This piece first appeared in German. The English version is a rewrite rather than a line-by-line translation, and the German original stays online in the archive: Was ist ein High Potential von morgen?.

In a field study of a customer support center, access to a generative AI assistant raised productivity by 14 percent on average. Among novice and low-skilled workers the gain was 34 percent. For experienced staff it was close to nothing.

Distribution matters more here than the average, and it is uncomfortable. The tool compresses the distance between someone in their first month and someone in their fifth year, and it compresses it on exactly the ground where the first five years used to be spent. Three in four knowledge workers were already using AI at work by the 2024 Work Trend Index from Microsoft and LinkedIn. That puts the old markers of talent, the grades, the logos, the mastery of routine, under a selection pressure they were never designed for.

The ladder is being pulled up from the bottom

AI takes speed into the zones where people used to learn the craft: repetitive research, standard analyses, summaries, contract and document checks. That is attractive on a cost line and it has a second face. Learning curves that consisted of repetition and correction get shortened, and the job of training moves from doing toward judging.

Audit is the case study, because the speed arrived before the measurement

Around three quarters of Deloitte UK's auditors use PairD, the firm's own assistant, according to reporting in Financial News. The time saved is real, and it forces the organization to secure the quality of its judgments by some other means than the one it has always used.

The Financial Reporting Council, the UK audit regulator, found in June 2025 that the six largest firms use automated tools and AI without formally monitoring how those tools affect audit quality. Productivity is present. Measurability has to follow.

Culture is the part a model does not inherit

Banking points the same way. Goldman Sachs rolled out a generative assistant across the firm, which makes the work faster and makes a harder question louder: how do you translate a culture into machine logic?

Marco Argenti, the bank's chief information officer, has said that making AI agents smart is not the hard part. Firms have to work out how to inject their cultural traits and leadership principles into them. An AI colleague learns facts quickly and does not automatically learn how a particular firm does things, which is what decides whether a recommendation is usable, defensible and effective.

Once the doing accelerates, judgment moves to the center, and an awkward question follows: how do you recognize a high potential now? The old answer, the stations and the grades and the logos, was never perfect, but it worked. What counts next is the ability to produce effect under uncertainty, through context, models, data criticism, clear language, integrity and the capacity to learn. The World Economic Forum's skills outlook points the same way: analytical thinking is the core skill employers value most, and AI and big data lead the fastest-growing skills, with creative thinking close behind.

What the good ones actually do

This stays abstract until you watch how strong people work. Strong people start with context, before tooling. The first question is what would have to be true for option A to be the best one. Assumptions go on the table, data gets checked for where it came from and how old it is, and the places where they distrust their own analysis get named out loud.

Their decision papers are written to fit the reader, and two pages of clear trade-offs beat ten slides of pictograms. Amazon institutionalized the principle: no PowerPoint, six-page narratives, read in silence at the start of the meeting. Writing forces clarity, and clarity is the raw material of good judgment.

Described operationally, these are people who structure complexity instead of decorating it. They carry a finding from one context into another without bending it. Where others reach for phrases, they speak in models, in causality and sensitivities and thresholds. Data gets handled critically, with sources, timestamps and stated uncertainty.

Their writing is decided, and it owns the options it puts forward. Red lines get named even at the cost of a short-term advantage. Learning happens visibly, in loops, with assumptions tested and discarded.

Selection has to make thinking visible

Unstructured conversations and brainteasers deliver very little. Anyone selecting seriously watches reasoning happen in real time: a live problem worked through out loud, a two-page decision paper with a recommendation, assumptions and risks, a work sample from the actual stack.

Then it gets measured, consistently. The audit debate is instructive well beyond audit. Wherever AI does the work, quality has to be defined, documented and checked, and without that a firm is producing speed with no warranty attached.

Four things a firm can do now

Strengthen the writing culture. Every decision of consequence starts with a short memo: problem, options, criteria, recommendation, assumptions, risks. It sounds bureaucratic and it is productive, because it forces precision and makes tacit knowledge checkable.

Rebuild the training. Where doing teaches less, learning has to be constructed: shadowing focused on how decisions get made, simulations with real target metrics, post-mortems that ask which assumption failed.

Operationalize AI governance with metrics that measure effect, such as error rates, rework, and the depth of the audit trail. And translate the culture into guardrails, because agents without them return answers that are correct and do not fit.

The evidence a candidate should carry

Anyone who wants to be a high potential needs a different portfolio. Start with a cleanly written decision memo. Then a one-page model on a hard question, or a transfer example of the kind that asks what a fundraising campaign can learn from SaaS pricing, and a short post-mortem correcting an assumption in public. That is the currency that convinces in an interview, because it shows how the person thinks.

The label only means something when it is observable

The term high potential is loaded, and it often labels early opportunity inside a narrow system. In an age of AI it acquires substance only if it becomes observable, which means living with the contradictions. AI does deliver acceleration. It does flatten differences in the doing, at first to the advantage of the younger. And that is precisely why judgment becomes the bottleneck.

What remains is craft and posture. Craft is the discipline of producing one page of prose before a decision, where ten pages of slides were the habit. Posture is a willingness to document assumptions and revise them. Both are learnable, and together they separate speed from effect.

This essay follows on from "Who teaches juniors to think", which set out the starting position; this one describes the consequence for talent. The next step is governance: who decides about talent tomorrow, when nobody knows precisely which future counts?

Sources

Grouped by the section they support.

Opening

  • Brynjolfsson, Li and Raymond (2023), Generative AI at Work, NBER Working Paper 31161, nber.org; published in the Quarterly Journal of Economics 140 (2), 2025. Supports the 14 percent average productivity gain, 34 percent for novice and low-skilled workers, and minimal impact on experienced staff.
  • Microsoft and LinkedIn (2024), 2024 Work Trend Index Annual Report, 8 May 2024, news.microsoft.com. Supports 75 percent of knowledge workers using AI at work.

Audit is the case study, because the speed arrived before the measurement

  • Financial News (2025), Deloitte triples number of auditors using AI chatbot, 8 April 2025, fnlondon.com. Supports the rise in PairD use among Deloitte UK auditors, reported elsewhere as from about a quarter to about three quarters.
  • Accountancy Age (2025), Big accountancy firms fail to monitor AI's impact on audit quality, says FRC, 27 June 2025, accountancyage.com, reporting on the Financial Reporting Council's thematic review of 26 June 2025, frc.org.uk. Supports the finding that the largest firms perform no formal monitoring of how automated tools affect audit quality.

Culture is the part a model does not inherit

  • Morris (2025), Goldman Sachs rolls out an internal AI assistant firm-wide, Fortune, 24 June 2025, fortune.com. Supports the firmwide launch of the GS AI Assistant.
  • Lee (2025), Making AI agents smart isn't the hard part — it's teaching them company culture, says a Goldman Sachs exec, Business Insider via Yahoo Finance, 26 March 2025, finance.yahoo.com. Supports Marco Argenti's view that firms must inject their cultural traits and leadership principles into AI agents.

Judgment is the signal that replaces the logo

  • World Economic Forum (2025), The Future of Jobs Report 2025, chapter 3, Skills outlook, 7 January 2025, weforum.org. Supports analytical thinking as the top core skill, and AI and big data as the fastest-growing skills.

What the good ones actually do

  • Bezos (2018), 2017 Letter to Shareholders, Amazon, aboutamazon.com. Supports Amazon's six-page narrative memos in place of slide presentations, read at the start of meetings. The description of how strong people work is the author's own.

All other sections

  • No external source. The recommendations on selection, training and governance draw on the author's professional experience.
The link has been copied!