-resourced. The bottleneck for non-US candidates is not the lab's willingness; it is the H-1B cap lottery, which makes US relocation a multi-year proposition unless you qualify for O-1 (extraordinary ability) or already hold a green card. London-based candidates have an easier path to UK AI roles than to US ones; many strong candidates have used a UK-based role as the bridge.

What domains are hottest within AI in 2026?

Three stand out by hire volume in early 2026. Agentic systems engineering (multi-step AI workflows that operate tools and APIs autonomously) tops applied hiring at startups. AI governance and model risk (driven by the EU AI Act and US sectoral regulation) tops hiring in regulated industries. Foundation-model evaluation and red-teaming, including adversarial robustness work, tops hiring at frontier labs because it is the bottleneck on shipping more capable models safely. Domain-specific verticals (legal AI, healthcare AI, defence AI) are growing fast off smaller bases.

What does an entry-level AI engineer interview week actually look like?

For a structured candidate running parallel loops in 2026, a normal week looks like this: two to three recruiter calls (30-45 minutes each), one or two technical phone screens (60-90 minutes each), one take-home assessment in progress (3-6 hours of work, often over a weekend), and one on-site loop scheduled for the following week. Above that volume, candidates typically lose interview quality because preparation per loop suffers. Below that volume, the search stalls because no parallel offers materialise. Two to three serious loops in flight at any given time is the right cadence for most candidates; we discuss the cadence in detail in our job-hunt playbook.

Are AI internships worth pursuing if I am already mid-career?

Sometimes, with one specific framing. Most AI internships at frontier labs are explicitly aimed at students; mid-career candidates rarely fit. The exception is the "residency" or "visiting researcher" programmes that several frontier labs run for mid-career switchers, typically six to twelve months long. These programmes have produced full-time hires in roughly 30-50% of cases in 2024-26, which is a strong base rate for a structured switcher path. They pay well below full-time roles but they are visible from the labs and they produce a credentialled track record that converts to interviews at peer companies even if the conversion to full-time at the host lab does not happen.

The bottom line

If you are reading this trying to decide whether to make the jump, the answer in 2026 is that the market is open, the doors are wider than they have been in any previous AI hiring cycle, and the salary numbers justify the effort even allowing for a 50-60% probability of rejection at any specific company. Pick one of the four paths. If you already build software, ship two visible AI features at your current job and write about them in detail. If you already work in data or ML, pivot toward product-shipping work and contribute publicly to a framework. If you are switching from a different domain, choose your vertical specialism and become the obvious hire for that wedge. If you are a senior PM or designer, build technical depth alongside your product instincts and target the AI PM track. Then target the right tier of company for your evidence: non-tech F500 first if your portfolio is two months old, top-tier startup if it is twelve months old, frontier lab if it is two years old. Read every guide in our AI careers hub and pick a target before applying. The candidates who get hired in 2026 are the candidates who can show, not tell.

Last updated: May 2026