AI can make client delivery more efficient while changing the mechanism through which firms develop future senior judgment
Senior judgment is what a professional firm sells, and it is also what the firm has least of. That makes professional services unusual: the product and the constraint are largely the same thing. Yet most visible AI investment is aimed somewhere else, at research, drafting, document review and compliance processing.
That does not make those investments wrong. If demand is weak, or if a firm genuinely lacks delivery capacity, cheaper and faster production may be exactly what it needs. But in a relationship-led firm with work to sell, the harder constraint is the finite number of people who can win a client's confidence, stand behind the work and make a call that nobody else can check. A firm can stretch that supply, or poach it from a competitor, but it cannot make more of it quickly.
The scale of investment below that constraint is already large. Deloitte has announced a $3 billion commitment through 2030, and EY has announced 150 agents for 80,000 tax professionals. The figures are not directly comparable, but they show where firms expect AI to make a difference: making the work around senior judgment faster.
If those efficiencies free senior people to spend more time with clients, the investment can reach the constraint. Faster research, drafting or compliance work does not automatically create more senior judgment, though. The firm may simply arrive at the bottleneck sooner.
The harder problem is that the same client work being made more efficient is also one of the ways firms have produced the judgment they are short of.
Judgment is not produced by exposure
Professional firms often explain the development of judgment through exposure. Put a bright graduate close to difficult work, let them watch enough senior people, give them enough years, and judgment accumulates.
Exposure helps, but it does not explain how the conversion happens. A junior can watch a thousand difficult decisions and become very good at recognising what experienced people do without becoming capable of making the same decision independently. A lot of the conversion, in my experience, happens through consequential error under supervision.
The error has to be the learner's own. They form a position and discover it is wrong: they decide what the number means, whether the clause works, which risk matters or what the client is likely to accept. If they never held the mistaken view themselves, there is nothing in their own reasoning to correct.
It also has to matter enough to register. It does not need to damage a client or cost the firm money; defending a recommendation in front of a team whose opinion matters is usually enough. A wrong answer with no felt consequence gets filed as information rather than experience.
And someone with better judgment has to explain why the call was wrong: which assumption failed, which signal was missed, what should travel to the next situation. Without that, people tend to learn to avoid the whole category of decision instead.
People do not learn only by making mistakes. Observation, simulation, correct calls and near misses all contribute. The narrower point is that professional firms have relied heavily on supervised, consequential error to turn knowledge into independent judgment without ever needing to name the mechanism, because paid client work created the conditions on its own.
AI can interfere with that mechanism without making the apprenticeship look very different.
The junior is still on the engagement. The client stakes are still real. The partner still reviews the work. The hours, titles and progression can look unchanged. But if an AI-first workflow supplies a competent analysis before the junior has formed one, there may be no genuine error of the junior's own for consequence and explanation to act on.
Judging an answer is a different activity from arriving at one. A reviewer asks whether an existing position is defensible. An originator has to form a position that might be wrong. If the first move in the workflow becomes reviewing a competent model output, a person may get very good at evaluation without experiencing enough of the mistakes that once helped make evaluation possible.
PwC has said something close to this out loud. A PwC assurance leader described new hires as "almost instantaneously becoming reviewers and supervisors", while also saying the firm would teach critical thinking and professional scepticism earlier. That leaves the question of what experience made the new hire capable of reviewing someone else's answer in the first place.
A junior who forms a position before seeing the model's answer keeps the first condition and gains a useful comparison. Someone who independently rebuilds an analysis, finds the model's error and defends a different view may learn a great deal. AI may also release senior time for better coaching. The question is how much of the learner's own first attempts a competent model output ends up replacing.
People who routinely begin with model output should eventually perform just as well on unaided consequential decisions as people who routinely formed their own position first. If they do, the mechanism is wrong or overstated. No published longitudinal comparison of that kind exists yet, and any failure would take years to become visible. Faster work appears immediately; weaker judgment may not appear until that cohort is expected to make the calls itself.
So firms are pointing AI at the work around the constraint, and in doing so may be weakening one of the ways they produce more of it.
The supply side is producing the wrong thing
Most firms of any size run a matrix. On one side is demand: sectors, markets, client relationships and problems to solve. On the other is supply: practices, capabilities and the people the firm can field.
The supply side has traditionally been organised to manufacture competence. It defines capabilities, builds methodologies, trains people, assesses proficiency and moves them through levels.
AI changes the economics of much of that. It is making codified knowledge and competent first-pass analysis cheap, and doing nothing of the kind for judgment.
A natural response would be to add judgment to the capability model: define levels of judgment, write behavioural indicators, build a curriculum and assess people against it. Judgment does not fit that model well. A framework can describe evidence of good judgment after the event, but it cannot manufacture the circumstances in which judgment develops.
If those experiences no longer arise reliably as a by-product of paid work, the supply side has to concern itself not only with what people know but with the situations in which they are placed.
A casting function is one way to think about that change. The question becomes which person should own which uncomfortable decision next, in front of whom, with what consequence and at what level of risk. The placement has to matter enough to register but remain survivable when the person gets it wrong.
Professional firms once did this more intuitively. People were placed slightly beyond their current competence under credible supervision. It did not need to be labelled judgment development because the structure of client work generated enough of these situations on its own.
As AI removes more routine first attempts, firms can no longer assume the next useful stretch will simply appear. Someone has to decide where it will come from, and making that decision well consumes senior judgment too.
Demand has a client in front of it today. The risk and the margin on the work are both measured today, so the safest staffing choice is the person most likely to get the answer right with the least supervision.
Supply sometimes needs the less experienced person to make the call, under supervision, because that person is not yet the safest choice. The firm needs them to become capable of making the same call several years from now.
Historically, client delivery could take priority without completely starving development. Even well-staffed engagements contained unavoidable first attempts. People still had to work things out for themselves, so judgment continued to develop as a side effect of serving the client.
AI-assisted work can reduce those moments. The demand-side risk remains immediate and visible in the engagement P&L, while the developmental loss is spread across a cohort and may not appear for years. Every local incentive points the same way: protect the client, protect the margin, use the best available answer.
The matrix has always chosen the client in front of it, and that was fine while judgment kept developing as a by-product of the work. The change is that it may no longer, and nobody in the structure currently owns the question.