Skip to content

Thoughts

Snake Oil in a Tailored Suit: Why Classic Consulting Is a Dying Model

Snake oil used to come in a brown bottle from a traveling salesman. Today it comes as a slide deck, with a day rate and a partner title. On an industry that claims expertise in everything, has rarely built anything itself, likes to leave responsibility for the result with the client and has lost track of itself. And on what works instead.

Jonas Höttler25 min read

What matters most

  1. Experts in everything, the risk left with the client. First communication, then digitalization, now AI. Always the same people, always instant experts, which is unlikely enough on its own. Add titles that claim competence where there is none. The people giving the advice rarely built any of it themselves. And they rarely answered for the result.
  2. The industry has lost track of itself. One firm warns of vendor dependency in an AI whitepaper. A week later, a partner at the same firm calls that “simply a fantasy.” The market is stagnating, and the bosses talk about a golden age.
  3. AI hits the business model twice. It replaces the junior work, and it is likely to thin out the middle management through which many consulting projects came in.
Contents14 sections

Snake oil used to come in a brown bottle. A traveling salesman rolled into town, climbed onto a crate and sold a miracle cure for everything. By the time anyone noticed it cured nothing, the wagon was already in the next town.

Today snake oil comes as a slide deck. The wagon is a company Audi, the crate is a steering committee, and instead of bottles there are day rates. The cure still works against everything: communication, digitalization, transformation, AI. And when it doesn't work, the consultant is long gone, already on the next project.

It was never the consultant's fault. It was the implementation, the culture, the market, the client. The vagueness is no accident. It is the business model: always leave yourself a back door, and let others take the damage.

You can politely call that expectation management. Or you can say what it is for many clients: they are not being advised. They are being played. And they pay the day rate for it.

What It Sounds Like

On October 12, 2026, I asked a simple question on LinkedIn: what does AI cost today, and what will it cost tomorrow? The post was pointed, and disagreement was welcome. It came the next day, from a partner at a management consultancy, responsible for Tech Strategy. First with a riddle, then with the core claim, and finally with an offer. In the original German, with all its errors, and my translation in parentheses:

Partner at a management consultancy · LinkedIn, Oct 13, 2026Red: claim without evidence
approx. 07:50

„Ich verspreche Ihnen, wenn Sie das zuende [sic] denken, ist das erhellend. 🙏🏻“
(“I promise you, if you think this through to the end, it is illuminating. 🙏🏻”)

A riddle instead of an argument.
approx. 13:50

„Der Wechsel von einem LLM aufs andere ist völlig banal und nahezu Kostenfrei [sic]; Jedes [sic] moderne KI system [sic] ist in der lage [sic], eine beliebige mit einer anderen KI gebaute Anwendung 1:1 zu betreiben. Die Lock-In Angst [sic] […] ist schlicht eine Fantasie. 🙏🏻“
(“Switching from one LLM to another is completely trivial and almost free of cost; any modern AI system is able to run any application built with another AI 1:1. The lock-in fear […] is simply a fantasy. 🙏🏻”)

No number. No example.
later

„Wenn Sie Beraten [sic] werden möchten, können Sie gerne ein Projekt beauftragen.“
(“If you would like to be advised, you are welcome to commission a project.”)

An offer instead of an answer.
Evidence missing

Asked for evidence, the partner offered no number. Instead: posts age quickly. In a large firm, not everyone sees everything the same way. My arguments were “bad and thoroughly refuted.” Refuted with what was never said.

The thread is public. Anyone can read it. I'm not interested in the person. I'm interested in what this exchange reveals about an industry. Because in miniature, it shows everything that goes wrong at scale.

Seven Days Earlier

On October 6, 2026, seven days before that comment, the partner's own firm had published a whitepaper on AI. Its message: companies risk “vendor dependency.” A capability that looks cheap at adoption can become expensive once it sits inside an important workflow. Companies should calculate the “total lifecycle cost,” “including data preparation, integration, usage, controls, monitoring, change and maintenance.” A colleague from the same firm had written in July that vendor dependency makes switching “increasingly difficult,” and recommended building up two vendors.

