In May 2023, IBM CEO Arvind Krishna announced the company would pause hiring for roughly 7,800 back-office roles — positions in HR, finance, and document processing — because he expected AI to replace them within five years. It was one of the first explicit, on-record statements by a major CEO connecting AI directly to planned headcount reduction in professional roles.
Six months later, Goldman Sachs published an internal analysis estimating that generative AI could automate 25–50% of current work tasks at legal, accounting, and consulting firms. The tasks most at risk were document review, first-draft writing, data summarization, and routine client correspondence.
Earlier AI waves primarily hit manufacturing and data entry. The large language model breakthrough of 2022–2023 moved the frontier into work that requires reading, reasoning about text, and generating coherent prose — work previously insulated by its cognitive complexity.
The key shift: LLMs do not need to be programmed for each specific task. A single model can draft a contract clause, summarize a 200-page discovery file, write a performance review, answer a benefits question, and generate a financial memo — often at speeds and costs that make the economics of human labor difficult to justify for lower-complexity versions of those tasks.
Three documented cases from 2023–2024 illustrate the scope:
| Company / Sector | AI Application | Documented Outcome |
|---|---|---|
| Klarna (fintech) | Customer service AI (OpenAI-powered) handling support queries | Replaced work of ~700 FTE agents; handled 2.3M conversations in first month (Feb 2024) |
| KPMG / Deloitte (consulting) | AI document review and due-diligence drafting tools rolled out firm-wide | Junior associate review time cut 60–80% on standard M&A document sets per internal pilots |
| BT Group (telecom) | AI for network monitoring, billing queries, and HR ticketing | Announced 55,000 job cut target by 2030; AI specifically cited for back-office reduction |
None of these cases involved AI eliminating an entire profession. What happened is more precise: specific task bundles within jobs were automated. A junior lawyer still exists — but fewer hours billing for document review. A customer service manager still exists — but managing an AI queue, not a team of 50 agents.
Economists Daron Acemoglu (MIT) and David Autor (MIT) have spent years building what they call a task-level model of automation. Rather than asking "will this job disappear?" they ask "which specific tasks within this job can a machine now perform?" This framing matches what is actually happening in the labor market.
Every job is a bundle of tasks. Some tasks are routine and codifiable — following rules, matching patterns, retrieving information. Others are non-routine — exercising judgment under ambiguity, building trust, improvising solutions, physical dexterity in novel environments. AI in 2024 is extremely capable at the first category and improving rapidly in the second.
The workers most affected so far are those whose jobs consist heavily of text-in, text-out tasks at a junior level: paralegals doing first-pass contract review, junior consultants building slide decks from data, entry-level copywriters producing boilerplate content, financial analysts drafting standard reports. These are often roles people use to enter professional careers and build foundational skills.
A 2024 study by researchers at Harvard Business School and the University of Chicago tracked actual hiring on Upwork — a major freelance platform. Categories most affected: writing and content creation (−21% posting volume in 18 months after ChatGPT launch), basic coding and web development (−17%), and data entry and research (−29%). These are real market signals, not projections.
If you are entering a field that involves significant amounts of routine knowledge work — drafting, summarizing, researching, templating — you are entering it at a moment when those specific tasks are being automated. That does not mean the field is closed. It means the path through it is changing. The skills that used to get you promoted — volume of output — are being devalued. The skills that matter now are judgment, relationships, and the ability to direct and evaluate AI output.
Pick a professional role you know well (your own job, a target career, or someone you know). Ask the AI assistant to help you break it into its core task components and assess which ones are currently automatable. The goal is to build a task-level map — not a vague "will AI take my job?" answer.
In August 2023, the law firm Allen & Overy — one of the world's largest by revenue — announced it had deployed Harvey, an OpenAI-powered legal AI, firm-wide across its 43 offices. Lawyers at the firm were using it to draft first versions of contracts, answer client questions from prior case data, and conduct preliminary regulatory research. The firm did not announce layoffs. It announced that its lawyers could now handle more matters with the same headcount.
That phrase — "more matters with the same headcount" — is how most legal AI deployment is being framed. It is, functionally, a way of saying fewer hires are needed to grow revenue.
The legal tasks most affected fall into three categories, all well-documented by 2024:
| Task Category | AI Tool / Company | What It Replaces |
|---|---|---|
| Document review (e-discovery) | Relativity, Kira Systems, Luminance | Hours of paralegal and junior associate time scanning documents for relevance, privilege, key clauses |
| Contract drafting (first pass) | Harvey, Ironclad AI, LexisNexis AI | Junior associate drafting time for standard agreements (NDAs, MSAs, employment contracts) |
| Legal research | Westlaw AI, Casetext CoCounsel | Associate hours hunting precedents, summarizing case law, building argument frameworks |
| Due diligence | Kira, Diligen, Luminance | Weeks of M&A due diligence reading — now done in hours on standard document sets |
The traditional law firm model is built on billable hours. Junior associates bill 1,800–2,200 hours per year, much of it on document review and research. If AI cuts that time by 60–80%, firms face a revenue model problem. The response in 2024: shifting to flat-fee engagements, reducing associate class sizes, or restructuring billing entirely. Yale Law School reported in 2024 that on-campus law firm recruiting for associate positions was down 12% from 2022 levels.
