AI will not just replace junior consultants. It will expose what consulting was really selling.
AI is unbundling what consulting really sold. The pyramid is becoming a barbell — generic analysis squeezed, judgment-heavy work repriced upward.
The standard story about AI and consulting is that the analyst layer goes first, the partner layer hangs on, and everyone in between holds out hope of being one of the survivors. That story is directionally correct and analytically lazy. It treats the consulting pyramid as a hierarchy of seniority when it is actually a bundle of services. What AI is doing is unbundling that bundle.
The old consulting bargain
For most of the modern era, consulting firms sold four things stapled together: scarce analytical labour, a structured way of converting ambiguity into a plan, institutional memory from having seen many similar problems, and senior credibility when the recommendation needed to survive a board meeting. The pyramid worked because the first item — analytical labour — was both expensive to produce and genuinely valuable to clients. Analysts gathered data, cleaned it, modelled it, benchmarked it, and turned scattered information into a coherent story. Managers framed and reviewed it. Partners sold it and took the heat.
Clients paid a premium because all four pieces were difficult to assemble in-house. The consultant’s edge was real. But it was always a bundle, and a bundle is only as defensible as its most defensible component.
AI attacks the analytical layer first
The reason “junior first” is directionally right has nothing to do with juniors being less intelligent. It has to do with the shape of their work. Most entry-level consulting tasks — first-cut market scans, interview synthesis, competitor benchmarking, spreadsheet clean-up, draft slide writing, meeting summaries, basic issue trees — share three properties: they are legible (you can specify what good looks like), repeatable (the same patterns recur), and output-shaped (the deliverable is a discrete document or chart). Those are exactly the properties that make work AI-friendly.
The available labour-market data points the same direction. Anthropic’s recent work on AI exposure shows the steepest pressure on programmers, customer service workers, and data entry. These are adjacent occupations whose tasks share that same legibility profile. Consulting analyst work is structurally close enough that the parallel is unavoidable. McKinsey’s CEO Bob Sternfels has talked publicly about a deliberate internal restructure cutting non-client-facing roles by roughly a quarter while growing client-facing ones by the same. BCG’s Deckster, McKinsey’s Lilli, KPMG’s Workbench and Deloitte’s Zora are all production tools doing what a strong second-year analyst used to do, faster and at scale.
But “junior first” is too crude as a stopping point. It misses that AI is not removing the people; it is removing the price on a category of output. That distinction matters because it implies a much wider blast radius than the org chart suggests.
The conversion layer is more exposed than the headlines admit
The most awkwardly positioned tier in consulting is not the entry-level analyst. It is the senior analyst, manager, and engagement manager whose value to the firm has been reliable conversion of client ambiguity into standard outputs. Call it the conversion layer. Its economic function is taking unstructured intent — “we want to be more AI-native” — and producing structured artefacts: a 90-day roadmap, a target operating model, a benefits case. That role exists because the partner cannot personally translate eight messy interviews into a clean problem statement and the analyst is too inexperienced to do it well. The middle layer ran the conversion.
AI compresses the conversion itself. Taking unstructured intent and producing structured first drafts is exactly what large language models are now best at; Lilli, Deckster, and the BCG- and McKinsey-built variants of GPT for internal use are conversion engines. The hypothesis that follows is that a partner with a good model, a good prompting habit, and a junior who can stress-test outputs will eventually need fewer senior analysts in the middle to translate.
That is hypothesis, not yet evidence. McKinsey, BCG, and Bain have not visibly flattened their middle layers. McKinsey’s Lilli is now used monthly by 75% of its 43,000 employees, an internal datapoint that cuts both ways. It shows the conversion engine exists; but McKinsey is simultaneously raising graduate hiring by 12% in 2026 (per Eric Kutcher, North America Chair). The precondition is forming, not the result. A reasonable test of this prediction: by 2027, look for Big Three engagement-manager-level promotion rates to compress, or for at least one MBB firm to publicly de-emphasise the EM grade. If neither happens by 2028, the thesis is wrong.
What is genuinely new, and what makes the standard “junior first, senior safe” framing unhelpful, is the directional pressure. Mid-tier consultants whose moat was speed, reliability, and ownership of the standard deliverable are the ones whose pricing power is most directly threatened by a tool that does conversion at scale.
In AI and data consulting work specifically, this pattern shows up in a slightly different form. Clients no longer pay a premium for a manager-level person to wrangle a team of analysts into a polished deck. They pay for someone who can decide what to build, sit with the discomfort of an ambiguous problem, and own the call when it is wrong. Output ownership is becoming cheap; problem ownership is becoming the thing.
