The Honest Executive's Guide to What AI Cannot Do

Current AI systems are extraordinarily good at pattern recognition and at generating outputs that are consistent with patterns in their training data. They are significantly less reliable when the situation is genuinely novel: when the right answer requires reasoning that goes beyond pattern matching to first-principles analysis.

This limitation has direct business implications. AI can summarise existing research very effectively. It is significantly less reliable at generating genuinely original strategic insight that is not derivable from patterns in existing content. AI can apply existing frameworks to familiar situations effectively. It is less reliable at recognising when a situation genuinely falls outside the frameworks it has been trained on.

For executives using AI for strategic analysis, this means that AI outputs that appear analytical are actually patterns of what analytical outputs look like, not genuine analysis from first principles. The human judgment required to recognise this distinction is itself a critical AI governance skill.

The hallucination problem is well-established but its practical implications for business use are still frequently underestimated. AI systems generate text that is linguistically fluent and contextually plausible. They do not have access to verified factual databases in the way that a search engine retrieves indexed content. When they generate specific facts, dates, citations, financial figures, and regulatory requirements, those outputs can be wrong in ways that are hard to detect without independent verification.

The practical implication is that AI outputs in factually sensitive contexts (legal, compliance, financial, medical) should be treated as drafts that require expert verification, not as authoritative sources. Organisations that have deployed AI in these contexts without adequate verification workflows are carrying risk that may not yet have crystallised in visible incidents.

One of the most persistent misunderstandings about AI capability is the belief that access to large language models reduces the value of deep domain expertise. In practice, the opposite is often true. The ability to evaluate AI outputs critically, to identify when AI has applied the wrong framework or made a plausible-sounding error, requires exactly the deep domain knowledge that AI is supposed to be making less necessary.

A junior lawyer who accepts AI-generated contract analysis without the domain expertise to assess whether the analysis is correct is not being augmented by AI. They are being exposed to AI-generated errors that they lack the knowledge to identify. The value of AI in professional contexts is realised when it is used by people who have sufficient expertise to use it critically, not when it is substituted for expertise that does not exist.

AI systems perform best with clear, specific, well-structured instructions. They handle genuine ambiguity less reliably. When an instruction is ambiguous, an AI system will typically make a choice rather than flagging the ambiguity and asking for clarification, and the choice it makes may not align with the user's actual intent.

In business contexts, the instructions given to AI are often ambiguous in ways that users do not recognise as ambiguous. Asking an AI to summarise the key risks in a document is ambiguous: key according to what criterion, at what level of detail, for what audience? The AI will make choices about these parameters that it will not make visible to the user. Understanding and managing this characteristic requires AI users to be much more precise in their instructions than most have been trained to be.