Back when I worked in consulting, the job I hated most wasn't the intense project analysis. It was pulling all-nighters to redo PowerPoints.
Not the content. The formatting. I remember working on a pitch for a financial services client once. At three a.m., the partner dropped a line in the WeChat group: "This page's color still isn't right." I fixed it and sent it back. At four a.m. he pinged again: "Change it back to the previous version." That kind of thing could happen five or six times in a single night. The final version carried exactly the same information as the 6 p.m. version. What actually changed was color, alignment, font size, chart style. None of it had anything to do with "business insight."
But no one at the time thought this was odd. Everyone accepted that this was what junior consultants were supposed to do. You had to grind through this phase.
Looking back now, the whole thing was already a signal: if a large share of a job's hours go into execution that has nothing to do with core value, sooner or later that piece gets handed off to something cheaper. Back then, the "something cheaper" was a younger person willing to stay up later. Now it's AI. Internal tools like BCG's Deckster and McKinsey's Lilli, according to public reporting, can already do roughly 80% of a junior analyst's research and deck-building work.
---
The longer you stay in a consulting firm, the more you realize a fact nobody likes to say out loud: on most projects, we don't actually understand the client's business better than their own people do.
So what were clients actually paying several million for? A few things, really. One was the ability to take a vague question and use frameworks to pull it apart — the client's own team is too buried in the day-to-day to do that well. One was a sense of what was happening on the outside: what competitors were doing, what the best practices in the industry looked like. We'd built up this external perspective across many projects. And then there was the final deck — turning all of the above into something the client's CEO could walk into a board meeting with next week.
Before 2023, these things had a real barrier to entry. A fresh graduate couldn't pull them off alone. But AI has basically flattened that barrier. Structured problem decomposition — a large model does it no worse than a consultant with three years of experience. Industry scanning — the volume it covers in a few minutes exceeds what a human team can cover in a week. Deck delivery — tools are already taking it over.
Which turns the question into this: when the client can get outside information, industry benchmarks, and a structured framework out of AI on their own, do they still need to pay several million for an external team to do it?
I can't quite say they don't. Complex organizational change, strategy rollouts entangled with internal politics, situations where decisions need the weight of an "outside authority" to actually move forward — consulting firms still have a place in those. But that value has nothing to do with "junior analysts pulling all-nighters on slide formatting."
In 2023, McKinsey launched its internal AI tool, Lilli. That was also the year a wave of layoffs began that has since cut more than 5,000 people, bringing total headcount from roughly 45,000 down to around 40,000. Over the same period, graduate recruiting at the UK Big Four also dropped off a cliff — KPMG UK's 2024 graduate intake was 29% smaller than 2023 (down from 1,399 to 942).
By 2026, McKinsey announced a 12% expansion in North America, with entry-level headcount possibly growing 15%–20%. But this isn't a traditional "refill the analyst pool" move. Eric Kutcher, chair of McKinsey North America, said in effect: a 20-year-old econ major is better at using AI than a 35-year-old consultant, so younger talent's value has actually gone up. The roles being cut and the roles being added aren't the same role.
---
Later I moved to a top Chinese internet company as a business analyst and operating analyst, thinking I'd switched lanes.
The shape of the pit turned out to be pretty similar. At the consulting firm, most hours went into polishing slides. At the internet company, most hours went into writing SQL to pull data. Then writing the analysis report after the numbers came in.
You'd think pulling data was a technical task, but it's more of a queueing task. The data platform is slow. A moderately complex query sits for half an hour. You come back and realize you wrote the condition wrong — back into the queue. By the time you finally have the numbers, the question your boss wants answered this week has already shifted.
The core skill set of this role is actually thin: slightly more structured thinking than the business people, slightly more data-query ability than the business people. Able to pull numbers out of the system. Able to organize scattered findings into something with a logical shape.
Its core weakness is obvious: the first-hand feel for the business is nowhere near what the product manager or ops lead who lives in that scenario every day has.
Put plainly, this role is a "translation layer in the middle" — translating business questions into data language, then translating data results back into business language. The whole layer exists only because the business side can't talk to data directly.
That premise is loosening. Text-to-SQL can already generate SQL straight from natural language. Models like DeepSeek and Claude can read large datasets, run descriptive statistics, and produce visualizations, collapsing what used to be hours of pulling and charting into a few minutes. If a business lead can talk to data directly in natural language and get answers back, the necessity of the translation layer has to be weighed all over again.
I'm not saying the layer disappears entirely. AI-generated SQL still makes mistakes. Business people can't always judge whether a conclusion is sound. Complex attribution and cross-validation still demand real expertise. But the threshold for "pulling data" has been meaningfully lowered — that's a fact — and pulling data used to eat up a large chunk of a BA's working hours.
---
Now I do marketing at an AI startup.
I used to be the one getting hit by AI. Now I'm on the supply side, spending my days helping enterprise clients work out what AI can actually do.
After talking to a lot of buyers, I've noticed something slightly counterintuitive: after an AI product demo, what worries a company most often isn't "can this thing replace my analyst." It's two other things — "who tells the AI what to analyze" and "can I trust the answer it gives back."
The first is problem definition. I was once demoing at a retail client, and their head of data said something that stuck with me: "Your agent is smart. Whatever I ask, it can answer. The problem is, on this particular day, I don't actually know what I should be asking it." AI executes analysis quickly, but it doesn't know which metric matters most at the current stage, which dip is a real problem and which is normal noise, or what the CEO wants to hear next week. Those judgments have to rest on a deep familiarity with the business and the organization.
The second is verification. AI hallucinates. It will fabricate insights that sound extremely reasonable. BCG and Harvard Business School's well-known "Jagged Frontier" experiment already made this point: on tasks within AI's capability, consultants using GPT-4 completed about 12% more tasks, 25% faster, at roughly 40% higher quality; but on complex tasks outside AI's capability frontier, the group that over-relied on AI got answers right at a rate about 19 percentage points lower than the group that didn't use AI at all.
Put differently: AI has automated the execution side of an analyst's work. The judgment side wasn't replaced — and precisely because execution is now cheap, judgment has become more valuable.
PwC's 2025 Global AI Jobs Barometer reports that roles requiring AI skills pay a 56% wage premium over otherwise comparable roles. A 2025 Harvard study based on the résumés of 62 million workers found that in firms actively adopting generative AI, junior-role employment shrank 7.7% relative to non-adopters within six quarters of adoption — and the drop came mostly from "not hiring," not from layoffs. Those two numbers are really describing the same thing from two angles.
---
My current view of this industry:
What AI is doing to business analysts is redrawing the line between "work that has value" and "work that doesn't."
The slide-format fixes we pulled all-nighters for in consulting, the hours spent writing SQL at the internet company — the value of those tasks was never high to begin with. There just wasn't a cheaper substitute, so by default a human did them. AI has simply pulled that fact out into the open. It can't be unseen anymore.
What's actually valuable is the stuff AI can't do — reading the business, asking the right question, judging whether a data conclusion is trustworthy, pushing something through a complex organization. None of that has depreciated. If anything, it's scarcer than before.
But here's the implication that makes me uncomfortable: if the core of value lives in business judgment and problem definition, then the hard skills a BA used to be proud of — SQL, statistical modeling, Excel — were never really a moat. They were an entry ticket under a specific set of technical conditions, and AI is driving the price of that ticket down toward free.
So the capability model for this whole profession needs to be rewritten. What's depreciating, what's appreciating, and the logic behind what appreciates — I'll get into that in the next post. Assuming, of course, I can work it out for myself first.