In July 2018, researchers from DeepMind and Moorfields Eye Hospital NHS Foundation Trust published results in Nature Medicine showing their AI system matched or exceeded the diagnostic accuracy of the world's leading ophthalmologists on over 50 sight-threatening eye diseases — using only optical coherence tomography scans. The system recommended the correct referral decision 94.5% of the time, compared to 93.4% for the top human retinal specialist. The radiologists in the study were not made redundant. Instead, the collaboration revealed a new role: clinicians were needed to handle edge cases, communicate diagnoses to patients, and apply ethical judgment the model could not replicate.
That same year, the FDA cleared IDx-DR — the first AI device authorized to provide a clinical diagnosis without a clinician's involvement — for diabetic retinopathy screening in primary care offices. A primary-care physician with no ophthalmology training could now run a definitive screening. The bottleneck shifted from specialist availability to patient willingness to be screened.
Radiology was the first specialty AI colonized at scale, because its inputs are already digital images and its outputs are structured reports — both machine-readable. By 2020, the FDA had cleared more than 30 AI-based radiology tools. The workflow change was specific and documented.
Triage acceleration: Aidoc's intracranial hemorrhage detection tool, deployed at hundreds of hospitals, flags critical findings and automatically moves those studies to the top of the radiologist's worklist. A 2020 study in the American Journal of Roentgenology found this reduced time-to-treatment for stroke patients by 52%. The radiologist still reads every study; the AI determines the order.
Second-reader function: iCAD's ProFound AI for mammography, cleared by the FDA in 2019, acts as a second reader — raising sensitivity for invasive breast cancer by 8 percentage points while reducing radiologist reading time by 52.7% in published trials. The radiologist remains the signatory; AI handles volume.
Measurement automation: AI tools like Imagen Technologies' OsteoDetect automate bone density measurements and fracture flagging. Tasks that required 20 minutes of a radiologist's time per study now take seconds of automated processing, freeing attention for complex interpretation.
A 2022 report from the Harvey L. Neiman Health Policy Institute found that despite AI adoption, radiologist demand remained high — the technology increased throughput rather than replacing positions. Subspecialty radiologists increasingly spend more time on AI exception handling and quality assurance than on routine reads.
Traditional drug development takes an average of 12 years from target identification to FDA approval. AI is attacking multiple steps in that pipeline simultaneously.
AlphaFold2 (DeepMind, 2020): The most significant scientific event in structural biology in decades. AlphaFold2 predicted the 3D structure of virtually every known protein — over 200 million — with accuracy that had previously required years of X-ray crystallography per protein. Pharmaceutical researchers immediately used it to identify potential drug binding sites. Insilico Medicine used AlphaFold predictions to help design a drug candidate for idiopathic pulmonary fibrosis that entered clinical trials in 2021 — the first AI-designed drug molecule to do so.
BenevolentAI's baricitinib identification (2020): During the early COVID-19 pandemic, BenevolentAI's knowledge graph platform identified baricitinib — an existing rheumatoid arthritis drug — as a candidate COVID-19 treatment by mapping viral entry pathways against approved compound libraries. The analysis took 48 hours. The drug was subsequently authorized by the FDA for emergency COVID-19 use in November 2020, one of the fastest repurposing validations on record.
Roles being redefined: Medicinal chemists who spent careers manually synthesizing and testing compounds now supervise AI-proposed synthesis routes. Computational biologists increasingly orchestrate ML pipelines rather than running individual simulations. The volume of hypotheses expands enormously; the human role shifts to experimental validation and regulatory navigation.
The American Medical Association documented in 2022 that physicians spend nearly 2 hours on administrative tasks for every 1 hour of direct patient care. AI is attacking this ratio directly.
Ambient AI scribes — systems that listen to doctor-patient conversations and generate structured clinical notes in real time — are now in use at hundreds of health systems. Nuance (acquired by Microsoft in 2022 for $19.7 billion, largely for its Dragon Ambient eXperience product) reported in 2023 that physicians using DAX complete documentation during or immediately after appointments rather than after clinic hours, recovering an average of 45 minutes per day.
Prior authorization — the process of getting insurer approval before a procedure — consumed 16% of physician administrative time according to MGMA surveys. AI platforms from companies including Cohere Health use clinical NLP to auto-generate prior auth submissions from medical records, reducing submission time from hours to minutes and approval rates in documented pilots at several health systems.
