AI-generated consulting reports can be trusted only to the extent that a named person has verified every source and signed the conclusion. In 2025 and 2026 the largest firms in the profession published reports with fabricated citations, and 70 percent of clients now say they would not trust a report prepared with AI.
What happened
In October 2025, Deloitte Australia agreed to partially refund the Department of Employment and Workplace Relations for a report that cost A$440,000. An academic reviewing the document found a quotation attributed to a federal court judgment that did not exist and citations to academic papers that had never been written. The corrected version, which disclosed that Azure OpenAI GPT-4o had been used in its preparation, addressed roughly twenty errors. The Associated Press covered the refund.
Nine months later, in July 2026, the Financial Times verified research from GPTZero showing that four published reports from PwC Middle East contained footnotes that pointed to sources which could not be found. By that point, industry observers noted that every one of the Big Four had been caught publishing generative content with invented references.
These were not obscure firms or careless juniors. They were the most trusted brands in professional services, and they published invented evidence under their own names.
Why it keeps happening
Three forces combine. First, the large firms have invested more than $10 billion in AI since 2023 and have deployed assistants to tens of thousands of staff. McKinsey's internal tool handles roughly 500,000 queries a month. Second, they have cut the junior layer that used to check the work: UK graduate intakes at KPMG fell 29 percent, at Deloitte 18 percent. Third, the people using these tools are not verifying the output. The KPMG and University of Melbourne trust study of 48,000 people found that 58 percent of employees rely on AI output without evaluating it, 57 percent had made mistakes as a result, and 53 percent had presented AI content as their own.
A language model produces text that looks like a citation because citations are what the training data looks like at that point in a document. It does not check whether the source exists. If nobody in the chain opens the source, a fabricated reference goes to the client.
What clients now believe
In August 2026 Source Global Research published a survey of 3,887 companies. Seventy percent said they would not trust a report they knew had been prepared with AI. Thirty-two percent said their trust in a firm would be shaken if it used AI for public-facing reports. The most striking result was about exposure: among clients who had been shown or had used their consultants' AI tools, 77 percent said AI was a bubble, compared with 55 percent of those who had not been exposed. Seeing the tools in action made clients more doubtful, not less.
"We're not going to keep paying $500K for a report that we know was generated largely by a machine."Chief Information Officer, quoted by FutureOfConsulting.ai, January 2026
The right question is not "did they use AI"
More than three quarters of UK consultancies now use AI in delivery, according to the Management Consultancies Association. Refusing to work with any firm that uses AI is not a practical position, and it misses the point. The Deloitte report was not wrong because a model drafted it. It was wrong because no human opened the sources before it shipped.
The question to ask is whether the firm can show you the chain of evidence. Can they name the people they spoke to? Can they produce the transcript? Did a human open every cited document? Who, by name, is accountable for the conclusion?
Five questions to ask any adviser
- Where did AI touch this work, specifically? A firm with a policy will answer in one sentence. A firm without one will talk about "responsible innovation".
- Can I see the sources? Ask for the interview list, verification method and document references before the engagement starts, not after.
- Who verified each citation, and how? The honest answer names a person and a process. "Our quality process" is not an answer.
- Who signs the recommendation? If the answer is "the firm", nobody does.
- What is your commercial relationship with the vendors you are advising me on? McKinsey, Bain, Capgemini and PwC all have formal partnerships with OpenAI or Anthropic. That does not make their advice wrong, but you should know.
How Theory handles this
Theory Intelligence publishes The Theory Standard, a one-page policy that commits every deliverable to a provenance appendix, a verification log and a named signatory. AI is used for scheduling, transcription, coding and drafting. It is never the source of a factual claim, and no citation reaches a client that a human has not opened. If you want to see what that looks like in practice, a Signal Brief takes ten business days.
Sources
- AP News, Deloitte partial refund, October 2025
- The Telegraph, Consultants face backlash over AI plans, 3 August 2026
- Resultsense, PwC reports carried fake citations, 29 July 2026
- KPMG and University of Melbourne, Trust, Attitudes and Use of AI, 2025
- FutureOfConsulting.ai, 2026 AI revolution update