One firm, three statements, 88 days
DateWhoWhat
Jul 17, 2026Colleague, Principal ConsultantVendor dependency makes switching “increasingly difficult.” Recommendation: two vendors.
Oct 6, 2026The firm's whitepaperRisk of “vendor dependency.” Calculate the “total lifecycle cost,” from integration to maintenance.
Oct 13, 2026Partner, Tech Strategy„völlig banal und nahezu Kostenfrei [sic]“ (“completely trivial and almost free of cost”). Lock-in is „schlicht eine Fantasie“ (“simply a fantasy”).

The whitepaper says: plan the exit, calculate the follow-on costs, or it gets expensive. The partner says: trivial, free, fantasy. One logo, one week, two pieces of advice that clash.

You could object: the whitepaper means vendor dependency in a broader sense, covering software, processes and contract terms, not just swapping one language model for another. True. But the partner didn't say a model switch is often manageable. The claim was that any application runs 1:1 on any other system, and the fear of lock-in is “simply a fantasy.” That absoluteness is the problem. It sweeps aside exactly the costs the partner's own firm tells clients to calculate.

Which one does the client get? The one in the whitepaper, or the one sitting in the steering committee? If that depends on which partner happens to be in the room, the client isn't buying analysis. They're buying someone's form on the day. And a mood. Yesterday this, today that, tomorrow something else again. Only the tone stays the same: I'm right, you're stupid. Some people are impressed by that. But inflated incompetence is still incompetence.

And it gets better. When a partner contradicts their own firm, those are two legitimate views. In a large firm, not everyone sees everything the same way. When someone from outside disagrees, it's “bad and thoroughly refuted.” Inside, disagreement is diversity. Outside, it's a thinking error.

And to the outside world, they act as if they were proclaiming laws of nature. Their own view is explained to people as if no second one existed. Outside, 1 + 1 = 2, and whoever doubts it hasn't thought it through. Inside, 1 + 1 may also be 3. That just happens in large organizations.

If the answer depends on which partner is in the room, the client isn't buying analysis. They're buying a mood.

Only the Key Is Trivial

To be fair, part of it is true. Swapping the API key really is easy. I said so in the thread. One line of configuration, and the application talks to a different model.

What hangs behind the key
Swap the API keyone line of configuration
Waterline
  • Retune promptsWhat was built for one model often runs worse on another.
  • Test againEvery workflow, every answer has to be checked again.
  • Monitor qualityResults shift without anyone noticing.
  • Interfaces and formatsTools, outputs and limits differ from vendor to vendor.
  • Your own models and dataFine-tuned models and vector data often can't come along.
  • Contracts and data protectionNew vendor, new review, new approval.
  • Contingency planWhat runs if a model is deprecated tomorrow?
  • CostsNew prices, new bill, new budget.

What hangs behind it is not trivial. Prompts tuned to one model. Tests that have to run again. Results that shift without anyone noticing. A study at the ICLR 2024 conference measured accuracy differences of up to 76 percentage points from the formatting of a prompt alone, and which format works well transfers only weakly between models. A paper from December 2025 calls this “model drifting” and says it is “both common and severe.” In a Zapier survey of 542 executives at large US companies, 89 percent believed they could switch AI vendors within a month. Of those who had tried, 58 percent said the switch had failed or was far more work than expected.

Belief and experience
believe they could switch AI vendors within a month89%
of those who tried: failed or far more work than expected58%

Zapier survey of 542 executives at large US companies, April 2026.

I've been through it myself. A provider deprecated a model, and one day the news search in my project came up empty for 178 businesses. Nothing big depended on it, and it was fixable. It was not trivial. Now picture the same thing at a corporation worth billions on the stock market: a customer service that suddenly answers with nonsense. A check that waves through wrong values. A report full of numbers that don't exist. Anyone building there without safeguards, because switching is “trivial,” is gambling with their client's reputation. And on the stock market, reputation is hard cash.

