AI and Your Career

Almost every job involves some reading, writing, analysis or communication, which is exactly what current AI is good at helping with. Understanding what the evidence actually shows, rather than the loudest predictions, helps you make sensible choices about what to learn and how to work.

Think in tasks, not jobs

A job is a bundle of tasks. A paralegal researches, drafts, files, schedules and talks to clients. AI can speed up some of those tasks a great deal, a few only a little, and others not at all. Economists call a job exposed to AI when a meaningful share of its tasks could be done or assisted by it. Exposure does not mean the job disappears; it means the job is likely to change.

Two outcomes are possible for any exposed task:

• Augmentation: AI helps a person do the task faster or better, and the person stays in charge.
• Automation: AI does the task with little or no human involvement.

An analogy: power tools did not eliminate carpenters, but they changed which skills mattered, how much one carpenter could produce, and how many were needed for some jobs.

What the research says

The evidence is still early and the studies measure different things, so modest claims are the honest ones.

These findings fit together better than they first appear. The overall labor market had not been upended as of late 2025, but effects show up in specific places: some tasks get much faster, newer workers can benefit most from AI help, and entry-level hiring in some exposed fields appears to be weakening. Employer surveys are forecasts, not measurements.

Worked example. Suppose a support agent resolves 2.0 tickets an hour. A 14% improvement brings that to 2.28, about one extra ticket every 3.5 hours. For a newer agent at 1.5 tickets an hour, a 34% improvement brings them to about 2.0, roughly the level of the experienced colleague. That narrowing of the gap between novices and experts is good for new workers, but it also means the experience premium in some roles may shrink.

Skills worth building

• AI literacy: knowing how models work, where they fail, and what they cost. The Beginner lessons in this academy cover this.
• Verification and judgement: checking claims, spotting errors and deciding when output is good enough. As AI produces more first drafts, reviewing becomes a bigger part of many jobs.
• Domain expertise: deep knowledge of your field is what lets you catch subtle mistakes and ask the right questions.
• Problem framing and communication: describing a task clearly, with context and constraints, is the core of good prompting and of good delegation to people.
• Workflow design: seeing which steps of a process to hand to AI, which to keep, and where a person must sign off.
• Data handling: basic spreadsheet and data skills remain widely useful, and are needed to evaluate AI-produced analysis.
• Human-centered skills: the WEF survey lists creative thinking, resilience, flexibility, curiosity and lifelong learning among the fastest-rising skills alongside AI and big data, networks and cybersecurity, and technological literacy.

Using AI at work responsibly

• Follow your organization's policy. Use approved tools and ask if there is no policy.
• Protect confidential data. Do not paste customer data, personal information or trade secrets into tools that are not approved for it.
• Verify before you rely. You remain responsible for anything you submit, whoever drafted it.
• Be open about AI use where colleagues, clients or rules expect it.
• Watch for bias when AI touches decisions about people, and keep a human decision-maker.
• Mind intellectual property. Check terms of use, and remember that purely AI-generated material may not be protected by copyright in some countries.
• Be careful with agents. An AI that can act on your behalf can be misled by content it reads; keep approvals on for sending, spending and deleting.

A learning path through this academy

1. Foundations: start with What Is AI?, How Machines Learn and How LLMs Work.
2. Daily use: Prompting Fundamentals, Context Windows, Hallucinations and Verification, Choosing a Model and Privacy and Your Data.
3. Responsible practice: this track, especially AI Security Risks, Bias and Fairness and AI Law and Regulation.
4. Going deeper, if your work involves technology: APIs, Tokens and Pricing, Embeddings and RAG, AI Agents Explained and Evaluating AI Systems.
5. Optional, for privacy or cost reasons: the Running AI Locally track, starting with Why Run AI Locally.

Going deeper

How researchers measure exposure

Exposure studies break jobs into tasks and ask, task by task, whether AI could substantially help. A widely cited example is Eloundou, Manning, Mishkin and Rock (2023), "GPTs are GPTs". Using the US O*NET database of occupations and tasks, they rated whether access to a language model, or to software built on one, could cut the time needed for a task by at least half at the same quality. They estimated that about 80% of the US workforce could have at least 10% of their tasks affected, and about 19% could see at least half of their tasks affected, with higher-wage, information-heavy jobs generally more exposed. Note what this measures: technical potential, not what employers will actually do, how fast, or whether jobs grow or shrink.

