Which Jobs Are Actually at Risk? AI Agents and the 2025 Labor Market

In January 2024, Klarna's CEO announced a number worth sitting with: their AI assistant — built on OpenAI — handled the equivalent work of 700 customer service agents in its first month of production deployment. Not a pilot. Not an internal experiment. Production. The company simultaneously paused external hiring and began shrinking its total headcount.

The same month, Duolingo laid off roughly 10% of its contractor workforce — primarily content creators and translators. CEO Luis von Ahn said it plainly: the reason was AI, not weak business performance.

If you are still asking "will AI really take jobs?", that question has already expired. The useful question now is: which roles, through what mechanism, at what speed?

This article goes straight to the data — no vague forecasts, no false reassurance.

---

Why This Wave Is Different

Earlier automation primarily displaced physical labor and repetitive clerical work. Factory workers, logistics staff, data-entry clerks. White-collar, educated workers were assumed to be protected by the cognitive nature of their work.

That assumption no longer holds.

Current-generation AI agents — Salesforce Agentforce, Microsoft Copilot Studio, custom pipelines on GPT-4o — can now complete multi-step cognitive tasks without requiring human approval at each step. Filing forms, drafting contracts, processing refunds, conducting legal research, generating code. These are the core functions of knowledge work, not the periphery.

McKinsey Global Institute estimates that generative AI could automate tasks accounting for 60–70% of employee time across occupational categories. This does not mean 60–70% of jobs vanish — it means 60–70% of the time within existing jobs could be redirected to machines. The practical outcome: fewer people needed to produce the same output.

That is a qualitatively different dynamic from every prior automation wave.

---

Who Is Being Displaced — at the Task Level, Not Just by Job Title

The standard way to frame automation risk is to list "jobs at risk." That framing is technically wrong. AI doesn't replace jobs — it replaces specific tasks within jobs. But when enough tasks are replaced, headcount falls.

Tier-1 Customer Service

The fastest and clearest category. AI agents handle password resets, order tracking, FAQ responses, and refund processing at near-zero marginal cost. Klarna is the most prominent example, but not the only one. Salesforce Agentforce is now deployed at thousands of companies handling tier-1 interactions without human escalation.

What remains for humans: escalation handling, emotionally complex situations, high-value customers. That's roughly 30% of prior volume — requiring significantly fewer people.

Volume Translation and Localization

Duolingo is the cleanest case study. AI translation — GPT-4, DeepL, Google Translate — has crossed the quality threshold for most commercial applications. Literary translation, legal precision, and nuanced cultural adaptation still need human judgment. But volume translation — product documentation, subtitles, user guides — is largely automated.

Junior Software Developers (Specific Tasks)

GitHub Copilot had more than 1.8 million developer users by 2024. GitHub's research found developers completed tasks 55% faster on average with the tool, and Copilot was responsible for approximately 46% of accepted code contributions on the platform.

The consequence: boilerplate generation, unit test writing, documentation, and basic CRUD functionality have been commoditized. Junior developers whose primary work fell in these categories face the most pressure. The junior-to-senior pipeline is disrupted — not because senior developers matter less, but because the entry-level ramp is narrowing.

Data Entry and Administrative Roles

Document processing, invoice extraction, email triage, form completion — all being handled by AI agents combined with OCR and workflow tools (UiPath, Microsoft Power Automate). In May 2023, IBM announced it would pause hiring for approximately 7,800 back-office and HR roles that AI could replace within five years. Reuters reported this as a deliberate strategic decision, not a temporary freeze.

Basic Content Creation and Copywriting

Product descriptions, social media posts, ad copy variants, and boilerplate content have reduced demand for entry-level copywriters. Strategic content, brand voice, and high-end creative work still require humans — but far fewer people can now produce far higher volume.

