What replaces the analyst-to-VP apprenticeship if the grunt work that taught judgment disappears?

The shift

Producing a competent first-pass model, comp set, or memo — the volume-of-reps work that used to be the only way analysts built pattern recognition — goes from scarce (years of analyst hours) to abundant (minutes, AI-generated, at near-zero marginal cost).

The axioms

• Judgment is tacit and gets acquired by grinding through thousands of low-stakes reps (comps, models, deck formatting) under a senior person who checks the work.
• The grunt work itself is a training exercise disguised as a deliverable — priced and staffed as if the output were the point, when its real function was building the analyst.
• Trust in a junior is earned incrementally: having built 200 models is a proxy for being reliable enough to hand harder calls to.
• Pattern recognition — this deal smells like that deal — comes from grinding enough live reps that intuition forms; it takes years because reps were rate-limited by headcount and deal flow.
• The pyramid (many analysts, few VPs) is an economic structure as much as a teaching one — cheap junior labor subsidizes senior time and gives seniors leverage and deniability.
• A VP's authority to override a model or call a client rests on having personally sweated the same kind of detail once, which is what lets them smell when something's off.

Invalid axioms

1. Building a model or comp set from scratch is how an analyst earns their seat. The scarcity was analyst-hours needed to produce a usable first draft; AI collapses that to minutes. The habit-trap: banks still staff and bill deal teams by analyst headcount and hours, as if drafting capacity were the bottleneck, and still promote based on who ground through the most models rather than who caught the most errors in one.
2. Reps accumulate through personally doing the grunt work. The volume of hand-built models used to be the only available proxy for "has seen enough patterns." Now a junior can generate that volume without internalizing anything — the habit-trap is mistaking output volume for exposure, and continuing to credit "did 300 models" as if it still meant "has pattern-matched 300 deals."
3. The pyramid's headcount ratio is set by how much grunt work exists to distribute. Deal teams are still sized as if someone needs to spend 60 hours building the base case. That labor is now abundant; the sizing logic that justified the analyst class hasn't been re-derived.

Unchanged axioms

1. Someone accountable has to own the number that goes in front of a client or IC. A model can generate a plausible DCF instantly; it can't be the person who answers for it being wrong. Accountability doesn't get cheaper because drafting did.
2. Smelling when a model is wrong under novel, ambiguous deal terms is still a human skill, and it's not obviously learnable by watching AI do the drafting. The open question is whether that judgment can be built any other way than by once having built the thing yourself — nobody has run this experiment at scale yet, and the honest answer is "we don't know," not "no."
3. Client trust and the standing to make a call in a live negotiation stay relational and human. A VP's authority in the room comes from a track record and a face people have dealt with before, not from having produced more tokens.

New axioms

1. If judgment used to form through the physical act of building, what's the substitute training loop when building is instant? Reviewing AI output at speed is a different cognitive task from constructing it from scratch — nobody has shown review-only reps produce the same intuition, and firms have no answer yet for what juniors should spend years doing instead.
2. When first drafts are abundant and instant, verifying them at the same volume becomes the bottleneck — and juniors aren't automatically good at this just because they can prompt a model. Catching a confidently wrong AI output requires the same pattern recognition the grunt work used to build, creating a chicken-and-egg problem: you need judgment to verify AI, but judgment used to come from the work AI now does.
3. Compressed timelines mean whatever apprenticeship survives has less runway before someone is trusted with real stakes. If juniors reach "VP-adjacent" judgment demands faster because AI absorbs years 1-3 of grunt work, the industry has no calibrated sense of how much faster is safe versus reckless.

Where it breaks

Banks keep the analyst class sized and priced as if drafting capacity is still the scarce resource (INVALID #1 and #3) while having no answer for how those same analysts are supposed to develop the verification judgment the new world actually needs (NEW #2). The two-year analyst program still exists to produce cheap labor for a task AI now does better and faster — but nobody has redesigned what those two years are supposed to teach instead, so the pyramid is being kept intact for the wrong reason at the exact moment its original reason disappeared.

ai-axioms.vercel.app/axioms/finance/finance-ib-apprenticeship-ai-axioms