Your own exposure, weighted by time

The same method works for one person. List your tasks, the hours each takes per week, and a realistic speed-up factor $s$ with AI help ($s = 1$ means no help; $s = 2$ means twice as fast, including checking). Your time-weighted exposure is the share of hours spent on tasks with $s > 1$.

Amdahl's law: why the overall gain is smaller than it looks

A principle from computer engineering, Amdahl's law, gives the overall speed-up when only part of the work gets faster. If a fraction $f$ of your time is sped up by a factor $s$, the remaining time is $(1-f) + f/s$, so the overall speed-up is

$S = \frac{1}{(1 - f) + f/s}.$

Worked example. If 40% of your week ($f = 0.4$) is drafting that AI makes twice as fast ($s = 2$): $S = 1/(0.6 + 0.2) = 1.25$. Your week shrinks by 20%, not 40%. Even if those tasks became instant ($s \to \infty$), $S = 1/0.6 \approx 1.67$. Checking costs time too: if verification adds 5% of your original week, $S = 1/(0.6 + 0.4/3 + 0.05) \approx 1.28$ for $s = 3$, instead of 1.36 without checking. The parts of a job AI does not touch, such as meetings, relationships, judgement calls and physical work, set the ceiling. This is one reason studies find task-level gains much larger than job-level or economy-wide effects.

tasks = [  # (task, hours per week, speed-up with AI; 1.0 = no help)
    ("Drafting reports",        10, 2.5),
    ("Client meetings",          8, 1.0),
    ("Data clean-up",            6, 3.0),
    ("Email and scheduling",     6, 1.5),
    ("Reviewing others' work",  10, 1.2),
]
total = sum(h for _, h, _ in tasks)
new_total = sum(h / s for _, h, s in tasks)
exposed = sum(h for _, h, s in tasks if s > 1)
print(f"time-weighted exposure: {exposed / total:.0%}")
print(f"hours: {total} -> {new_total:.1f} (overall speed-up {total / new_total:.2f}x)")

Here 80% of hours are exposed, yet the week falls from 40 to about 26.3 hours, a 1.52x speed-up, because meetings are untouched and review only speeds up modestly. Use your own numbers, and estimate $s$ from real trials rather than hopes.

The jagged frontier

AI capability is uneven in ways that are hard to predict: a model can draft a strong strategy memo and then fumble a simple-looking question. Dell'Acqua and colleagues (2023) studied consultants at Boston Consulting Group working on realistic tasks. On tasks inside the AI's capability "frontier", consultants using GPT-4 completed more work, faster, and at higher rated quality. On a task deliberately chosen to fall outside the frontier, consultants using AI were less likely to reach the correct answer than those working without it, because the AI's confident but wrong analysis was persuasive. The practical lesson: learn where the frontier lies for your own tasks by testing, and apply the most scepticism exactly where the output sounds most plausible.

What the productivity studies imply for skill-building

Several experiments point the same way: the customer-support study (Brynjolfsson, Li and Raymond) found the largest gains for less experienced workers, and a controlled writing experiment by Noy and Zhang (2023) found that professionals completed writing tasks faster and with higher rated quality using ChatGPT. Together with the jagged frontier, this suggests a strategy: use AI to get up to speed faster, but build the domain judgement that lets you recognise when it is wrong. That judgement is what employers can least easily replace.

Make your skills visible

Because AI makes polished output cheap, evidence of judgement becomes more valuable. Keep a short record of concrete results: a process you redesigned, hours saved (measured, not guessed), errors you caught, an evaluation set you built. For technical roles, a small public project (for example one of this academy's capstones) shows more than a list of tools.

Common pitfalls

• Equating exposure with job loss. Exposure measures what tasks AI could affect, not what employers will do.
• Overestimating your speed-up. Amdahl's law and checking time cap the overall gain; measure before promising.
• Trusting AI most where it sounds best. The jagged frontier means fluent answers on unfamiliar ground deserve the most scepticism.
• Letting fundamentals slide. If AI does every first draft, you can lose the skill needed to judge drafts; keep practising core tasks yourself.
• Hiding AI use where your employer, clients or rules expect disclosure.
• Chasing tool names. Products change every few months; categories, evaluation habits and domain knowledge last.

In practice

• List the main tasks in your job and mark each as likely to be automated, augmented or unaffected.
• Pick one augmented task and practice doing it with AI for two weeks, checking every output.
• Strengthen the skills AI makes more valuable: verification, domain depth and clear communication.
• Keep learning in small, regular steps; the tools will keep changing, but the fundamentals in this academy change slowly.