Healthcare AI is not eliminating physician roles in the near term — it is eliminating the routine pattern-matching portion of those roles and expanding the exception-handling, ethics, and communication portions. Radiologists who understand AI outputs and can audit model failures will be more valuable than those who cannot. Clinical informaticists — who bridge medicine and machine learning operations — are among the fastest-growing roles in hospital systems.
In this lab you will explore how AI is reshaping specific healthcare roles. Consider the cases from Lesson 1 — DeepMind at Moorfields, Aidoc in radiology triage, AlphaFold2, BenevolentAI's drug repurposing, and Nuance DAX for clinical documentation.
Choose one healthcare role (e.g., radiologist, medicinal chemist, primary care physician, clinical informaticist) and analyze how AI is changing the day-to-day tasks, decision-making authority, and skills required in that role. The AI tutor will help you think through trade-offs and career implications.
In February 2017, JPMorgan Chase revealed that its Contract Intelligence (COIN) platform had automated the interpretation of commercial loan agreements — a task that previously required 360,000 hours of lawyer and loan officer time annually. COIN used machine learning to extract 150 data attributes from each document, doing in seconds what took teams of people all year. JPMorgan did not announce mass layoffs of lawyers. Instead, it announced a redeployment: legal professionals were shifted toward complex judgment work, regulatory strategy, and dispute resolution that required contextual reasoning the model could not provide. The firm simultaneously disclosed it was investing $10.8 billion in technology that year — a figure that has since exceeded $15 billion annually.
The COIN revelation was diagnostic: it showed that a significant fraction of high-skilled financial services labor had been consumed by high-volume document pattern-matching — a task AI handles cheaply and quickly. The actual legal judgment work — the reasoning that determined whether a contract clause created unacceptable risk — remained human.
By 2023, algorithmic trading accounts for approximately 60–73% of all U.S. equity trading volume, with some estimates by TABB Group reaching 80% during high-volatility periods. This is not a new development — quantitative hedge funds like Renaissance Technologies have used statistical models since the 1980s — but the sophistication has reached a qualitatively different level with deep learning.
Two Sigma and Man Group have publicly described deploying NLP models that parse earnings call transcripts, Federal Reserve statements, and news feeds for sentiment signals milliseconds before human traders can read the first paragraph. The competitive advantage of human qualitative analysis at speed has largely disappeared at the sub-second scale.
What changed for human traders: The Bureau of Labor Statistics documented a 13% decline in securities, commodities, and financial services sales agents between 2010 and 2022. Goldman Sachs reduced its cash equity trading desk from 600 traders in 2000 to roughly 2 by 2017 — replaced by 200 engineers managing trading algorithms. The surviving human roles shifted to strategy design, model validation, risk parameter oversight, and client relationship management.
Market microstructure roles grew: Simultaneously, quantitative researcher, market structure analyst, and algorithmic execution strategy roles expanded significantly. The Financial Industry Regulatory Authority (FINRA) reported a net increase in compliance technology roles of 22% from 2018–2023, partly driven by the need to audit AI trading systems.
Visa's AI fraud detection system processes over 500 million transactions per day across 200 countries, applying deep learning models that score each transaction in under 100 milliseconds. In 2022, Visa reported its AI saved clients approximately $27 billion in prevented fraud annually — a figure that would have been impossible with human analysts reviewing individual transactions.
Mastercard's Decision Intelligence network uses a neural network trained on over 75 billion transactions to generate a real-time risk score. In a documented 2022 case study, a major European bank reduced false-positive fraud alerts (which freeze legitimate transactions) by 50% after switching to AI scoring — directly improving customer experience while simultaneously catching more genuine fraud.
Fraud analyst roles: Rather than disappearing, fraud analyst positions shifted from reviewing individual flagged transactions to managing model performance — investigating cases where the AI was wrong (both false positives and false negatives), retraining models on new fraud patterns, and designing rules for edge cases the model handles poorly. The job became more technical and more strategic, but did not disappear.
Upstart, an AI lending platform, uses over 1,000 variables (vs. the traditional FICO model's handful) to underwrite personal loans. In a 2022 Consumer Financial Protection Bureau study, Upstart's model approved 27% more applicants than traditional models while experiencing 16% fewer defaults — demonstrating that the traditional human underwriting heuristic was leaving creditworthy borrowers underserved. However, the CFPB simultaneously opened inquiries into whether AI underwriting introduced new forms of discriminatory bias — a human oversight function that became critical.