AI in medicine is disrupting differently — less through cost-cutting at law firms and more through genuine capability expansion. Three documented deployments stand out:
Radiology AI: Google's DeepMind published results in 2023 showing its AI detected breast cancer in mammograms with greater accuracy than the average radiologist (fewer false negatives, fewer false positives). The UK's National Health Service began limited deployment of AI radiology screening tools. Radiologist employment has not dropped — but radiologist growth has slowed and new graduates are entering a field with AI as a permanent colleague, not a future threat.
Ambient clinical documentation: Microsoft's Nuance DAX and similar tools now transcribe and structure doctor-patient conversations in real time, auto-generating clinical notes. Over 200 US health systems had deployed some version of ambient AI documentation by late 2024. The task being automated: the 1–2 hours per day physicians spend on administrative charting. This is widely seen as augmentation — doctors see more patients, burn out less. But it also reduces demand for medical scribes, a job that employed tens of thousands.
Pathology AI: Paige.AI received FDA clearance in 2021 for AI-based prostate cancer detection in pathology slides. By 2024, multiple AI pathology tools were FDA-cleared. The pattern mirrors radiology: AI handles initial screening and flags, pathologist confirms and handles edge cases.
What both fields reveal is a split between credentialed knowledge tasks and judgment-under-uncertainty tasks. AI can pass the bar exam (GPT-4 scored in the 90th percentile in 2023). It can interpret a radiology scan at average-radiologist accuracy. What it cannot reliably do: advise a client whose situation is emotionally complex, navigate the trust relationship with a frightened patient, or make the call when the facts are genuinely ambiguous and the stakes are high.
The lesson for career planning: the credential is no longer enough. What you can do beyond what a well-prompted AI can do is the question that will shape compensation and job security in both fields.
Many firms and hospitals deploying AI describe it as "augmentation" — making existing workers more productive. This framing is often genuine. But augmentation at scale means fewer people are needed to do the same volume of work. If a junior lawyer can now do in two hours what previously took ten, firms need fewer junior lawyers. Augmentation at the individual level and displacement at the workforce level are not contradictions — they can be the same thing.
This lab explores the boundary between "what AI knows" and "what requires human judgment." Ask the AI to handle a professional task from law or medicine — something that seems like it requires credentials. Then probe its limitations. Where does it hedge? Where does it confidently produce something useful? Where does it fail or appropriately decline?
The 2023 Hollywood writers' strike lasted 148 days — the longest in WGA history. AI was explicitly named in the contract demands. Writers sought protections against studios using AI to generate scripts or outlines that human writers would then "punch up" at lower pay. The eventual agreement included provisions limiting how AI-generated content could be used in the writing process and requiring disclosure when AI was involved.
It was the first major US labor contract to directly regulate AI in creative work. The fight was not hypothetical — studios had already begun experimenting with AI script generation tools, and the WGA documented specific cases where AI-generated outlines had been shared with writers for revision.
By 2024, the capability landscape for creative AI had expanded dramatically. Key documented capabilities:
| Creative Task | AI Capability Level | Evidence |
|---|---|---|
| Short-form copywriting | High — often production-ready | Upwork writing postings down 21%; major ad agencies running AI copy in A/B tests |
| Stock illustration / image generation | High — disrupted the market | Shutterstock, Getty saw volume decline; iStock reported contributor income drops 2023–2024 |
| Journalism (data-driven articles) | Medium — in use at scale | AP has auto-generated earnings reports since 2014; Washington Post's Heliograf covered 850 stories in 2016 alone |
| Feature writing / investigative journalism | Low — AI drafts are weak | No documented case of AI replacing long-form investigative capacity |
| Music composition (background/functional) | Medium-High | Epidemic Sound, Soundraw, Udio offer AI music; used in advertising and YouTube extensively by 2024 |
| Film/video VFX | Growing rapidly | Runway Gen-2, Sora; studios using AI for background generation and de-aging in 2024 productions |
The most complete case of AI disrupting a creative market is stock photography and illustration. The timeline is stark:
2022: Midjourney, DALL-E, Stable Diffusion release. Quality is impressive for many commercial use cases.
2023: Shutterstock announces it will sell AI-generated images and pays a fund to contributors whose work trained the model — implicitly acknowledging their work was used. Getty Images sues Stability AI for scraping its library without compensation.
2024: Multiple independent illustrators report 40–70% income drops from stock platforms. The market for generic commercial illustration — characters, backgrounds, product mockups — has largely shifted to AI generation. The market for distinctive artistic style, editorial illustration, and work requiring a known creator identity has held better but not been unaffected.
In November 2023, Futurism reported that Sports Illustrated had published articles under fake AI-generated author names with AI-generated profile photos. The publisher, The Arena Group, initially claimed the content was from a licensing partner. The revelation caused significant reputational damage and the CEO was subsequently ousted. It illustrated both that AI content was already being published at scale at major outlets — and that the lack of disclosure carried serious consequences.
