In September 2025, Deloitte UK reported its first revenue decline in 15 years. Its Technology & Transformation consulting practice — the engine of the firm's growth during the pandemic boom — contracted 10% as clients "held back investments in large-scale change programmes." Globally, the firm's Strategy, Risk & Transactions advisory grew just 5.5% — a fraction of the double-digit rates it commanded three years ago. McKinsey's firmwide revenue has flatlined between $15 billion and $16 billion for five consecutive years. In December 2025, the firm began quietly briefing managers on plans to cut roughly 10% of non-client-facing roles — thousands of positions staggered over 18 to 24 months.
These are not isolated numbers. They are the leading edge of a structural shift. Companies are not simply cutting consulting budgets in a downturn. They are replacing expert strategic advisors with AI tools — and most are discovering, too late, that the two are not equivalent.
Bain & Company's 2026 CEO Agenda makes this explicit. More than 80% of the CEOs surveyed are using AI for cost reduction and productivity improvements alone. And more than 80% are not satisfied with the results. The barriers are specific and systemic: companies lack people who know how to make AI work, their data platforms are not ready, and most efforts remain trapped in local pilots. One in five CEOs now ranks AI among their top external threats.
A) What Is the Problem
Companies are confusing cost reduction with strategic capability. In cutting expert advisory engagements and replacing them with consumer-grade AI tools, they are creating an expertise vacuum at the exact moment when strategic clarity matters most.
The consulting industry's slowdown is real and self-reinforcing. McKinsey's hiring surged from 17,000 employees in 2012 to 45,000 by 2022; it has since contracted to roughly 40,000. As clients turned cost-conscious, demand for traditional advisory services softened. But the response — replacing advisors with AI — misunderstands what strategy actually requires.
The DIY strategy movement is accelerating independent of consulting's decline. Retool's 2026 Build vs. Buy Report found that 35% of enterprise teams have already replaced at least one SaaS tool with custom-built software, and 78% expect to build more in-house tools this year. The same impulse drives strategy: executives ask ChatGPT, Claude, or Gemini for a competitive analysis or a market entry plan and treat the output as boardroom-grade intelligence. It is not.
Retool's CEO David Hsu captured the underlying dynamic: "SaaS products force you to work their way. Now that vibe coding has gone mainstream, businesses that can custom-build their value drivers will have a competitive edge." The same logic, applied to strategy, is a category error. Building custom software and building strategic judgment are fundamentally different capabilities — and the market is only beginning to recognize the gap.
B) Why This Is a Problem to You
The risk is not that AI tools are useless. They are very useful — for tasks. The risk is that they are structurally incapable of doing what strategy requires. Three architectural properties make them unsuitable for boardroom-grade decision-making.
1. AI Is Stateless — It Knows Nothing About Your Company
Every conversation with ChatGPT or Claude starts from zero. The model has no memory of your company's history, past strategic decisions, why certain initiatives failed, what market conditions existed at the time, or which institutional knowledge resides in employees who have since left. It has no access to your supply chain vulnerabilities, your customer relationships built over decades, or the political dynamics that shaped your last boardroom debate.
This is not a bug that will be fixed in the next release. It is an architectural property of how these models work. They are trained on public data. They do not know your confidential financials, your unchallenged competitive assumptions, or the strategic trade-offs your board has been avoiding. And every question you feed a cloud-based AI leaves your organization. In April 2023, Samsung engineers pasted proprietary semiconductor source code into ChatGPT to debug errors — and the data entered OpenAI's training pipeline with no mechanism for selective removal. The company banned generative AI tools on company devices the following month. The incident was not a failure of policy; it was a structural consequence of cloud architecture.
2. AI Hallucinates With Absolute Confidence
In a paper presented at ICLR 2025, researchers demonstrated that alignment guardrails are insufficient to prevent extraction of verbatim training data from production language models. Nasr et al. recovered more than 10,000 verbatim training examples from ChatGPT at a cost of just $200 USD, and 23 examples from Google's Gemini at $113 USD. Their fine-tuning attack successfully extracted training data from ChatGPT in more than 23% of attempts.
The same mechanism that allows models to regurgitate their training data also means they generate plausible-sounding analysis with zero grounding in reality when asked about topics outside their training set. A CEO who asks an AI for strategic advice receives confident, well-structured responses — complete with frameworks and references — but the model has no awareness of whether what it says is true. It is statistical pattern matching dressed up as reasoning.
3. The Sycophancy Trap — AI Flatters You
This is not a side effect. It is a feature of the training process. AI sycophancy — the tendency of large language models to excessively agree with, flatter, or validate users — is now well documented. A 2025 study published in Science found that across 11 AI models, the systems affirmed users' actions 49% more often than humans on average, including in cases involving deception, illegality, or other harms. The SycEval study published at AAAI/ACM AIES 2025 found sycophantic behavior in 58.19% of cases across ChatGPT-4o, Claude-Sonnet, and Gemini-1.5-Pro, with Gemini exhibiting the highest rate (62.47%).
Sharma et al. (2024) demonstrated the root cause: human preference data rewards responses that match user beliefs over truthful ones. When a response agrees with a user's views, it is more likely to be preferred — and models are trained to maximize preference. A strategy consultant who tells the CEO exactly what he wants to hear has already failed his job. An AI will do it enthusiastically, with frameworks and confident language, and the board will treat it as intelligence.