Junior Legal and Routine Financial Analysis

Harvey AI, Casetext (now part of Thomson Reuters), and LexisNexis AI are automating contract review, legal research, and due diligence. Major investment banks deployed AI tools that reduced time spent on routine analytical tasks by 30–50% in some workflow measurements — shrinking paralegal and junior analyst headcount at law firms and financial institutions.

---

Task-Level Risk Summary

Task Category	Risk Level	Primary AI Mechanism
Tier-1 customer support	Very High	AI agents (Agentforce, GPT-4o)
Volume translation	Very High	GPT-4, DeepL
Data entry / form processing	Very High	OCR + workflow agents
Boilerplate code generation	High	GitHub Copilot, CodeWhisperer
Junior legal research	High	Harvey AI, Casetext
Routine financial reporting	High	Automated report generation
Entry-level copywriting	High	GPT-4, Claude
Strategic content	Low–Medium	AI-assisted, not replaced
Software architecture	Low	Requires holistic judgment
Complex customer escalation	Low	Requires empathy + context

---

Who Is Gaining Ground

The market isn't simply contracting — it's bifurcating sharply.

AI/ML Engineers and Researchers: Demand far exceeds supply. Median total compensation for senior ML engineers at major US tech companies reached $300,000–600,000+ in 2024, according to Levels.fyi data. The competition between Anthropic, OpenAI, Google DeepMind, Meta AI, and Microsoft is creating extraordinary conditions for this talent tier.

Prompt Engineers and AI Product Specialists: A role that barely existed in 2022 is now a mainstream job title, paying $100,000–200,000+ at US companies. The core skill set: designing AI workflows, evaluating model outputs, and bridging business requirements with AI capabilities.

Domain Experts Who Use AI: This is the most relevant category for most professionals reading this. The WEF Future of Jobs Report 2025 found that 44% of workers' core skills will be disrupted in the next five years — and identified analytical thinking, creative thinking, and AI/big data literacy as the three skills employers most expect to grow in value. Lawyers using Harvey AI complete due diligence faster. Developers using Copilot ship faster. The consistent pattern: professionals who adopt AI tools outperform peers who don't, and employers prefer them. Early adopters within traditional fields have converted productivity gains into real career leverage.

Data Curators and AI Trainers: Human feedback remains essential for model alignment. Scale AI and similar platforms have expanded their contractor networks, particularly for high-skill curators who can produce quality feedback on complex reasoning tasks.

---

The Augmentation vs. Replacement Reality

The "AI augments, doesn't replace" framing dominated 2022–2023 industry messaging. By 2025, the picture is clearer and more honest.

Where augmentation is genuine: High-complexity, judgment-intensive work benefits from AI assistance without displacing workers. A doctor with AI diagnostic support sees more patients; demand for doctors has not declined. The same applies to architects, therapists, senior researchers, and educators. AI raises the quality ceiling and throughput of skilled professionals without eliminating the human role.

Where replacement is real: Routine cognitive tasks that were previously performed by specialized workers — translation, data entry, tier-1 support, standard code generation, boilerplate content — are being replaced, not augmented. When Klarna's AI replaces the equivalent of 700 agents, those are displaced workers — not workers doing less because they're being helped.

The honest middle ground: Most jobs contain a mix of automatable and non-automatable tasks. AI handles the automatable portion, which changes the nature of the job and typically reduces required headcount. A customer service team that needed 100 people may now need 30 — those 30 handling escalations and complex situations while AI manages 70% of volume. That is workforce-level replacement even if the remaining individuals are individually "augmented."

The speed problem: Even if net jobs are positive over a ten-year horizon — the optimistic WEF framing — the pace of transition is causing real dislocation now. A worker displaced from customer service in 2024 cannot become an ML engineer by 2025. The mismatch in skills, geography, and transition time is the central problem the 2025 labor market has no clean answer to.

---

Common Mistakes

1. Assessing risk by job title instead of task composition

"I'm a Data Analyst — that's not high risk." But if 70% of your time is writing recurring SQL reports and populating dashboard templates, those specific tasks are being automated. The correct frame: audit your actual daily tasks, not your job title.

