Generative AI is fluent by design. It will answer almost any question in polished prose, whether or not the answer is true. That makes it most dangerous in the hands of someone who cannot tell the difference, and most useful to someone who can. Three capabilities separate the two: deep knowledge of a field, the habit of testing what the machine says, and the skill of bringing other people along.

The clearest test comes from a preregistered experiment with 758 consultants at Boston Consulting Group. On tasks inside what the researchers call AI's “jagged frontier,” consultants using GPT-4 completed 12.2% more tasks, 25.1% faster, and their work was rated about 32% better on average.¹

Then the researchers set a task outside the frontier: a business problem designed so that the AI, given the case materials, would reach the wrong conclusion. Working without AI, 84.5% of consultants reached the right answer. With AI, 70.6% did, and among those also given training in how to prompt it, only 60%.¹ These were, in the authors' words, “highly specialized knowledge workers working within their domain of expertise.”¹

A further finding matters as much. In the authors' words: “AI-supported participants produced recommendations that were more coherent and persuasive, underscoring that AI improves presentation and argumentation even when the underlying analysis is flawed.”¹ A fluent mistake can be harder to catch than a clumsy one. The lesson is not to avoid AI. It is to know enough to overrule it, and to check.

AI does not only make mistakes. It makes them persuasively, and people who should know better pass them on. Exhibit 2 lists cases in which a confident AI answer cost someone money, a ruling or a reputation.

Someone is keeping count. A public database compiled by the legal researcher Damien Charlotin listed 2,097 court decisions involving AI-invented material as of September 30, 2026: fake cases, fake quotations, fake citations. Self-represented litigants account for 58% of them and lawyers for 40%. Judges account for 33 decisions.¹¹

The quieter risk is misplaced trust. In a survey of 319 knowledge workers by researchers at Carnegie Mellon and Microsoft, “higher confidence in GenAI is associated with less critical thinking, while higher self-confidence is associated with more critical thinking.”¹² Workers more confident in their own abilities, in other words, reported checking more. The authors note that the survey cannot show cause and effect.

An early study from MIT's Media Lab points the same way. Of 18 participants who wrote an essay with ChatGPT, 15 could not correctly quote a sentence from the essay they had just written, against 2 of 18 in each of the comparison groups. The gap narrowed in later sessions, and the study is a small preprint that has not yet been peer-reviewed.¹³

The habit can also run the other way, with the tool questioning the person. McKinsey's global managing partner, Bob Sternfels, built an AI agent he calls Contra. Every talk he gives goes to Contra first, with two questions: “What am I missing? What's the other argument to this?”¹⁴

An answer becomes a decision only when people agree to act on it. Cornell's committee makes the point directly: “People must be good at deciding what to do and why. They will need to figure out how to negotiate those whats and whys with other people.”²

In consulting, winning that agreement can take most of a project. Luis Garicano, an economist at the London School of Economics and a co-author of the 2026 book Messy Jobs: The Work That AI Cannot Reach, describes a consultant redesigning a large airline's loyalty program: “His team finished the analysis in two weeks. Months later the program still does not exist.” The months went to the parties with something to lose: departments inside the airline, and partners such as hotel chains and credit-card companies. “Once the analysis is cheap,” Garicano writes, “what remains is to learn what each party will actually accept, and to obtain commitments from those who are authorized to make them.”¹⁵

Some employers now hire for it directly. McKinsey puts “more of a premium on working with other people,” Sternfels told the New York Times in October 2026. One marker the firm looks for is success in a team sport. Another is a retail job as a teenager, where “you don't have positional authority. You have to deal with people.”¹⁴

Employers say graduates arrive short on several of these skills. In surveys by the National Association of Colleges and Employers, 78.1% of students rated themselves very or extremely proficient in communication, but only 53.5% of employers rated new graduates that highly. The gaps are wider for leadership and professionalism and close to zero for teamwork. On technology, employers rate graduates more highly than graduates rate themselves.¹⁶

