Who teaches juniors to think
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: AI ersetzt das Doing – aber wie lernen junge Berater noch das Denken?.
David Solomon, who runs Goldman Sachs, has put a number on it. An AI system can draft 95 percent of an S-1 prospectus in minutes, work that used to occupy a team of six for two weeks. What matters now, in his words, is the last 5 percent, because the rest has become a commodity.
Read that as a statement about documents and it is a productivity story. Take it as a statement about careers and it is something else, because the 95 percent is precisely the work that first-year bankers were given in order to learn the job. Consulting, audit, tax and law all work the same way. Whatever goes first is what the apprenticeship was made of.
The work juniors learned from is the work being automated first
Routine is disappearing across the professional services. Research, data analysis, benchmarking, financial modeling and document review now run through specialized systems, faster and often more accurately than a person managed.
Market analyses and benchmark slides that took days take minutes. In audit, standardized reconciliations of accounts and contracts increasingly run with AI support. Financial News reported that use of PairD, Deloitte UK's internal chatbot, tripled to around three quarters of its auditors. Tax practices prepare declarations, bookkeeping and reporting automatically. In law, tools such as Harvey and Kira Systems review documents, locate clauses and produce first drafts of an argument.
The craft that entry-level people used to be handed is being redistributed, and nobody planned what replaces it.
The apprenticeship was the training, and it is going
Young consultants built their skills through repetition and correction. Building endless spreadsheets taught logic, a feel for numbers, and what it costs to be wrong by one cell. Reading two hundred pages of due diligence taught how contracts are structured, how deep to go, and how an outlier looks when it appears.
When a machine takes the routine, that loop goes with it. Mistakes made, corrected and thought about were the mechanism, and the years of doing them were the training. The risk is a generation that arrives in senior roles faster and without the foundation those roles assume.
Judgment is what is left, and it is harder to teach
The center of the job moves from execution to interpretation. Four capabilities carry it.
Understanding context, which means reading data and text closely enough to say what they mean. Transfer, which means carrying a finding from one situation into another. Critical thinking, which means interrogating what a model returns instead of forwarding it. And communication, because no system holds a persuasive conversation with a client, a bank or a supervisory board.
Marco Argenti, chief information officer at Goldman Sachs, framed the same problem from the other side. Making AI agents smart is not the hard part, he has said; firms have to work out how to "inject" their cultural traits and leadership principles into them.
Machines supply content, and people supply meaning, significance and the cultural setting the content has to fit.
Every one of these professions keeps its last five percent
McKinsey and Boston Consulting Group have been running AI-supported research platforms for a while. Value still forms in the synthesis, in deciding which numbers matter and how they turn into a strategy.
Tax advice lives on precision and routine, both of which automate well. Clients increasingly want a trusted adviser: someone who reads the numbers, shows where the room to maneuver is, and carries part of the decision. AI creates the time for that work without doing it.
Large law firms have had Harvey and its competitors in contract review long enough to know the efficiency gains are real. No client relies on the tool alone. Lawyers are expected to interpret the output, place it in a legal framework and recommend a course.
Audit is the clearest case of all. Deloitte's juniors, who once produced lists and reconciliations, now examine the anomalies, which moves the learning curve without removing it. That holds only where seniors take the time to explain what they are doing. In banking, Solomon's last 5 percent is structure, timing and negotiation, and that is where a deal succeeds or fails.
Learning has to be designed now that it is no longer a by-product
If the doing goes, organizations have to build the learning deliberately. Case shadowing puts juniors alongside seniors on live projects to observe and reflect, with no deliverable of their own. Simulations and stretch assignments overload people on purpose, close enough to reality to force a rethink. Mentoring and feedback require leaders to schedule time for explaining how they arrived at a judgment, and not only for correcting the result.
The tool market grows faster than any of this. Harvey for lawyers, Kira Systems for contract analysis, AlphaSense for market and research data, ChatGPT Enterprise for text and data work across a company. Each one accelerates the doing. None of them builds judgment, and relying on them alone is how a person reaches year five without having formed any.
The question is who does the teaching
AI takes the routine away, which is welcome. The belief that it also takes over the thinking is where the mistake sits. Critical reflection, an understanding of context and the capacity to judge are what decide the next decade in consulting, law, audit and banking alike.
Responsibility for that lands on organizations and the people who lead them. They have to build the rooms in which young talent learns, or watch a generation slide into the Midjourney mentality, taking what a model returns without examining it. The open question in every one of these firms is the same: how do you train the next generation of consultants, lawyers, auditors and bankers to work with AI and still think for themselves?
Why I write about this: I have spent many years in restructuring, turnaround work and the core professions around consulting, and in all of them what mattered was the thinking, more than the processing. The training of young talent is the part I care about most. Those apprenticeship years are what AI is now removing, and unless we build new places to learn, we risk a generation that consumes results without understanding them. That is a competitive question as much as an educational one, because complex crises are only survivable with people who reflect.
Sources
Grouped by the section they support.
Opening
- Confino (2025), Goldman Sachs CEO says that AI can draft 95% of an IPO prospectus in minutes, Fortune, 17 January 2025, fortune.com. Supports Solomon's remarks at the Cisco AI Summit: 95 percent of an S-1 drafted by AI in minutes, a task that took a six-person team two weeks, and "the last 5% now matters because the rest is now a commodity."
The work juniors learned from is the work being automated first
- Financial News (2025), Deloitte triples number of auditors using AI chatbot, 8 April 2025, fnlondon.com. Supports the tripling of PairD use among Deloitte UK auditors.
- Harvey, harvey.ai, and Kira by Litera, litera.com. The legal tools named in the text.
Judgment is what is left, and it is harder to teach
- 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 remarks on injecting cultural traits and leadership principles into AI agents.
Every one of these professions keeps its last five percent
- McKinsey & Company (2023), Meet Lilli, our generative AI tool that's a researcher, a time saver, and an inspiration, mckinsey.com. Supports McKinsey's AI research platform. The assessments of tax, law, audit and banking are the author's own.
Learning has to be designed now that it is no longer a by-product
- AlphaSense, alpha-sense.com, and OpenAI, ChatGPT Enterprise, chatgpt.com. The further tools named in the text. The training formats are the author's recommendation.