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: KI in Beratung: Konkrete Skills im praktischen Kontext.
Ezekiel Akinsanya spent four years at Princeton aiming at one employer. He co-founded a consulting club, did the summer internships, drilled the cases. When he graduated in spring 2026, he turned consulting down and planned to return to England for work in government or the military. Bloomberg told his story under a blunt headline: consulting was a dream first job until AI threatened those roles.
Anyone aiming at consulting knows the doubt behind that choice, and the signals have been accumulating. McKinsey's headcount fell by more than ten percent in eighteen months to about 40,000, as reported in May 2025. The large firms are not hiring fewer MBAs so much as different ones, with specialist backgrounds favored over the generalist profile. Some observers expect the classic consulting pyramid to become a diamond, thinner at the base and wider in the middle.
The conclusion runs differently: one particular kind of junior work is being automated. Aggregating data in Excel, building a standard market analysis, or writing an executive summary in bullet points. McKinsey's internal AI platform, Lilli, already takes on work of this kind: it drafts proposals and builds slides from simple prompts.
What survives shows up in the firm's own interview format. McKinsey has begun piloting an interview round in which candidates work with Lilli. The point is not whether they can avoid AI, it is how they work with it. Candidates are expected to prompt the tool, review its output and apply judgment, and the firm says it is testing skills such as collaboration and reasoning. What gets tested is judgment in dialogue with a machine, more than prompt craft.
So the question for anyone entering the field has changed. Being replaced is no longer the interesting part of it. Which work is still needed, and how you prepare for that, is.
The clearest answer comes, of all places, from Palantir projects.
What an AI consulting project actually consists of
Bain and Palantir run a partnership that produces use cases at large corporations. Boston Consulting Group's own research finds that around three quarters of companies struggle to achieve and scale value from AI. The projects that succeed are the interesting ones, because they make visible what human work sits between "we have AI" and "we generate measurable value with it".
Take three such projects, at a large meat processor, a packaged-food group and a consumer brands company. Look at them side by side and ask what the consultant actually did, and the same pattern shows up in nearly every project. Not "deployed AI". Four very specific thinking movements.
Those four are structurally identical to what you do when you build a master prompt, meaning a structured, multi-stage prompt. In a consulting project they turn into an implementation with a budget in the millions, while inside your Claude or ChatGPT window they become three to five prompt stages, divided into sections by markup tags. The thinking is the same. Only the scale differs.
That is why prompt engineering, taken seriously, is compressed consulting thought at the smallest available scale.
First: recognize the fragmentation, before the machine calculates anything
Every one of these cases starts with the same diagnosis. Data sits scattered across enterprise resource planning systems, laboratories, logistics tools and local spreadsheets that nobody maintains centrally. Before any model does anything, somebody has to understand where the company's knowledge is stuck and how it connects, or fails to connect.
At the meat processor the consultants first had to bring several separate planning systems into a common model before the AI layer could calculate at all. The value sits in the diagnosis rather than anywhere near the model: somebody has to see that the fragmentation is the actual problem, and pull the pieces of information together.
You do the same thing in miniature when you build an agent. Say you want a research assistant to prepare literature for your thesis. Before prompting anything, the questions are which sources count in your field, what your supervisor expects stylistically, which arguments you already have, and which counter-positions exist. That is a context inventory, and it separates an agent that produces generic summaries from one that does something for your specific project.
Most students skip the step. They write "help me with my thesis" and are surprised by a generic answer. The consultant on that project would have discarded that chat in thirty seconds, because the diagnosis comes before the solution.
Second: build an ontology, which is the quiet core of the work
This is the central and usually invisible part of every case. An ontology, in plain terms, is an explicit model of which entities exist in a system, what properties they have and how they relate. What is a supplier? How does it relate to a product? When does an order become a commitment?
For its trade promotion work, the consumer brands company built an ontology out of syndicated market data, existing tools and unstructured documents such as retailer rules and brand strategies. Only once that conceptual work was finished could elasticity models calculate anything meaningful, and only then could the tool help steer decisions about advertising spend. None of that is technology. It is the precise fixing of what this company means when somebody says promotion.
A well-structured master prompt is the smallest version of the same work. Working with tags such as <role>, <context>, <rules> and <output_format> is defining entities. Making explicit inside those tags what a qualified lead is in your context, what makes a section complete, at what point research counts as sufficient, is exactly the ontology work done on that project.
A concrete example. Filtering candidates for a position with AI requires criteria that say what qualifies:
<qualified_contact>
- at least 50 employees
- actively hiring in engineering
- budget holder reachable via LinkedIn
</qualified_contact>
Three lines of conceptual work, and they are the three lines that stop your agent treating every contact identically, and the point at which the machine stops guessing.
Third: classify the decisions you are willing to delegate
A company of that size makes an enormous number of operational decisions every year. The critical consulting work is the classification: which decisions are delegable, and which need human judgment.
Some warehouse reallocations can be decided algorithmically: the criteria are clear, the consequences reversible, the risk contained. A price adjustment across an entire product line is a different order of magnitude, where a person has to check the recommendation against political and strategic context and then decide. The useful systems distinguish these levels themselves and treat them differently, acting autonomously on some and only proposing on others.