The labor market has paid for these skills for decades. Between 1980 and 2012, jobs requiring high levels of social interaction grew by nearly 12 percentage points as a share of U.S. employment, while math-intensive jobs that needed little social skill shrank by 3.3 points.¹⁷ Pay followed. Real wages rose about 26% in jobs that demanded both math and social skills, against about 5.9% in math-heavy jobs with little social content.¹⁷

Picture most professions as a curve. Most practitioners sit in a broad middle of competent work. A smaller group sits far out on the right, and it accounts for a disproportionate share of what gets done. A study of 633,263 researchers, entertainers, politicians and athletes found performance in most of its 198 samples far more top-heavy than a bell curve. Applied to academic researchers, that shape implies the top 1% produce 10% of the output and the top 5% produce 26%.¹⁸ Other researchers argue that in many cases the extreme skew is an artifact of how performance is measured, and that better measures produce a curve closer to normal.¹⁹

This briefing is about what puts people in that tail, and what AI does to everyone else. In work that serves people, this briefing argues, the best are set apart by their attention to the person in front of them: the patient, the guest, the client. The examples below are suggestive rather than proof. Exhibit 6 shows the shape the rest of this briefing uses.

Medical knowledge is the entry ticket. A study published in JAMA suggests that what doctors do in the room matters too. Researchers recorded routine office visits with primary-care physicians who had no malpractice claims and with those who had two or more. The doctors with no claims spent longer with each patient, 18.3 minutes against 15.0. They told patients more about what to expect, and they laughed and used humor more.²⁰

The study was observational, and among surgeons it found no such differences.²⁰ For primary care, though, the association is striking: the doctors who took time to orient and engage their patients were the ones without a history of claims.

At a luxury hotel, checking a guest in is routine. In this briefing's view, what sets the best front-desk agents apart is what happens around it: noticing the exhausted guest, remembering the returning one, fixing a problem before the guest raises it. St. Regis has offered butler service, introduced by John Jacob Astor IV at The St. Regis New York, for more than a century.²¹

The Ritz-Carlton, a sister brand within Marriott, turned the same idea into a rule. Under co-founder Horst Schulze, every employee, down to the dishwasher, had up to $2,000 to spend on a guest's problem, according to a profile in Chief Executive.²²

Some companies write the trait into their rules. Amazon lists “Customer Obsession” first among its 16 Leadership Principles: “Leaders start with the customer and work backwards. They work vigorously to earn and keep customer trust.”²³

Employers across industries rank the same qualities. In the World Economic Forum's Future of Jobs Report 2025, “empathy and active listening” and “service orientation and customer service” both make the list of the ten core skills employers value most, with analytical thinking at the top.²⁴

AI can write the words. In a 2023 study in JAMA Internal Medicine, licensed health-care professionals compared chatbot and physician answers to 195 questions patients had posted online. They preferred the chatbot's answer 78.6% of the time. They rated its answers good or very good in quality 78.5% of the time, against 22.1% for physicians, and empathetic or very empathetic 45.1% of the time, against 4.6%.²⁵

That is the distinction this briefing draws. Empathy in a reply is a sentence. Empathy at the top of a profession is an act: noticing what the other person needs, deciding what to do about it, and answering for the result. No chatbot notices the guest at the desk, or stands accountable for a patient's recovery.

The best evidence on what AI does to work comes from studies that compared workers with and without AI, then measured the difference. They span customer support, management consulting and professional writing. On one point they agree: AI helps the least skilled the most.

That is good news for anyone starting out, and a warning for anyone planning a career. When most of a profession can turn out the same competent work with the same tool, that work stops being scarce. The bottom 80% of the curve in Exhibit 6 starts to look alike.

Why would AI help the weakest most? The customer-support study is the clearest case. The researchers followed 5,172 agents at one company as an AI assistant was rolled out, and suggest the tool worked “by exposing lower-skill workers to the best practices of higher-skill workers.”²⁶ In effect, AI turned yesterday's best practice into today's baseline.