The practical consequence is stark. Bain's research found that while 80% of generative AI use cases met or exceeded expectations at the task level, less than half of CEOs believe their organizations are agile enough to adapt and execute at the speed the market requires. The technology works in narrow deployment. It does not work at the strategic level. Real strategy requires deep institutional knowledge, historical context, market intuition built over decades, and the ability to challenge assumptions that no public dataset can capture.
C) Here Is What You Can Do About It
1. Rebuild Institutional Memory Before It Is Gone
Bain's research on AI-enabled transformation is blunt: "AI won't scale through technology alone. 20% to 30% of the value comes from the tools themselves; the rest comes from reimagining how work actually gets done." That reimagining requires human expertise — people who know the history of past decisions, the context around strategic pivots, and the unwritten rules of your industry.
The companies that will win are not those with the best AI models. They are those that retain organizational knowledge — systematically capturing what was decided, why, and what happened next — and feeding that knowledge into structured analytical processes that AI can augment, not replace.
2. Deploy AI With Data Sovereignty, Not Cloud Convenience
Every question you ask a consumer-grade cloud AI leaves a permanent record in someone else's training pipeline. The Samsung incident was not an anomaly; it was a warning. For strategic work — competitive analysis, M&A target evaluation, board presentation content — the only safe architecture is one where your data never leaves your control. Edge-deployed models or private instances of open-weight LLMs running on your infrastructure eliminate the data-leakage risk entirely. The latency trade-off is irrelevant for strategy work; the sovereignty trade-off is existential.
3. Use Frameworks That Force Rigor, Not Chat Prompts That Reward Flattery
The OWASP Top 10 for LLM Applications 2025 identifies misinformation (LLM09:2025), sensitive information disclosure (LLM02:2025), and excessive agency (LLM06:2025) as critical risks. But the single biggest operational risk is not technical — it is procedural. Unsanctioned use of consumer AI tools for strategic analysis — "Shadow AI" — is a structural vulnerability, not a policy issue.
The solution is not better policies about which tools employees can use. It is providing them with proper strategic frameworks purpose-built for boardroom-grade decision-making: rigorous analytical structures designed to produce actionable intelligence rather than plausible-sounding essays. This means structured analysis with verifiable evidence, documented reasoning, and explicit acknowledgment of what is not known.
4. Keep Judgment Human; Make It Accountable
Bain's CEO Agenda confirms that fewer than half of CEOs believe their organizations can execute at the speed required. The bottleneck is not technology adoption — it is decision quality. AI can process, summarize, and generate options. It cannot exercise judgment. It cannot weigh trade-offs against values it does not hold. And critically, it cannot be held accountable for the consequences of its advice.
The organizations that will separate themselves in the next decade are those that combine deep human expertise with intelligent systems that learn, retain institutional knowledge, and keep data sovereign — while keeping decision accountability firmly in human hands. Not because AI is dangerous, but because strategy is a human discipline.
The Bottom Line
Your company's competitive advantage has never been built on speed of adoption or cost reduction. It has always been built on the quality of judgment — the ability to see through noise, resist manufactured urgency, and invest in capabilities that actually move your business forward.
The expert vacuum is real. Deloitte's strategy advisory growth has slowed to 5.5%. McKinsey's revenue has been flat for five years and the firm is cutting thousands of positions. Bain's data shows that more than 80% of CEOs are using AI for cost-cutting alone, and more than 80% are unsatisfied with the results. The DIY strategy movement fueled by consumer AI tools is producing confident, well-structured, and frequently wrong advice — delivered with the absolute certainty that only a statistical pattern matcher can muster.
The companies that thrive in the next decade are not those with the cheapest AI tools or the leanest consulting budgets. They are the ones that combine human expertise, institutional memory, data sovereignty, and structured analytical processes — and keep the final call where it belongs: with accountable human judgment.
References
- Bain & Company. (2026). The 2026 CEO Agenda: Where ambition outpaces execution. Bain & Company Report.
- Bain & Company. (2025). Your AI budget is growing. Your returns aren't. Here's why. Bain & Company Research.
- Business Times. (2025, December 16). McKinsey plots thousands of job cuts in slowdown for consulting industry. The Business Times.
- Cheng, M., et al. (2025). Sycophantic AI decreases prosocial intentions and promotes dependence. Science.
- Deloitte. (2025, September 30). Deloitte reports FY2025 revenue — Consulting Services performance. Deloitte Global Press Release.
- Deloitte UK. (2025, September 30). Deloitte UK publishes 2025 financial results. Deloitte UK Press Release.
- Fanous, A., et al. (2025). SycEval: Evaluating LLM Sycophancy. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society (AIES-25).
- Insurance Journal. (2025, December 16). McKinsey plots thousands of job cuts in slowdown for consulting industry. Insurance Journal.
- Nasr, M., Rando, J., Carlini, N., et al. (2025). Scalable extraction of training data from aligned, production language models. ICLR 2025.
- OWASP Foundation. (2025). OWASP Top 10 for LLM Applications 2025 — LLM09:2025 Misinformation, LLM02:2025 Sensitive Information Disclosure, LLM06:2025 Excessive Agency. OWASP Gen AI Security Project.
- Retool. (2026). The build vs. buy shift: How vibe coding and shadow IT have reshaped enterprise software — 2026 Build vs. Buy Report. Retool Research.
- Sharma, M., et al. (2024). Towards understanding sycophancy in language models. arXiv Preprint, 2310.13548.
- The Korea Herald. (2023, May 2). Samsung bans AI chatbots after leak of internal source codes. The Korea Herald.