2. Confusing "AI isn't perfect" with "AI isn't good enough to matter"

AI translation doesn't reach literary quality, but it clears the bar for 80% of commercial applications. GitHub Copilot generates buggy code, but developers still ship 55% faster. The "good enough" threshold is lower than most people assume — and low enough to change headcount decisions.

3. Waiting for AI to "stabilize" before engaging

The most common trap. Stack Overflow's 2024 Developer Survey found 76% of developers already using or planning to use AI tools. The gap between users and non-users has become a productivity gap, and employers are measuring it.

4. Treating technical skill alone as sufficient

In an AI-saturated environment, judgment about AI — knowing when to trust output, when to verify, when to override — is more valuable than the ability to produce output manually. Domain expertise plus AI literacy beats either alone. The combination is what the market is actually pricing.

5. Reading "augmentation" as unconditional protection

A team augmented by AI can still be dramatically smaller than the previous team. Individual-level augmentation does not protect against organizational-level restructuring. Klarna's remaining workers are augmented. The 700 displaced workers were replaced.

6. Applying US or European data directly to other markets

Labor market dynamics differ significantly by region. Markets with large BPO and outsourcing sectors, younger workforce demographics, or different regulatory contexts will experience AI displacement at different rates and through different mechanisms. The directional signal is real everywhere; the specific numbers need local calibration.

---

Five Concrete Actions — Not Vague "Learn AI" Advice

1. Audit your task portfolio this week

Write down your 10 most time-consuming tasks from last week. For each one, ask honestly: "Can AI do this independently? With my direction? Or does it require me fully?" If the first group accounts for more than 50% of your time, you need an active strategy, not a wait-and-see posture.

2. Use the AI tool specific to your domain — not just ChatGPT in the abstract

GitHub Copilot if you write code. Harvey AI or Casetext if you work in legal. Claude or GPT-4o for analysis and writing. The goal isn't to "know AI exists" — it's to build hands-on fluency within your actual workflow. Abstract familiarity is not leverage.

3. Practice evaluating AI output, not just generating it

The most valuable skill in an AI-augmented environment isn't prompt writing — it's knowing when AI is wrong. Build this by using AI on tasks you know deeply, then systematically analyzing its errors. This is the foundation of AI supervision, which is increasingly the core human role in AI-augmented workflows.

4. Move deliberately toward the hardest 20% of your field

Whatever AI still does poorly in your domain — the most complex, ambiguous, judgment-heavy tasks — that's where to invest your development time. In customer service: VIP account management and crisis handling. In law: litigation strategy and courtroom argumentation. In data science: identifying the right business question, not just answering the wrong one. AI is most capable in the middle of the complexity spectrum; humans remain most critical at the top.

5. Build T-shaped skills with AI literacy as the horizontal bar

Your domain expertise is the vertical — preserve and deepen it. AI literacy is the horizontal — add it deliberately. A lawyer who understands how Harvey AI evaluates contracts is more valuable than one who doesn't. A data analyst who can build AI pipelines is more valuable than one who only runs SQL. That combination is not replaceable; it's the new definition of domain expertise.

---

Key takeaways:

• Klarna (700 agents), Duolingo (10% of contractors), IBM (7,800 roles paused), BT (55,000 over seven years) — these are not forecasts. They are disclosed, documented decisions made by companies that cited AI explicitly as the driver
• AI replaces tasks first and headcount second — the right unit of analysis is specific daily tasks, not general job categories
• Tier-1 support, volume translation, boilerplate coding, data entry, and routine analysis are the highest-risk task categories, and the displacement is already measurable
• Professionals who use AI tools outperform peers who don't, and that productivity gap is becoming a career gap — the window for early adoption advantage is narrowing
• "Augmentation vs. replacement" is not a binary — most jobs contain both, and being individually augmented offers no protection against team-level or organizational-level restructuring