Patterns in AI-generated business writing: what we observe across thousands of drafts

The most common pattern: the framing opener

Across the drafts we work with daily, the single most consistent feature of AI-generated business writing is an opening sentence that frames rather than informs. 'In today's competitive landscape...' 'As businesses navigate an evolving market...' 'It's worth noting that effective communication...' These openers occur in well over half of unedited business drafts we see, regardless of the underlying model. They are statistically the safest way for a model to begin — uncontroversial, on-topic, hard to be wrong about — which is precisely why they read as boilerplate. Cutting them is the single highest-impact edit available.

Sentence length clusters tightly

When we measure sentence-length distribution across unedited AI business drafts, the standard deviation is consistently narrower than in equivalent human writing. The dominant cluster sits between sixteen and twenty-four words, with longer sentences (above thirty words) and shorter sentences (below eight) noticeably underrepresented. Skilled human writing distributes more broadly — short punches and longer developmental sentences appear in roughly the same ratios across genres. The narrow cluster is what readers feel as 'flat' even when they cannot name it.

Generic noun frequency is high

A small set of nouns — solutions, experiences, opportunities, insights, capabilities, strategies, journey, ecosystem — appear in AI drafts at roughly two to three times the rate of comparable human writing. These nouns are not wrong; they are simply the model's safe choice when a specific noun would require knowledge of the specific situation. Replacing one of these per paragraph with the actual concrete referent is consistently the change that most increases reader trust.

Hedging accumulates

AI drafts tend to use multiple hedges per paragraph — 'can,' 'may,' 'often,' 'generally,' 'in some cases.' Each hedge individually is defensible. Their accumulation produces writing that takes no position. The pattern is downstream of the model's training: claims that take a position can be wrong, so the statistically safer output softens them. Edited writing is willing to claim, while marking the genuine uncertainties; unedited AI output softens almost everything.

Bullet lists where prose would serve better

We see frequent overuse of bullet lists in AI business drafts — three or four bullets where a single well-built sentence or a short paragraph would carry the same information with more rhythm. Bullets are visually safe; they signal organisation; they require no transition. They also flatten emphasis. The right rule is: bullets when the items are genuinely parallel and discrete, prose otherwise.

Endings that summarise rather than land

The closing sentence in AI business drafts is almost always a summary. 'In summary,' 'In conclusion,' or an implicit summary that restates what was already said. Skilled human writing more often lands — a short final sentence that delivers a single image, a sharp claim, or a concrete next step. The summary ending is the model's safe choice; the landing ending is the editorial one. The change at the close of a piece has outsize effect on what the reader retains.

What this means for editing defaults

A short, repeatable editing pass — cut the framing opener, break the longest sentence in each paragraph, replace one generic noun per paragraph with a specific one, remove half the hedges, replace bullet lists with prose where the items are not genuinely parallel, replace the summary close with a landing — addresses most of the patterns we observe. The pass takes minutes on a short draft. Teams that build it into their workflow consistently report that their AI-assisted output passes informal 'sounds human' tests almost immediately.

Related research