Not all creative work is equally disrupted. The pattern that has emerged:
Most affected: Generic, commodity creative work — stock images, boilerplate ad copy, template-based design, routine data-journalism articles, background music. These were already low-margin, high-volume tasks.
Less affected (so far): Work where the creator's identity, voice, or relationship is the product. A columnist's following is built on their perspective. A brand photographer's clients buy their eye. A novelist's readers want their world. AI cannot replicate established creative identity — yet.
The dangerous middle: Mid-career creatives who specialized in "good enough" commercial work — competent but not distinctive. They are losing ground fastest because AI output is increasingly "good enough" for the use cases they served.
For anyone in a creative field, the question to ask is: Is what I produce valuable because of its content — or because of who made it? Content-value is increasingly replicable by AI. Identity-value is not. Building a distinctive voice, documented expertise, or audience relationship is now a career survival strategy, not just a nice-to-have.
This lab uses the AI as a creative tool — then has you analyze the output critically. You'll test AI's performance across different creative tasks and identify where it produces "good enough" commodity output vs. where human creative identity still wins.
In 2022 and 2023, major retailers made high-profile bets on automation. Amazon's "Just Walk Out" cashierless technology was deployed in Amazon Fresh stores and licensed to third-party retailers globally. Walmart expanded its use of shelf-scanning robots. Ahold Delhaize deployed robots in Giant Food stores for inventory scanning.
By early 2024, Amazon had quietly removed Just Walk Out technology from all its Amazon Fresh stores, replacing it with "Dash Carts." The reason cited internally: the system relied heavily on human reviewers in India watching video feeds to classify purchases — reportedly more than 1,000 workers. The promised fully-automated checkout had not materialized. Physical retail proved harder to automate than advertised.
Robotics and AI face a fundamental challenge in unstructured physical environments. This is sometimes called Moravec's Paradox — named after roboticist Hans Moravec, who observed in the 1980s that the tasks hardest for humans (complex math, chess, language) are easiest for computers, while tasks easy for humans (walking, picking up a cup, recognizing an object in unusual lighting) remain extremely hard for machines.
This paradox has not been fully resolved. LLMs solved the language problem but did not solve the physical-world manipulation problem. The result: jobs requiring dexterous physical work in unpredictable environments have held better than predicted.
| Job Category | Automation Prediction (circa 2017) | Actual Status 2024 |
|---|---|---|
| Truck drivers | Highly at risk; autonomous vehicles would replace within a decade | Still in strong demand; autonomous trucking limited to specific routes, not widespread deployment |
| Fast food workers | Robots would replace order-taking and food assembly | Kiosks common for ordering; food assembly automation limited and expensive; net employment largely stable |
| Construction workers | Modular and robotic construction would reduce labor needs | Labor shortages, not surpluses; robotics adoption slow due to site variability |
| Plumbers, electricians | Eventually automatable as robotics advanced | Severe shortages; wages rising significantly; no meaningful robotics deployment |
| Home health aides | Uncertain; social interaction seen as barrier | Fastest-growing US occupation 2023–2024; AI tools assist but do not replace |
The US Bureau of Labor Statistics Occupational Outlook Handbook (2024 edition) projects home health and personal care aides will grow by 22% from 2022–2032 — adding over 800,000 jobs. This is among the fastest growth of any occupation. Care work requiring sustained human presence, emotional attunement, and physical assistance has proven highly resistant to automation.
In the same period that AI disrupted knowledge work entry-level hiring, skilled trades faced the opposite problem: severe labor shortages. The reasons are structural and long-developing, but the contrast is striking.
Electricians: The US needs approximately 80,000 new electricians per year. Current training pipelines produce far fewer. The electric vehicle transition and data center buildout (driven in part by AI infrastructure) are creating enormous additional demand. The Bureau of Labor Statistics projects electrician employment to grow 11% from 2023–2033 — faster than average for all occupations.
HVAC technicians: Similar story. The transition to heat pumps and more efficient systems requires skilled installation and maintenance. Employment projected to grow 9% through 2033.
Plumbers: An aging workforce is retiring faster than new workers enter. The combination creates wage pressure upward — journeyman plumbers in many US cities earned $80,000–$110,000 in 2024, with master plumbers often exceeding $120,000.
The AI wave is actually increasing demand for some physical trades, because AI infrastructure (data centers) requires massive electrical, HVAC, and plumbing installation.
Based on where AI is and where robotics is in 2024, the following job characteristics correlate with lower current disruption risk:
The AI disruption story is not "all jobs are at risk." It is more precise: text-based, routine, scalable knowledge tasks at the junior level are the most disrupted right now. Physical, relational, variable-environment, and high-stakes judgment work is holding — and in many cases growing. Career resilience means understanding which side of this line your work sits on, and whether you can move toward the resilient side if needed.
This is the most personal lab in Module 2. Use the AI to do a structured resilience assessment of your own career position — or a career you're considering. The goal is to produce an honest, specific assessment: which tasks are exposed, which are resilient, and what moves could increase your resilience.