Betterment launched the first mainstream robo-advisor in 2010. By 2023, robo-advisor assets under management globally exceeded $2.5 trillion (Statista). Vanguard's Personal Advisor Services — a hybrid human-AI model — manages over $280 billion, making it the largest robo-advisor by assets.
The impact on financial advisor employment was less severe than early predictions suggested. The Bureau of Labor Statistics projects personal financial advisor employment to grow 13% through 2032 — faster than average — because the robo-advisor segment predominantly captured first-time investors and small-account holders who had not previously used any advisor. The AI expanded the market rather than purely substituting advisors.
However, the skill profile changed substantially. Financial advisors at major wirehouses like Merrill Lynch and Morgan Stanley are now expected to use AI portfolio analytics platforms (Merrill introduced its AI advisor tool "Merrill Advisor Match" in 2022) and to spend proportionally more time on behavioral coaching and complex life-event planning that AI cannot navigate.
Finance is the sector with the longest AI history — and the clearest pattern: high-volume repetitive analytical tasks (loan document review, transaction scoring, routine portfolio rebalancing) automate quickly. Roles requiring contextual judgment, client trust, regulatory navigation, and AI model oversight expand. Finance professionals who combine domain knowledge with data fluency — able to interpret model outputs, catch model failures, and explain AI decisions to regulators — are consistently the most sought-after in the current market.
Financial services AI creates clear winners and losers at the task level — but the role-level picture is more complex. In this lab, you will analyze a specific finance career path and map which tasks AI is automating, which tasks remain human, and what new skills that role now requires.
Choose a finance role: loan underwriter, fraud analyst, equity trader, financial advisor, or compliance officer. Tell the tutor your choice and your initial analysis of what AI has changed about it.
In 2012, Amazon acquired Kiva Systems — a warehouse robotics company — for $775 million. At the time, Amazon's fulfillment centers employed humans to walk miles of aisles picking individual products. By 2023, Amazon had deployed over 750,000 robots across its fulfillment network under the Amazon Robotics brand. The robots move entire shelving pods to human pickers, who now stand in place and pick items that arrive at them — eliminating the miles of walking without eliminating the picking task itself.
Amazon's human fulfillment workforce also grew during this period, reaching approximately 1.5 million employees globally by 2023. But the nature of the roles changed. A 2021 independent analysis by the Strategic Organizing Center found that injury rates in Amazon's robotic facilities were higher than in non-robotic facilities — because the pace of work is set by the robotic system's throughput, not by human sustainable rates. The automation changed the productivity floor, which changed the working conditions, which changed the workforce profile that could tolerate those conditions.
In 2017, Elon Musk declared Tesla's Fremont factory would become "the machine that builds the machine" — a fully automated manufacturing system. By mid-2018, Tesla's Model 3 production was severely behind schedule. Musk publicly acknowledged the problem: over-automation. The Gigafactory had deployed automated conveyors and robotic assembly for steps that required human adaptability — a specific flawed conveyor system had to be disassembled and a human workforce substituted.
By Q3 2018, Tesla hit its 5,000-vehicle-per-week production target — by adding human workers back to the assembly process. The documented lesson: robotics at the time could not reliably handle the variability in parts orientation and assembly sequencing that experienced human hands manage intuitively. The AI-controlled robotic cells required tightly controlled input conditions that real manufacturing environments do not always provide.
Tesla's subsequent approach became hybrid: AI-driven quality inspection cameras at every assembly station (replacing manual quality checkers), robotic spot welding (replacing human welders in specific high-volume, high-precision tasks), and human assembly workers retrained as robot cell operators and maintenance technicians.
Traditional manufacturing quality control relied on human visual inspection — sampling perhaps 1–5% of production for defects. AI computer vision systems now enable 100% inspection at line speed.
Landing AI's manufacturing deployments: Andrew Ng's Landing AI has deployed AI visual inspection systems at electronics manufacturers (including Foxconn), automotive suppliers, and food producers. In documented deployments, the systems detect defects missed by human inspectors — specifically catching subtle surface anomalies in circuit boards at sub-millimeter scale that human fatigue causes inspectors to miss after the first 30 minutes of a shift.