Consultancies love to show off the big names on their client list. So as a client, I would ask one simple question: what exactly has my consultant built into my systems, and what happens to it when the vendor switches off a model tomorrow? Anyone who answers “trivial” hasn't understood your risk. Or doesn't care.

That's the point. Not whether switching is possible. But that you have to protect yourself against deprecations, price jumps and outages, no matter how easily the switch works in the end. That takes effort. It costs money. It is not trivial. Telling a client “trivial” takes away exactly the budget they will need later.

Only the key is trivial. Everything behind it is expensive.

Experts in Everything

Classic consulting has a remarkable talent: it is always already the expert in whatever is hot right now.

First it was communication across hierarchy levels. Then processes, lean, reengineering. Then change management. Then digitalization, digital transformation, agility. Now AI, with sustainability, resilience and whatever next quarter brings on the side. The topics change every few years. The slide deck stays.

New label, same bottle
  1. Communication across levels
  2. Lean and reengineering
  3. Change management
  4. Digitalization
  5. Agility
  6. AI
  7. next quarter: ?

Same people, same certainty, instant experts every time.

What's remarkable is who the experts were each time. Often the same people. Whoever explained communication across levels yesterday explains AI agents today, with the same certainty. That someone becomes an instant expert in something completely new every few years is unlikely enough on its own. Titles fill the gap. Within a few years the analyst becomes a senior consultant, an engagement manager, an associate partner, a principal. Titles that sound like decades and claim competence where there is often just practice at performing. And people celebrate themselves for it, on LinkedIn and on their CV: change expert, digitalization expert, AI expert. Serial expert, a new field every three years. Other professions call that a career change.

What stands out: the people giving the advice rarely built any of it themselves. They never ran software, never ran a factory, never put a model into production and fixed it at three in the morning. They told the client how it's done. The client had to do it.

Of course the big firms build things too now: BCG with BCG X, McKinsey with QuantumBlack. But that doesn't mean the partner making a specific recommendation in the steering committee has ever built or run anything like it. Someone writing code somewhere in the firm doesn't make the advice at the table any more practical.

To be fair: consulting used to offer something many companies lacked. It carried information through hierarchies, got departments around one table and pushed decisions nobody inside wanted to touch. That had value.

But communication is not deep technical experience. If you want a say on AI, you have to know how a model behaves in production, what a switch really costs and where it breaks. You don't learn that in a steering committee. Pumping hot air through the hierarchies and going to lunch until a “decision” emerges is not a future. It's bureaucracy with a day rate.

How well that works, the industry says itself. In an interview on McKinsey's website: “The academic research is really clear that when corporations launch transformations, roughly 70 percent fail.” That applies to transformations in general, not only those with consultants involved. But transformations are consulting's core business. So it is remarkably candid for a firm that sells them.

The 70 percent are, of course, always someone else's clients. Your own clients you lead to success. And if it went wrong anyway, the client should have bought more consulting.

Consultancies stand for everything. We can do everything, we know everything. But an opinion that fits every topic is as wipe-clean as the suits are slick. No special knowledge, no risk, no value. Slides and hot air.

An opinion that fits every topic is as wipe-clean as the suits are slick.

Smoke Screens in Slide English

Then there's the language. The longer the buzzword, the more complex the problem looks, and the more indispensable the person saying it. Translated into plain words, much of it sounds remarkably simple:

Smoke screen translator
On the slideIn plain words
Target Operating ModelHow the company should work from now on.
Value Creation RoadmapA plan for how we make money.
AI-enabled Transformation JourneyAI project.
Stakeholder AlignmentEveryone agrees.
North StarGoal.
Capability BuildingTraining.
Quick WinsWhat was planned anyway.

Good consulting makes a problem simpler. It explains it so the plant manager understands it, and the clerk too. Bad consulting makes it more complicated, so that only it can solve it. It doesn't explain. It lays down smoke screens.