The assistant learned from the company's past chats, and its developer gave extra weight to conversations handled by top performers.²⁶ The least-skilled fifth of agents gained 0.53 resolved issues per hour, a 36% increase. The gains shrank at every step up the skill ladder, to nothing measurable for the most skilled (Exhibit 11).²⁶ Experience showed the same pattern: agents in their first month gained 0.71 issues per hour, while those with more than a year on the job gained essentially nothing. With AI, agents with two months on the job performed as well as agents without it who had more than six.²⁶

For the most skilled agents there was little left to teach. AI added no significant productivity, and their resolution rates and customer satisfaction slipped slightly.²⁶ Whatever made them the best was not something the assistant could hand back to them.

Economists have a name for what is hard to copy. In 1966 the philosopher Michael Polanyi observed, “We can know more than we can tell.” MIT economist David Autor called this Polanyi's paradox: “engineers cannot program a computer to simulate a process that they (or the scientific community at large) do not explicitly understand.”²⁸ Machine learning gets around part of the paradox by learning from examples instead of rules. That is how the support assistant picked up top agents' habits. What it cannot learn is what never made it into the examples: the read of one customer in one moment, a relationship built over years, and the responsibility for what happens next.

On routine enough work, AI does not just help. It can do the job. In the writing study, a third of the professionals given ChatGPT said they submitted its first draft without editing it, and those who said they edited were active on the task for only 3.3 minutes on average. The ChatGPT users' work scored no higher on average than ChatGPT's raw output.²⁹ The authors leave open whether that means AI will displace such workers or augment them.²⁹

The machines are also getting better at the middle of professional work. GDPval, a benchmark OpenAI built from real assignments in 44 occupations, asks industry experts to grade AI deliverables against those of professionals with an average of 14 years' experience. In September 2025, the best model's work was graded as good as or better than the professional's 47.6% of the time. By April 2026, OpenAI reported 84.9% for GPT-5.5, its newest model at the time.^30,31

The benchmark's limits matter as much as its scores. Its tasks are, in OpenAI's words, “precisely-specified and one-shot,” and tasks involving “extensive tacit knowledge” or “communication between individuals” are out of scope.³⁰ The work it measures is the kind AI increasingly matches. The work it leaves out is the work this briefing argues will hold its value.

The labor market is moving the same way, starting with the youngest workers. Stanford economists tracking payroll records found that employment of 22- to 25-year-olds in the most AI-exposed occupations fell about 11% between November 2022 and June 2026. In the least-exposed occupations, employment for the same age group grew about 10%.³² Measured against where it would be had it kept pace with less-exposed peers, young workers' employment in AI-exposed jobs stands 19% short.³²

Experienced workers in the same occupations show no comparable gap, and the shortfall comes mostly from less hiring rather than more departures.³² The authors call these early, descriptive indicators rather than causal estimates.³² But the direction matters for anyone choosing a path. The declines are concentrated where AI substitutes for human tasks rather than complementing them.³² Some of the stakeholders Cornell's committee consulted gave the same warning: “when AI can cheaply replace an intern,” graduates “must attain a higher level of experience than before to even get in the door.”²

Economists have a theory of why the first rung moves first. Apprenticeships, as Garicano and Kellogg's Luis Rayo put it in a 2025 working paper, “let juniors pay for training by doing menial work.” Rayo explains the problem: “Now that AI does that [work] essentially for free, the currency that apprentices are using to buy knowledge is vanishing.” In their model, AI pushes juniors onto harder work from the start. Apprenticeships remain safe as long as a fully trained worker using AI produces more than about 2.7 times what the AI produces alone. Below that line, training gets shorter, and “advances in AI threaten wholesale apprenticeship collapse.”^33,34 It is a model, not a measurement, and the payroll data above cannot show cause, but the two point the same way.