The corollary is uncomfortable for senior people, and it is the part of this story most consulting commentary avoids. Tenure does not automatically protect anyone. A partner whose value rests on pattern recognition from twelve years ago, applied to industries that have since shifted under their feet, is exposed in a specific way: AI makes stale expertise visible. Howard Marks, in his February 26, 2026 Oaktree memo AI Hurtles Ahead, made the point with unusual bluntness for an investing context: “readily available, quantitative information about the present can’t hold the key to superior investment performance for the simple reason that everyone has it. Now, to the fact that everyone has it, we have to add the fact that AI can probably do a better job than everyone of [analysing it].” The same logic transfers cleanly. When a model can produce a defensible synthesis of the last eighteen months of an industry in twenty minutes, a senior consultant whose mental model is older than that is no longer the room’s most current source. Reputation is a lagging asset; it can persist for years past the point where the underlying judgment has decayed. The barbell that follows works only for the population whose judgment is live, which is meaningfully smaller than “everyone with a partner title.” This is a narrower category than most partnerships are publicly willing to admit.
The new barbell
Marks made the canonical version of this argument for investing in the same memo. His broader point: AI will dominate analysis of widely available information, which makes it harder for traditional active investors to win on data alone. What survives, and gets repriced upward, is judgment in genuinely novel situations, qualitative discernment, and the willingness to bear concentrated risk. AI, Marks notes, has no skin in the game.
The transfer to consulting is direct. AI compresses the value of competent-but-generic analysis. It increases the value of people who can decide what matters, make tradeoffs, carry delivery risk, and get an organisation to act. The result is a barbell. AI-native boutiques and judgment-heavy advisors gain leverage; the middle of the market, where firms compete on standard process and standard output, gets squeezed.
Concretely, a top consultant in a post-AI world is not valuable because they can produce a market map. AI can do that. They are valuable because they can say this market map is not the decision you need to make, or the board will not buy this unless we remove two priorities, or this operating model is rational but the country heads will block it, or, most consequentially, this is the recommendation I am willing to stand behind.
That last sentence is doing a lot of work. AI produces logically airtight recommendations and bears no consequences for them. A partner who will publicly vouch for a call when the CEO is being challenged at the next board meeting is offering something AI structurally cannot. Sternfels has talked about roughly a third of McKinsey’s revenue now coming from outcome-underwriting: fees tied to whether the recommendation works. That is the institutional version of the same shift: firms are being pushed to put their own money where their advice goes.
The apprenticeship problem
The barbell creates a quiet structural problem most firms have not solved. The traditional path to becoming a senior consultant ran through grunt work: do a thousand market scans, sit through five hundred client interviews, build two hundred models, see what good looks like, slowly accumulate judgment by repetition. If AI removes the grunt work, the apprenticeship breaks. A junior who never built the model and never wrestled with the messy interview transcripts has no obvious route to becoming the senior who can sense when a number looks wrong.
Harvey’s Winston Weinberg has been candid that this worries him more than almost any other question facing his company in legal services. He told StrictlyVC in November 2025 that “the most important thing financially for a law firm is to make sure you’re hiring, training, and developing lawyers that get to being a partner as fast as humanly possible.” The consulting equivalent is unspoken but identical. Firms bullish on hiring more graduates are betting that the apprenticeship can be reconstructed around managing AI rather than producing the work. McKinsey’s North America Chair, Eric Kutcher, has said the firm will expand North American entry-level hiring by 12% in 2026. That bet is not obviously wrong, but it is unproven, and “let the junior manage agents” is not a real training programme if the junior cannot tell when the agents are wrong.
There is no graceful answer to this yet. The honest version is that firms are running an experiment on their own future senior bench, and the results will not be visible for five to seven years.
The question for anyone still in this work
The career framing this all collapses to is not “what hyphenate role do I become.” Strategist-slash-product-manager-slash-workflow-designer is a true description and a useless prescription. The sharper question is: do you own a slice of the problem, or do you produce artefacts about it?
That is the line. AI gets very, very good at artefacts. It does not yet, and on current trajectory will not soon, own problems. Owning a problem means deciding what is worth working on, choosing what to cut, carrying the consequence when it is wrong, and being the person an organisation actually trusts to push the recommendation through its own internal politics.
The honest answer for the AI and data consultants and the strategy generalists who are paying close attention to this shift is that the role is being rewritten in real time, and most of the rewrite is on the human side. The MIT NANDA finding that 95% of enterprise AI initiatives produce no measurable P&L impact is partly about model limits, but it is mostly about clients not knowing how to convert capability into operational reality. That gap is large, and it is not closing on its own. Whoever sits inside it productively, with both technical fluency and the willingness to own outcomes, is doing what consulting was always supposed to be about, before the bundle stopped making sense.
The pyramid is becoming a barbell. The work that survives is the work that was never really about the slides.
Prediction tracker
Claim: By the end of 2027, at least one Big Three firm publicly de-emphasises the engagement-manager grade or visibly compresses EM-level promotion rates.
Confidence: Medium. The tools (Lilli, Deckster, GENE) and the commercial logic are in place; the partnership politics are the obstacle.
Falsifier: If, by the end of 2028, MBB engagement-manager headcount as a share of total firm headcount has grown versus the 2025 baseline, this article was wrong about the conversion layer.