The word isn't mine, by the way. In the thread, the partner used it for exactly the switching costs the partner's own firm had warned about a week earlier:

„Im übrigen [sic] sind die von Ihnen genannten "Wechselkosten" eine Nebelkerze.“

(“Incidentally, the ‘switching costs’ you mention are a smoke screen.” Translated from German.)

What counts today is the opposite. Clear words. And the courage to say openly: “I was wrong.” Then you sit down with the client and find the solution together. In the thread, I told the partner several times where the partner had a point. That costs nothing. It makes the rest more credible.

Anyone who thinks they know everything and proves themselves right with nebulous words is not an expert. They're a risk with a business card.

Good consulting makes a problem simpler. Bad consulting makes it more complicated, so only it can solve it.

Zero Accountability

Now the real question. What is a consultant liable for when the advice was wrong?

Legally, consultants can be liable for culpable breaches of their consulting contract. In practice, though, it's hard for a client to trace a failed project back to a specific breach. The advice was a recommendation, the client implemented it. What gets paid for is the service, not the success. The fee is certain, the benefit isn't.

And here lies a function of consulting that is rarely discussed openly: it gives executives a justification. The decision has often long been made. The report comes afterwards. If it goes well, it was leadership. If it goes wrong, it was the consultants' recommendation. A liability umbrella with a day rate.

What comes of it is on the record:

The balance sheet: delivered, paid, liable?
CaseCostWhat was deliveredConsequence
Deloitte Australia, 2025about A$440,000A review with sources that don't exist and a fabricated quote from a court rulingFinal installment refunded, less than a quarter
Deloitte Canada, 2025about C$1.6 millionA 526-page health workforce plan. At least four of the cited studies don't existProfessional regulator investigating, no known refund
KPMG, 2026own publicationReport on AI agents. UBS: “factually incorrect,” Transport for London: “misleading.” According to GPTZero, 5 of 45 sources correctReport withdrawn
EY Canada, 2026own publicationStudy with 16 of 27 sources fabricated, misattributed or deadStudy withdrawn
MD Anderson, 2013 to 2017more than $62 million, $21 million of it to PwCCancer project with IBM Watson, in the end “not in clinical use”According to the audit, invoices paid in full, delivered or not

Deloitte Australia came with an extra twist. In the same week, Anthropic announced that Deloitte was rolling out Claude to more than 470,000 employees. Australian Senator Deborah O'Neill said that buyers who don't check who actually does the work would be “better off signing up for a ChatGPT subscription.”

And Rotterdam spent years scoring welfare recipients with a risk model for benefit fraud, according to WIRED originally developed by Accenture, which praised the approach as “an example to other cities.” Investigations showed that the model discriminated against people based on origin and gender. In 2021, the city stopped using it.

Consultancies sell AI expertise and deliver reports whose sources an AI made up. Clients only notice when a professor, a journalist or a senator actually reads them.

And ethics? In 2024, McKinsey paid $650 million to the US Department of Justice. According to the department, the firm had helped Purdue Pharma “turbocharge” sales of the painkiller OxyContin, in the middle of the opioid crisis. A former partner pleaded guilty to destroying related records.

The research fits the picture. In Britain's National Health Service, hospitals that spent more on consultants became less efficient on average, not more.

The consultancy writes the press release. Someone else writes the audit report. The client pays for both.

AI Hits the Consultants' Best Customer

Now for the part consultancies don't like to talk about. AI doesn't just replace the junior work in consulting. It also hits the people who bought consulting.

Big strategy mandates come from the board. But in my observation, many other projects came in through middle management: division heads, department heads, program leads who had a budget and needed a justification. Anyone who has worked in a large corporation knows the pattern. Whoever moved up took their consultants along. The project grew, and so did the mandate.

That is exactly the layer AI is supposed to thin out. In 2024, Gartner predicted that by the end of 2026, one in five organizations would use AI to flatten its structure and eliminate more than half of its middle management positions. That's a forecast, not a measurement. And it isn't irreversible: according to another Gartner forecast, by 2028 three quarters of the organizations that cut such roles will rehire at least half of them. Reports, planning, coordination and status meetings are still being automated, and my forecast is this: with them, demand shrinks for the consulting that used to deliver exactly that work.