Building an agent means making the same classification explicit in the prompt. Instructions along the lines of: for tasks of type A act independently; for type B show me your plan before you implement anything; for type C always ask first. Without that boundary an agent is either too passive or too intrusive, useless in the first case and dangerous in the second. With it, you have a tool holding controlled autonomy.
That is the same skill a consultant applies when scoping an AI project. Smaller scale, identical logic.
Fourth: design the handoff, because adoption is the real metric
The number worth watching in these projects is the acceptance rate: the share of AI-generated recommendations that human decision-makers actually adopt. That is a design result and not an AI result.
The consultant builds the handoff so that a person can follow the recommendation, interrogate it and reject it. Its shape is not "the model says X, so we do X". It is "the model says X, shows us the two assumptions carrying it, and we decide in three seconds whether we agree". That design separates a tool people use from one they ignore.
In a master prompt, that is the instruction dividing a useful agent from a dangerous one. Before you make a recommendation, name two assumptions I should check. Make your reasoning explicit, not only your result. Say plainly what you do not know and what you guessed. Lines like these look unspectacular after three prompts, and they are the difference between an agent you follow blindly and one you can actually think with.
They are also, unsurprisingly, what McKinsey tests in its Lilli interviews. Not the ability to build a clever prompt. The ability to receive an output critically and install the handoff mechanics so that the person ends up making a better decision than either party would alone.
The short form
What these cases make visible is a plain thesis. AI consulting means structuring problems so that AI can solve them, and knowing precisely enough when the machine cannot be trusted that you notice at the right moment.
Those four movements, recognizing fragmentation, building an ontology, classifying decisions and designing the handoff, are not prompt tricks. They are the structural form of what a good consultant does in 2026. Master them and you sit on the growing side of the diamond, and skip them and you sit in the shrinking part.
The good news is that they are learnable. Because they show up the same way at small scale as at large, they can be practiced on a laptop, in a chat window, in a single evening. A master prompt that recognizes fragmentation, carries a clean ontology, sets decision tiers and designs handoffs is the same thing as a corporate implementation, at a scale of one to ten thousand.
Anyone who has understood the difference between using AI and working with it in a structured way has anticipated the most important shift in consulting for the next several years. The best part: none of it requires a consulting job. It requires a good master prompt, and the willingness to rebuild it until it holds.
Sources
Grouped by the section they support.
Opening
- Boyle (2026), Consulting Was a Dream First Job Until AI Threatened Those Roles, Bloomberg, 15 April 2026, news.bloomberglaw.com. Supports Ezekiel Akinsanya's path at Princeton, the consulting club, the internships and his decision after graduating in spring 2026.
- Merritt (2025), McKinsey's headcount drops more than 10% in 18 months, Fortune with Morning Brew, 29 May 2025, fortune.com, following a report by Business Insider. Supports the fall to about 40,000 employees.
- Lem (2026), The MBB Aren't Hiring Fewer MBAs – They're Hiring Different Ones, Poets&Quants, 1 July 2026, poetsandquants.com. Supports the shift from generalist to specialist profiles.
- Pye (2026), AI may up-end the consulting pyramid, Consultancy.uk, 19 February 2026, consultancy.uk. Supports the diamond shape, thinner at the base and wider in the middle, as one possible outcome.
- El Chmouri (2025), McKinsey leans on AI to make PowerPoints faster, draft proposals, Fortune with Bloomberg, 2 June 2025, fortune.com. Supports what Lilli does.
- Angelo (2026), McKinsey challenges graduates to master AI tools as it shifts hiring hunt toward liberal arts majors, Fortune, 14 January 2026, fortune.com. Supports the pilot interview with Lilli and what candidates are expected to do in it.
What an AI consulting project actually consists of
- Bain & Company (2026), Bain & Company announces expansion of lead global management consulting partnership with Palantir, press release, bain.com. Supports the Bain and Palantir partnership.
- Boston Consulting Group (2024), AI Adoption in 2024: 74% of Companies Struggle to Achieve and Scale Value, press release, 24 October 2024, prnewswire.com, on the report Where's the Value in AI?, bcg.com. Supports the roughly three quarters of companies.
- The three projects are described from the author's own knowledge; the companies are not named. The link between consulting work and master prompts is the author's own argument.
First: recognize the fragmentation, before the machine calculates anything
- No external source. The project example comes from the author's own knowledge; the transfer to agents is the author's own.
Second: build an ontology, which is the quiet core of the work
- No external source. The project example comes from the author's own knowledge; the prompt example is the author's own.
Third: classify the decisions you are willing to delegate
- No external source. The decision tiers are the author's own analysis.
Fourth: design the handoff, because adoption is the real metric
- Angelo (2026), Fortune. Supports the point that McKinsey's Lilli interview looks at how candidates review output and apply judgment.
- The acceptance rate as a design result and the handoff instructions are the author's own analysis.
The short form
- Pye (2026), Consultancy.uk. Supports the diamond image. The thesis is the author's own.