What is left for juniors is the part of the job they used to learn last. For knowledge workers, MIT Sloan labor economist Anna Stansbury suggested at a Peterson Institute discussion of Messy Jobs in September 2026, “the job has become a management job even at the entry level.” Someone still has to define the task, set the goals and delegate the work, then review it, handle the edge cases and judge its quality; what machines can now do is the execution in between. Juniors used to learn that judgment by doing the execution. The open question, Stansbury said, is how to build a system in which a newcomer can “come straight in” and do the goal definition and quality evaluation: “I would call those sort of management skills.”³⁵

Who is using AI shapes what it does. Anthropic, which makes the Claude models, classifies a sample of its own traffic. In February 2026, 53% of conversations on Claude.ai were augmentation, where people worked with the AI by learning, iterating or checking their own work, and 44% were automation, where they handed it a task to complete. Among businesses and developers using Claude through its API, automation was 68% and augmentation 17%. That API traffic includes Claude Code, Anthropic's coding tool, which accounts for a large share of it.³⁶

Klarna, the payments company, shows what automation looks like at scale. In February 2024 it said its AI assistant was handling two-thirds of customer-service chats, “doing the equivalent work of 700 full-time agents.”³⁷ About 14 months later its chief executive said cost “seems to have been a too predominant evaluation factor,” and “what you end up having is lower quality.”³⁸ Klarna now describes a “dual-track approach.” Its AI assistant handled 80% of chats in 2025, and every customer can still choose a person.³⁹ Klarna kept AI for most conversations and kept a person available for any customer who wants one. That line, between the machine's work and the person's, is where Part IV begins.

Exposure to AI is not the same as replacement. Goldman Sachs economists estimated in 2023 that roughly two-thirds of U.S. jobs are exposed to some degree of automation by AI, and that generative AI could take over as much as a quarter of current work. Their illustrative assumptions were consistent with 63% of U.S. employment being complemented by AI, 7% substituted and 30% unaffected.⁴⁰

The Stanford payroll study points the same way. Where AI “primarily complements workers, employment is flat or rising, especially for experienced workers.”³² The careers in this part all end in a moment of trust between people. That, this briefing argues, is why AI tends to make the people who do them more productive rather than unnecessary.

Most business buyers say they would rather buy without a sales rep. In a Gartner survey of about 645 B2B buyers in August and September 2025, 67% said they prefer a rep-free experience. Yet 69% prefer to validate AI-generated insights with sales reps, and reps remain buyers' most important source of information when they choose a preferred supplier, secure internal support and finalize the purchase.^41,42

Complex purchases are made by groups. Gartner's 2025 research found buying groups “ranging from five to 16 people across as many as four functions.”⁴³ Earlier, it found purchase regret 23% higher among customers who preferred to buy without a rep than among those who worked with one.⁴⁴ Reading a room of 16 people with different agendas, and moving them to one decision, is not something a buyer can download.

Gartner forecasts that by 2030, 75% of B2B buyers will prefer sales experiences that put human interaction ahead of AI, which it says “could mean fewer but more specialized sales roles, with greater investment in training and interpersonal skills.”⁴⁵ If that holds, the sellers who keep their seats will be the ones who bring judgment to the buying group, not information it can find for itself.

The operating room already runs on advanced machines, with surgeons at the controls. Surgeons performed about 3.15 million procedures with Intuitive's da Vinci robotic systems in 2025, up from about 1.23 million in 2019, on an installed base of 11,106 systems.⁴⁶ The U.S. Food and Drug Administration is blunt about the label: “The device is not actually a robot because it cannot perform surgery without direct human control.”⁴⁷

Autonomy is arriving one step at a time. In 2025 a Johns Hopkins robot performed the clipping-and-cutting phase of a gallbladder removal on its own, on eight pig specimens outside a living body. It was slower than a surgeon, and a human assistant loaded clips and changed its tools.⁴⁸ Robots have begun to carry out narrow steps on living patients: a peer-reviewed study in May 2026 described an autonomous system placing dental implants in 11 patients.⁴⁹ Experts reviewing the field in July 2026 wrote that “the highest level of autonomy has not yet been achieved.”⁵⁰