With a thinner middle layer, one of the channels consulting came in through is gone. And so is the need for the justification it was paid to provide.

Who Still Signs Up

The business used to be easy to explain. Hire McKinsey, BCG or Bain, and you rented the best graduates of the year, who spent their nights gathering data and building slides. The partners' knowledge plus the pyramid's labor.

AI is now taking over the pyramid's labor. According to Revelio Labs, the big firms' demand for consultants is about 40 percent below 2023. Entry-level hiring has fallen by about 10 percent since its 2023 peak, and there are now more AI roles than entry-level roles at these firms. Starting salaries, according to the Financial Times, have been frozen for the third year in a row. McKinsey's technology chief Kate Smaje: “Do we need armies of business analysts creating PowerPoints? No, the technology could do that.” And according to Bloomberg, around 150 former consultants from McKinsey, Bain and BCG are training AI models on exactly the tasks their careers started with.

And the elite? They still sign up. But no longer to stay. At Stanford GSB, only 11 percent went into consulting in 2025, 35 percent into technology. At Harvard Business School, tech overtook consulting for the first time. And of Harvard College's class of 2025, one in seven of those going straight into work starts in consulting. But less than one percent of all respondents want to still be there in ten years. That's nothing new and not an AI effect. But it shows what consulting has long been for many: a transit stop, a logo on the CV, not a destination. And so the logo loses weight too. If the best only stay two years, the client paying the big brand's day rate increasingly gets the ones who stayed.

Add to that a generation that asks what its work is for. Slides for a decision that has already been made answer that question poorly.

What's left shows up in the project. There's the consultant with three weeks of onboarding, sitting next to the engineer with fifteen years at the plant. He has the stage presence. She has the knowledge. Who is explaining the world to whom here?

The bosses see it differently. McKinsey head Bob Sternfels in July 2026: “Maybe I'm nuts, but when I think about the rest of this decade, I think there's a potential that this is a golden age for us.” The numbers: McKinsey's revenue has been stuck at $15 to $16 billion for five years. The German consulting market grew by 0.5 percent in 2025. The BDU, the German association of management consultancies, had forecast 6.4. Lünendonk, a German market research firm, titled its 2026 ranking „Konsolidierung statt Wachstum“ (“Consolidation instead of growth”). Revenue per head is falling, even though AI supposedly makes everything more efficient. For an industry that sells intelligence, foresight and genius, five years of stagnation is no advertisement. If a client had been treading water for five years, a team would be on site tomorrow, with a transformation program.

The numbers behind the golden age
German consulting market 2025, growth
Forecast by the industry association BDU6.4%
Actual0.5%
Stanford GSB, class of 2025: where to?
Technology35%
Finance33%
Consulting11%
Harvard College, class of 2025: consulting
of those starting work begin in consulting14%
of all respondents want to be there in ten years< 1%

Sources: BDU via Börsen-Zeitung, Stanford GSB via ClearAdmit, Harvard Crimson Senior Survey.

The best punchline: in early 2026, according to the Financial Times, KPMG pressed its own auditor to pass on AI savings and threatened to switch. The fee dropped. The moment consultants are the client, they believe firmly in the AI discount.

An industry that has spent decades explaining disruption to its clients can't recognize its own.

At the Next Table

I live in Munich. The city is a consulting hotspot, and if you eat lunch in the right places, you get the industry's successes thrown in for free. At full volume, from the next table. Which client, which project, which board is currently “completely out of its depth.”

After the second glass of white wine, discretion is gone.

For the client, this is no anecdote. It's a danger. Their numbers, their plans, their weaknesses, served between vitello tonnato and espresso, audible to anyone who happens to sit nearby. As long as the ego is front and center.

