What Is Vibe Consulting? The Risks of AI-Driven Advisory Work

In February 2025, Andrej Karpathy coined the term vibe coding. Vibe coding is where you describe, in natural language, what you want AI to build and watch the AI generate code, and accept what comes back without reading the diffs. Collins Dictionary made it Word of the Year for 2025. By late that year, developers on LinkedIn were renaming themselves "vibe code cleanup specialists" as a joke that had gotten all too real.

A parallel pattern is forming in consulting. Some have started calling it vibe consulting. The category is still being defined, so let's start there.

What Vibe Consulting Actually Is

Vibe consulting is the practice of using AI to extrapolate a consultant's expertise. Research, market sizing, competitive analysis, strategic frameworks, and implementation plans, all of which used to require years of pattern recognition, are now a prompt away. At its best, this expands what a small expert team can deliver. At its worst, it lets people who lack the underlying expertise sound like they have it.

The early warning signs are public and not coming from small or unknown firms. In May 2026, Sherwood News reported on a GPTZero investigation that found 60% of the references in a 44-page EY Canada advisory report were hallucinated. Citations linked to articles returning 404 errors. A McKinsey attribution traced back to a "Loyalty Economics Report" that did not exist. The same investigation found 19 hallucinations in a separate Deloitte report. Deloitte later issued a partial refund to the Australian government. A few weeks earlier, the law firm Sullivan & Cromwell acknowledged hallucinations in a filing it submitted to a federal bankruptcy court.

These are not no-name shops. EY's consulting arm generated $16.4 billion in revenue last year. Deloitte and Sullivan & Cromwell are equally established. If it is happening there, it is happening everywhere.

Human Judgment Is Not Going Away

The dominant story we keep hearing is that AI will eventually replace the individual consultant entirely. The data so far points the other way.

In July 2025, METR published a randomized controlled trial on AI coding tools. Sixteen experienced open-source developers worked through 246 real tasks from their own mature projects. Half the tasks were randomly assigned to use AI (Cursor Pro with Claude 3.5 and 3.7 Sonnet), half were not. Before starting, the developers predicted AI would make them 24% faster. After finishing, they estimated AI had sped them up by 20%. The actual measured result: developers using AI took 19% longer to complete their tasks.

Read that twice. Experienced developers, using the best available coding tools on their own code, all believed AI was speeding them up. It was slowing them down. The researchers attribute the gap to validation overhead. AI produced plausible-looking suggestions that needed careful checking, and the checking took more time than the suggestions saved.

It is fair to ask what this means for the rest of this post. If experts get slower with AI, what is the case for using it at all?

The honest answer is that METR does not say AI never helps experts. It says AI does not automatically help experts. The developers in the study used AI on tasks where verification cost more than the assistance was worth. The skill that produces gains, when gains are real, is the skill of choosing which tasks AI should touch and which it should not. That is itself a judgment call, and it is exactly the call that vibe consultants are not equipped to make.

The implication for consulting is direct. The work AI is "doing for you" is work you still have to verify. If you have the expertise to verify it and the judgment to know when to apply AI in the first place, you can compound your output. If you do not, you are publishing whatever the model gives you and hoping. That is vibe consulting, and the EY and Deloitte cases are what it looks like in the wild.

Where AI Actually Helps

This is not a case against AI in advisory work. It is a case against using AI without judgment to know where it can be best applied.

The win, when it exists, looks something like this. The operator decides what to draft from scratch, what to outline before handing off to AI, and when to skip using AI. Inside that frame, AI surfaces precedents from adjacent industries that a consultant would not have searched on their own. It generates counterarguments to a recommendation faster than a workshop room could. The operator filters what is useful from what is wrong, and ships work that would otherwise have taken a team of analysts a month.

The difference is the human in the loop, and the difference before that is which tasks the human hands to the model. An experienced operator will notice that the AI's recommended cost-cutting plan eliminates a small team whose work does not appear in financial reports but quietly keeps the company compliant with regulations. A consultant without that operating background will not. The first uses AI to compound their judgment. The second uses it to manufacture the appearance of having it.

The Overconfidence Problem

There is one more pattern worth knowing about, because it makes the verification problem harder. A Carnegie Mellon study published in Memory & Cognition in July 2025 found that LLMs do not adjust their confidence after performing poorly. They often become more sure of themselves, not less. In one experiment, Google's Gemini predicted it would identify about 10 out of 20 hand-drawn sketches correctly. It got fewer than one right. Afterward, it estimated it had gotten about 14 right.

That is the model you are taking advice from. Even when it is wrong, it sounds certain. Humans signal uncertainty through tone and hesitation. AI does not. A confident-sounding LLM answer with confident-sounding citations is exactly what the EY and Deloitte reports were made of.

The Optimistic Future

None of this is an argument against AI in consulting. It is an argument about who should be using it and how.

PwC's April 2026 AI Performance Study of 1,217 senior executives found that 74% of the economic value being created by AI is being captured by just 20% of organizations. The pattern in that data is consistent. The winners are not the ones with the most advanced tools. They are the ones who know their domain and use AI to do more of what they already do well.

The future of consulting is not vibe consulting. It is quieter and more selective. Experienced operators, choosing carefully where AI earns its keep and where it does not, delivering work that used to require teams several times their size. A small firm with deep expertise can compete with global brands because the volume work, on the tasks where AI actually accelerates it, is no longer the constraint. The expertise is.

The discipline is the one programmers are working out right now. Review what the model produces and test it against your own knowledge of the domain. Trust it where you can verify it.

The consultants who hold that line will scale. The ones who let the model do their thinking will publish reports with citations that do not exist.

The choice is not hard. The value is in requiring a human to make it.