Meanwhile AI is giving doctors time back with their patients. At Kaiser Permanente's Northern California medical group, 7,260 physicians used ambient AI scribes, which draft visit notes from the conversation, in 2,576,627 patient encounters over 63 weeks. The group estimates they saved 15,791 hours of documentation.⁵¹ In surveys, 84% of physicians said the scribes had a positive effect on their visits, and 56% of patients said the same about visit quality; none said their visit was worse.⁵¹ A randomized trial at UCLA found more modest time savings, so the gains vary by tool and setting.⁵²

On paper, law is among the most exposed professions. Goldman put 44% of legal tasks within AI's reach, second only to office and administrative support.⁴⁰ Firms are adopting it for exactly those tasks. In a Thomson Reuters survey of 1,514 professionals in late 2025, 41% of law firms said they already use generative AI. Among legal professionals using it, the top uses were legal research (80%), document review (74%) and summarizing documents (73%).⁵³

The profession's ethics guidance keeps a person in charge. The American Bar Association's 2024 ethics opinion warns that “lawyers' uncritical reliance on content created by a GAI tool can result in inaccurate legal advice to clients or misleading representations to courts and third parties.” Lawyers “may not leave it to GAI tools alone to offer legal advice to clients, negotiate clients' claims, or perform other functions that require a lawyer's personal judgment or participation.”⁵⁴

The courts have drawn the same line for themselves. Federal judges have been cautioned “not to delegate core judicial functions to AI, including decision-making or case adjudication.”⁵⁵ A proposed federal rule on machine-generated evidence, still under study, rests on a plain observation: “a machine cannot be cross-examined.”⁵⁶ The Bureau of Labor Statistics rates lawyers “very high” in AI exposure, yet projects their employment to grow 4.7% over the next decade, faster than the 3.5% average for all occupations.^57,58

Human contact alone is no shield; the line runs through each job, not between jobs. Microsoft's analysis of 200,000 Copilot conversations ranks sales representatives of services fourth of 785 occupations for AI applicability, customer service representatives seventh and concierges fourteenth.⁵⁹ What those scores capture is the information in the job: answering questions, explaining options, drafting messages. The authors caution that reading high scores as job loss “would be a mistake.”⁵⁹

The government's projections show where the information work is thinning. BLS expects customer service representatives to decline 5.3% by 2035, and cites AI tools in the sales process as one reason it expects sales jobs overall to shrink 1.4%. Lawyers, registered nurses and physicians, whose work ends in advice, care and accountability, are projected to grow at or above the average.⁵⁸

Messy Jobs gives that line a name. “A hard job is not a messy job,” Garicano told the same Peterson Institute audience. Chess and math are hard, but they are clean: single tasks where “you can actually give the computer feedback and say whether it is or it's not correctly done.” Clean work is where AI arrives first. A job is messy, the authors argue, when its tasks “must be combined under incomplete knowledge, conflicting objectives among the different parties and binding constraints on who has the authority to make decisions.”^35,15

Radiology is the classic test. In 2016 the computer scientist Geoffrey Hinton, a Turing Award winner, said on stage: “People should stop training radiologists now.”^60,61 In 2025, U.S. diagnostic radiology residency programs offered a record 1,208 positions, and radiology was the second-highest-paid medical specialty, averaging $520,000 a year.⁶¹ Reading an image is one task; the radiologist's job is the bundle around it. “A radiologist who has already spoken with the patient and reviewed the clinical record may be able to read the scan better than one who sees only the image,” Garicano writes.⁶²

The Boardroom is a five-day business program in Boston for high-school students. It is taught by the case method top business schools use, and led by practitioners with backgrounds in investment banking, private equity and management consulting. Students do not sit through lectures. In every session they make a call and defend it aloud while peers and instructors push back.

That format matters more as AI spreads. Cornell's Committee on the Future of the American University notes that in classes built on hands-on projects and cases, assessment is “more robust against the widespread availability of AI because students are demonstrating their capacities in action every time the class meets.”² A chatbot can write an essay. It cannot take a board's questions for a student.

Timing matters too. Part III shows entry-level employment falling behind in the jobs most exposed to AI, and stakeholders warning that graduates now need more experience than before to get in the door.^32,2 The Boardroom gives students their first real practice years before that door.