Foxconn's lights-out production: Foxconn — Apple's primary iPhone manufacturer — has been converting specific production lines to "lights-out" automation since 2016. By 2020, its Kunshan facility reported replacing 60,000 human workers with robots in assembly tasks, while redeploying remaining workers to quality control, robot maintenance, and logistics coordination. CEO Terry Gou stated publicly that 80% of Foxconn's manufacturing work could eventually be replaced by automation — a statement that prompted significant workforce transition planning in the regions where Foxconn operates.
UPS's ORION (On-Road Integrated Optimization and Navigation) system uses AI to optimize delivery routes for its 66,000+ drivers. By 2017, UPS reported ORION saved 100 million miles of driving annually, reducing fuel consumption by approximately 10 million gallons and cutting CO2 emissions by 100,000 metric tons. The saving per driver is roughly 8 minutes per day — but at UPS's scale, that equals approximately $300–400 million in annual savings.
The driver role changed: ORION provides optimized routes, and drivers who deviate from suggestions are flagged for performance review. The human judgment about route sequencing — which experienced UPS drivers had developed over years — was substantially automated. The remaining human value in delivery is customer interaction, package handling in complex environments (stairs, locked buildings, unusual addresses), and problem resolution when the route assumption is wrong.
Maersk's AI supply chain platform: The world's largest container shipping company deployed AI demand forecasting and container allocation optimization beginning in 2019. Documented outcomes included a reduction in empty container repositioning (a major cost) of approximately 20% in specific lanes, by predicting demand mismatches weeks in advance. Supply chain analyst roles at Maersk increasingly involve AI model validation and exception management rather than manual demand forecasting.
Predictive maintenance AI — systems that analyze vibration, thermal, and acoustic sensor data to predict equipment failure before it occurs — is one of the most commercially deployed AI applications in manufacturing. Siemens, GE, and Honeywell have all published documented ROI cases. The role of maintenance technician has shifted from reactive repair (fix it when it breaks) to predictive response (investigate the anomaly the AI flagged). The skill profile now requires data interpretation alongside mechanical expertise.
Manufacturing and logistics automation follows a consistent pattern: AI and robotics handle repetitive, high-volume physical tasks with predictable inputs first. Tesla's failed full-automation attempt at Fremont shows the current limits — variability, adaptability, and judgment in unpredictable environments remain largely human advantages. The growing roles are at the human-machine interface: robot cell operators, AI system technicians, computer vision QA specialists, and supply chain data analysts who can validate and override AI recommendations.
The lesson showed that manufacturing and logistics automation is well underway — but that full automation has real limits (Tesla's Fremont failure) and often changes workforce composition rather than eliminating it entirely (Amazon's workforce grew while deploying 750,000 robots).
Your task: choose a manufacturing or logistics role and analyze the automation boundary — specifically, what tasks AI and robotics now handle, what humans must still do, and what new hybrid skills are emerging in that role. Consider cases from the lesson: Amazon Robotics, Tesla Gigafactory, Foxconn's lights-out lines, UPS ORION, or Maersk supply chain AI.
In 2014, the Associated Press partnered with Automated Insights to deploy Wordsmith — a natural language generation platform — to automatically produce quarterly earnings reports for publicly traded companies. By 2016, AP was generating 3,700 financial stories per quarter, up from approximately 300 written manually by journalists. The volume increase was roughly 12x. AP leadership described the arrangement as "automation augmenting journalism" — freeing reporters from routine data transcription to pursue investigative and feature work.
The outcomes were more complicated than the press releases suggested. AP did not lay off financial journalists immediately, but the role of financial beat reporter changed permanently — routine earnings coverage became an AI function, and human reporters who remained in that specialty had to demonstrate differentiated value through analysis and sourcing that the template-based system could not provide. The automation also revealed that much of what appeared to be skilled journalism was actually structured data formatting dressed in prose.
The 2022 release of ChatGPT accelerated AI adoption in media at a pace that caught newsrooms unprepared. CNET became a high-profile case study in January 2023, when The Verge and Futurism reported that CNET had been publishing AI-generated financial explainer articles — under a byline reading "CNET Money Staff" — since November 2022. A subsequent review of 77 AI-generated articles by CNET's own editors found that a majority contained factual errors, including incorrect interest rate compounding examples in personal finance articles. CNET paused the program.
The Sports Illustrated AI scandal emerged in November 2023: Futurism reported that SI had published product review articles attributed to fake bylines — complete with AI-generated author portraits and biographies — that appeared to use AI-generated text. The Arena Group, SI's parent company at the time, acknowledged the issue and the editor-in-chief resigned.