And it fits the thread. Loud when it comes to their own successes. Deaf when someone asks questions. The thesis is set, the disagreement “thoroughly refuted,” the evidence stays in the suit. Whoever behaves like that can no longer be reached by whatever proves them wrong.

Loud about their own success, deaf to disagreement. That's not a consulting style. That's a risk for the client.

What Works Instead

If you're reading “outside help is useless,” you're reading it wrong. The research shows fairly clearly what works.

In a famous field experiment, Nicholas Bloom and colleagues brought consultants into Indian textile factories. Productivity rose 17 percent in the first year, and the lead still held years later. In Mexico, small and medium-sized firms had about 50 percent more employees five years after subsidized consulting. Both were hands-on work on the shop floor: quality, inventory, maintenance, processes. No strategy slides.

MIT's NANDA project found in 2025 that AI projects with external partners make it into regular operation twice as often as purely in-house builds. The study is small, relies on self-reports and doesn't prove causation. The two field experiments above are more robust. But the direction fits: outside help isn't the problem. The problem is which kind.

What works is close to operations, measurable and done by people who know the problem firsthand. What doesn't work is opinion as a product. The dying model isn't outside help. It's the suit: titles instead of evidence, pyramid instead of practice, performance instead of liability.

How It Can Work Differently

That's why I deliberately work differently. With Balane, my software company in Munich, we build things ourselves: no white label, nothing outsourced. And with KI-Managementsolutions, KIMS for short, in Switzerland, we help companies introduce AI responsibly. Small, focused, and with three perspectives at one table that big firms spread across three departments:

  • Ethics. An ethicist whose subject is people in digital change. Not as a fig leaf at the end, but as a question at the start: should this system be allowed to do what it's meant to do? Rotterdam shows what happens when nobody asks that question.
  • Governance and business economics. Structural change, processes, an AI management system under ISO/IEC 42001. Every recommendation comes with a bill, including the one for tomorrow: what does it cost if the vendor raises prices, deprecates a model or goes down?
  • Technology and implementation. That's my part: automation, integration, AI in live operation. Anyone who has switched models, readjusted prompts and rewritten tests themselves doesn't say “trivial.”

On top of that comes a network of partner firms that know our clients' problems from their own work. That's the real difference. A partner from practice has a completely different motivation than a consulting firm. They don't live off the follow-up mandate. They know the mistake because they made it themselves. And they tell you what didn't work, because they have nothing to sell.

We don't sell hours. We sell results. Not claimed on a slide, but measured, and measured by the client. That makes our job harder, and the risk sits with us. If we don't deliver, the client sees it immediately. Success is part of our offer, not part of our marketing.

We can't do everything. But where we work, we give the client a commitment on the result. If it doesn't work, we have to get better. Not the client buy more. That forces us to swim ahead of the wave instead of chasing it, and to think hard about what makes sense and what doesn't. Including where AI makes no sense.

And we stay close to the client. No partner who sells and a team that delivers. The people a client talks to are the people doing the work. We speak plain language, not slide English. When we're wrong, we say so before the client notices. And what we know about clients stays with us, even after the second glass.

And we write down what we think. This blog is part of that: what I thought, how I got there and where I think differently today. Changing your mind is good if there's a reason and you name it. Switching constantly without a reason isn't flexibility. It's arbitrariness.

The Snake Oil Test

You don't spot snake oil by its label. You spot it by the fact that nobody checks. So check, with everyone who wants to sell you advice on AI. Including us. Six questions before you sign:

  1. What does your own firm say about this? Put its publications next to the proposal. If they contradict each other, ask which line applies to your project. In writing. Opinions may change, but for a reason and traceably. Anyone who can't explain why they're saying something different today than last week has no line.
  2. What have you built yourself? Not recommended, not supported. Built and run. Anyone who can't show anything has nothing. They're an armchair pilot: explains flying without ever having taken off.
  3. Who does the work, and with which AI? Demand that every source be verifiable. And ask how the sources were chosen. What is advice worth if there are 200 theses and you pick whichever one fits at the moment? If all you want is a justification for your prejudices, you don't need a day rate of 2,000 euros. A ChatGPT subscription for about 20 euros a month will do.
  4. Does your advice still hold in three months? If the answer is “posts age,” negotiate the fee accordingly.
  5. What do you earn from, and what are you liable for? Hours, results, a vendor? The answer says more than any reference.
  6. When were you last wrong? And what did you do then? Anyone with no answer to that is wrong right now.

