Nvidia just compressed a ten month, eight engineer GPU design task into an overnight job. Meta is cutting 8,000 people on 20 May and wants 65% of its engineers writing 75% of their code with AI by mid-2026. Tech shed nearly 80,000 roles in Q1 2026 alone, and about half of those cuts were pinned on AI. This is the short, practical read on which IT jobs are exposed and what to become instead.
You will learn how to
- Spot which five IT roles AI is already eating.
- Evolve each one into work AI cannot do alone, with concrete next steps.
- Pick five adjacent careers that stay safe through the next decade.
The rule of thumb is simple. If the task is a closed loop with clean inputs, AI will do it. If the loop closes with a human, a building, or a body, it will not.
5 IT jobs at risk
1. Junior software developer
Anthropic's Economic Index puts programmers at the top of the AI exposure list, and Meta's internal target of 75% AI written code in months is not an outlier, it is where the industry is headed. The routine junior ticket we all cut our teeth on, rename a prop, add a button, wire a REST endpoint, is exactly the shape of work Copilot, Claude Code, and Cursor now finish in a single prompt. Entry level software engineering postings dipped across the US and UK in Q1 2026, and the ladder's first rung is quietly being sawn off.
How not to be replaced?
Evolve into a systems thinker who owns distributed design, data modelling, and the why. AI still flounders on concurrency, on failure modes at scale, and on trade offs that do not fit in one context window. That is the work that still pays.
Actionable paths:
- Read Designing Data Intensive Applications by Martin Kleppmann cover to cover.
- Work through MIT 6.824 Distributed Systems and implement the Raft labs in Go.
- Pair that with our own guide, data pipeline design patterns, to ground it in real workloads.
2. Tier 1 IT support and help desk
Password resets, printer drivers, and FAQ triage are textbook automation. Zendesk and Intercom are reporting north of 60% deflection on tier 1 tickets with their 2026 AI agents, and the deflected ones never touch a human. Anything that fits on a runbook page is already a bot's lunch, and the helpdesk tier most of us started on is vanishing quietly.
The way up is to stop being a human router and own the work the router cannot see.
Evolve into incident command and on-call infra, where judgement and physical presence win. That work lives in racking a server, chasing a bad cable in a cold aisle, or standing in a room with a furious CFO whose laptop is bricked before a board meeting. No LLM does any of that over chat.
Actionable paths:
- Read the free Google SRE Book and the Site Reliability Workbook.
- Study the PagerDuty Incident Response documentation.
- Build a home lab with Prometheus and Grafana, then run a fake outage against it.
3. Manual QA tester
Generative tools now write, run, and maintain test suites faster than we can. Playwright's codegen plus an LLM oracle turns a product spec into a working end to end suite in a morning, and self heals when selectors drift. Exploratory clicking was a real craft, and we have to be honest that it is also a commodity now.
So where do humans still win? Wherever the test is not already sitting inside the spec.
Evolve into SDET, chaos engineer, or security tester who breaks systems in ways AI cannot model. Adversarial thinking, race conditions across services, auth flows that an attacker has not seen yet, performance cliffs under real traffic, these are open ended problems with no training set and no obvious oracle.
Actionable paths:
- Ship a real test suite with Playwright against a live app.
- Run chaos experiments with LitmusChaos or Chaos Mesh on a Kubernetes sandbox.
- Work through the OWASP Web Security Testing Guide and grind a month of rooms on TryHackMe.
4. Basic data analyst and report builder
Dashboards, SQL boilerplate, and narrative summaries are table stakes for copilots now. Anthropic and the Washington Post both place analyst roles near the top of the exposure list, and a modern warehouse copilot already writes the weekly KPI deck, flags anomalies, and drafts the commentary. The job that used to be "pull last week's numbers and explain them" is a single prompt on the right dataset.
Long story short, the numbers are easy, the plumbing is not.
Evolve into data engineer, analytics engineer, or ML platform owner. The defensible work sits upstream, owning the pipeline, the schema, the SLAs, and the data contracts that decide whether tomorrow's numbers can be trusted at all.
Actionable paths:
- Complete the free Data Engineering Zoomcamp.
- Take dbt Fundamentals and ship a modelled warehouse on public data.
- Read Fundamentals of Data Engineering by Reis and Housley, then review our Hive partitioning layout guide for hands-on lake design.
5. Template front end developer
Prompt to UI tools now ship production React in an afternoon. v0, Lovable, Bolt, and Claude's artifacts all generate working interfaces from a one line brief. Landing pages, marketing sites, and CRUD screens that used to take a sprint are commodity output, and the portfolio move of "look, a nice dashboard" is no longer proof of much.
Shipping a button is free now. Shipping a trustworthy interface is not.
Evolve into design systems engineer, accessibility specialist, or performance expert. Multi tenant state, WCAG 2.2 compliance, Core Web Vitals budgets, internationalisation, and the hydration bug that only shows up on a slow phone in Lagos, these are the problems AI ships into and cannot debug alone.
Actionable paths:
- Complete web.dev Learn Accessibility and Learn Performance.
- Study the GOV.UK Design System and rebuild one component with full a11y coverage.
- Run Lighthouse CI on a real production site and close the top ten issues.
Writers, customer service reps, and translators top the non IT exposure list. The pattern repeats: language heavy, single turn work is first to go.
5 jobs AI will not replace soon
- Site reliability and on call infra engineers. 3am pages, physical data centres, cross team incident command. BCG and Anthropic both place these outside the automation frontier. Nobody wires an autonomous agent into a production database during a Sev 1.
- Security and red team engineers. Adversarial thinking against a human attacker is open ended by definition. Novel auth flows and supply chain attacks do not sit inside a training set, and regulators will not sign off on an AI CISO.
- Staff and principal engineers, platform architects. Ambiguous requirements, political trade offs, and roadmap calls across a dozen teams. AI assists, it does not decide. The work is as much stakeholder management as it is code.
- Skilled trades: electricians, HVAC, field technicians. The US Bureau of Labor Statistics projects 11% growth for electricians through 2033 and 79,900 annual openings. Robots do not rewire a hospital at night.
- Nurse practitioners and clinicians. Projected 45.7% growth by 2032. Messy bodies, messy conversations, and a regulatory wall that no LLM will cross this decade.
If the job ends with a human looking us in the eye or a building catching fire, AI is not coming for it this decade.
Conclusion
The Nvidia and Meta cuts are not an anomaly, they are the shape of the next five years in IT. The safe move is not to hide from AI, it is to climb above the tasks it does well and anchor the work to systems, people, and physical reality. Pick one evolution above, block out one hour a day on one of the linked resources, and start this quarter.
Resources
- Tech industry lays off nearly 80,000 in Q1 2026, half due to AI, Tom's Hardware
- Nvidia: AI cuts 10 month GPU design task to overnight, Tom's Hardware
- Meta to cut 8,000 jobs on 20 May as AI reshapes engineering, The Next Web
- Which jobs are most at risk in the age of AI, Inside Higher Ed
- Anthropic, jobs most at risk from AI, Futurism
- AI job losses, interactive exposure map, Washington Post
- AI will reshape more jobs than it replaces, BCG
- Top 65 jobs safest from AI and robot automation, US Career Institute
- 100 AI resistant careers backed by BLS data, Excel High School
Comments