What actually works in AI journalism: The Washington Post's Heliograf system — deployed since 2016 — generates structured sports and election results stories (e.g., "The [team] defeated [opponent] 24-17 in [stadium] on [date]") without factual controversy because they are template-constrained, not generative. Reuters' Lynx Insight tool flags statistical anomalies in financial data for human journalists to investigate — AI as a research assistant rather than author. These constrained applications have clean track records.
Legal document review — the process of examining thousands to millions of documents in discovery to find relevant evidence — has been transformed by AI. Before AI-assisted review (eDiscovery), litigation teams billed enormous hours for junior associates to read individual documents. By 2022, AI-powered eDiscovery platforms from Relativity, Logikcull, and Everlaw were standard in major litigation.
The McKinsey 2023 survey finding: McKinsey's 2023 State of AI report found that legal was among the functions with the highest AI adoption rates, with document review and contract analysis as the primary use cases. The survey found that law firms using AI review tools in discovery completed first-pass document review 60–70% faster than traditional methods.
Specific documented case — Allen & Overy: In February 2023, magic circle law firm Allen & Overy announced a partnership with Harvey, an AI platform built on GPT-4, deployed across its 43 offices globally. The firm described Harvey as being used for drafting, research, and document analysis — not replacing lawyers but described as equivalent to "adding a junior associate to every lawyer's team." This framing has become standard in large law firm AI announcements.
The Mata v. Avianca case (2023): A cautionary case in which attorney Steven Schwartz submitted a brief to federal court containing citations to six non-existent cases — hallucinated by ChatGPT — without verifying them. Judge P. Kevin Castel sanctioned the attorneys involved. The case became widely cited as evidence that AI legal tool usage requires the same professional verification responsibilities as all legal work product — AI does not transfer the lawyer's duty of care to the model.
In April 2023, McCann Japan's AI creative director "AI-CD β" — developed in 2016 using historical campaign data — won a Bronze Spike at Cannes Lions for a Clorets gum ad concept. It was the first AI creative to win at Cannes. The human team directed, produced, and refined the output, but the concept originated from the AI system.
By 2023, WPP, Publicis, Omnicom, and Dentsu — the four largest global advertising holding companies — had all announced proprietary generative AI platforms for client campaign production. WPP's partnership with Nvidia and Adobe targets AI-generated visual advertising assets. Publicis's CoreAI initiative involves retraining all 100,000+ employees on AI tools.
Copywriting: Jasper AI (formerly Jarvis) reported over 70,000 business subscribers generating marketing copy as of 2022. The impact on freelance copywriting rates and volume has been significant — freelance platforms including Upwork reported declines in short-form copywriting job postings of 21% from early 2023 to mid-2023 as businesses shifted to AI-assisted generation for routine marketing content.
What expanded: AI prompt engineering, creative direction, brand strategy, and content quality assurance became growth areas within marketing departments. The volume of content an AI-enabled small team can produce increased by 10–20x — but the brand judgment about what to produce and what to suppress remains deeply human.
A September 2023 MIT Sloan study by Fabrizio Dell'Acqua et al. found that Boston Consulting Group consultants using GPT-4 completed 12.2% more tasks, 25% faster, with 40% higher quality scores versus controls — on tasks within the model's capability range. Critically, on tasks outside the model's capability range, AI-using consultants performed 19% worse — because they trusted incorrect AI outputs. The study's core finding: AI raises the floor of performance for everyone but can lower the ceiling for uncritical users.
The white-collar AI disruption is real but uneven. Template-constrained, high-volume content generation (earnings summaries, sports scores, product descriptions) automates cleanly. Complex judgment — legal reasoning, investigative journalism, strategic brand decisions, client negotiation — remains human-dependent. The MIT BCG study's warning is critical: uncritical AI use can make skilled workers worse, not better. The premium increasingly goes to professionals who can both use AI effectively and identify when not to trust it.
The MIT BCG study's core warning: AI raises the performance floor but can lower the ceiling for uncritical users. The CNET and Mata v. Avianca cases show what happens when AI outputs go unverified in high-stakes contexts. But the AP, Washington Post Heliograf, and Allen & Overy cases show AI delivering genuine productivity gains.
In this lab, choose a knowledge work role in media, law, marketing, or consulting. Analyze the trade-off: where does AI genuinely help, where does uncritical AI use create risk, and what verification and judgment practices should professionals in that role adopt?