Anyone who won't answer these questions is selling snake oil. No matter how well the suit fits.

In Closing

It started with a question about AI costs. The answer was that it's trivial, and anyone who disagrees hasn't thought it through. If I wanted to be advised, I could always commission a project.

I think many companies have thought exactly that through by now. They want to be advised. Just not like this anymore. And not by people who tell them they haven't thought it through.

That's why consultancies are fighting for their place right now. In their position papers, they demand that clients have a clear line, measured results and an exit plan. None of that applies to their own advice. It may age, contradict itself and turn with the wind.

Changing your opinion with every wind is not expertise. It is intellectual bankruptcy.

The dying model isn't consulting. It's the suit that mistakes condescension for expertise and hasn't noticed its clients now read the fine print.

For context

What this is. This is a polemic. It targets a business model, not people. Many smart and decent people work in consultancies. “Snake oil” is an image for promises that cannot be kept.

What is documented. The partner's quotes appear verbatim in the original German, with all their errors. The English translations are mine. They come from a public LinkedIn thread under my post of October 12, 2026. I don't name the partner, because this is about the pattern. All cases described involve other firms and are sourced below. The Munich scenes I witnessed myself.

What's in it for me. I'm not neutral. In Munich I run Balane, a company for software development and AI, and alongside it KI-Managementsolutions (KIMS) in Switzerland. I have a financial stake in who helps companies with AI. That's why the test is up there. It applies to us just the same.

Update, October 26, 2026. I went back and counted: 41,594 texts from 28 consulting firms, plus the staffing, divestments and acquisitions of the big firms. The data is in Quo vadis, consulting?.

Sources

The thread

  • LinkedIn post „Was kostet KI heute? Und was kostet sie morgen?“ (“What does AI cost today? And what will it cost tomorrow?”), Oct 12, 2026, with comments from Oct 13, 2026. linkedin.com

The partner's firm

  • Whitepaper “Closing the AI ambition gap,” Oct 6, 2026, pp. 5 and 8. PDF
  • “The hidden cost of 'AI enabled',” Jul 17, 2026. argonandco.com

Switching between models

  • Sclar et al.: “Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design,” ICLR 2024. arxiv.org
  • Wang et al.: “PromptBridge: Cross-Model Prompt Transfer for Large Language Models,” Dec 1, 2025. arxiv.org
  • Zapier survey, Apr 2, 2026 (via IT Brief). itbrief.news

Consulting, liability, ethics

  • McKinsey: “Why do most transformations fail? A conversation with Harry Robinson.” mckinsey.com
  • Fierce Pharma: McKinsey, $650 million to the US Department of Justice, December 2024. fiercepharma.com
  • Kirkpatrick et al.: “The impact of management consultants on public service efficiency,” Policy & Politics 2019. bristol.ac.uk

Cases

Middle management

  • Gartner: Top Strategic Predictions for 2025 and Beyond, Oct 21, 2024 (via Network World). networkworld.com
  • CV Viverito (Gartner) in HR Executive: middle management rehiring by 2028. hrexecutive.com

Talent, pyramid, market

What works

  • Bloom et al.: “Does Management Matter? Evidence from India,” QJE 2013. RePEc · Bloom et al.: “Do Management Interventions Last?,” AEJ Applied 2020. aeaweb.org
  • Bruhn, Karlan, Schoar: “The Impact of Consulting Services on Small and Medium Enterprises,” JPE 2018. MIT DSpace
  • MIT NANDA: “The GenAI Divide,” July 2025. PDF

Part of: Advice is worth exactly as much as its evidence.

Jonas Höttler builds software for mid-sized companies with Balane.