What AI Does Poorly — and Why It Matters

AI writing failures are predictable once you understand the model's fundamental limitations. AI is a pattern matcher trained on text. It predicts what the next token should be based on patterns in its training data. It does not know whether the claim it is making is true. It does not know the organization's actual strategy, values, or relationships. It does not know what was said in last week's client meeting. It generates the most statistically likely continuation of the text it has been given — and the most statistically likely continuation is frequently the most generic one.

1. Factual accuracy AI generates factual-sounding claims with high confidence regardless of whether those claims are accurate. Statistics, names, dates, company facts, and technical details are all potential sources of AI fabrication. Every factual claim in an AI-generated draft must be verified against primary sources before the document is distributed. This verification is not optional for high-stakes business writing — it is the minimum standard of professional responsibility.
2. Contextual judgment AI does not know what the client relationship history is. It does not know that the recipient of this email was just passed over for a promotion and is in a fragile professional state. It does not know that the competitor mentioned in the analysis is the recipient's former employer. Contextual judgment — the kind that comes from actually knowing the situation — is the dimension where AI most reliably fails in business writing. It writes for a generic professional context, not for the specific one the writer inhabits.
3. Voice and register AI defaults to a competent, professional, neutral voice. This voice is appropriate for many document types and entirely wrong for others. A client relationship that has been built on informal, direct communication will be damaged by AI-generated correspondence that reads as formal and slightly distant. A board memo that needs to carry the credibility of a specific executive's voice should not read like a template. AI cannot match individual voice without extensive fine-tuning — and even then, it approximates rather than replicates.