AI-assisted business writing
The patterns that make readers stop trusting a text
Some habits of phrasing and structure make a draft sound generic, whoever wrote it. We look at the common ones and how to edit each one out.
When someone reads your report or proposal, they’re forming a view of you as they go, as well as taking in what the document says. Some habits of phrasing and structure work against you in that judgment: they make the text sound as though whoever wrote it took the first words that came to hand. Language models fall into many of these habits readily, but so do people writing in a hurry. The advice below applies to any draft, however it was produced.
We don’t think that making a draft pass as human-written is a useful aim. In a series of experiments, people couldn’t reliably tell whether self-descriptions such as professional profiles had been written by AI. The cues they relied on, including contractions and first-person pronouns, proved to be poor guides (Jakesch et al., 2023). The reason to edit out the patterns below is simpler: each one makes a sentence say less than it could. We’ve grouped them by where they tend to appear and suggested a remedy for each.
Openings that delay the point
Many drafts open with a sentence that could begin almost any document, such as “In today’s rapidly evolving business landscape, organizations face unprecedented challenges.” Nothing in the sentence is specific to your reader or your subject. Delete it and open with the finding or the problem that prompted the document.
A related opening restates the title. A report called Improving Supplier Onboarding that begins “This report examines how supplier onboarding can be improved” squanders its first sentence on what the reader already knows. Another maps out the route ahead: “This document will first outline the background, then discuss the methodology, before presenting recommendations.” Readers can see your headings, so in both cases you can start with what you found. If the structure seems to need explaining, the headings may need revising.
When hedging leaves nothing asserted
A single qualifier is often the accurate choice. Stack several together, as in “This approach may potentially offer some benefits in certain contexts,” and the sentence stops claiming anything. Decide whether the approach offers benefits. If it does, state what they are and where they apply; otherwise, cut the sentence.
Phrases like “It is important to note that…” do the same damage while posing as emphasis. Readers tend to assume that everything in a document is there for a reason, so you can delete the frame and keep the substance that follows.
The passive voice can also obscure who’s responsible for a claim. “Mistakes were made” and “it was decided” leave the reader wondering who made the mistakes and who did the deciding. When the actor matters, name them. Readers may lose confidence in a writer who seems reluctant to say who did what.
Where the actor is obvious or unimportant, though, the passive is often the natural choice. Models may also use it less than people do: one study found that GPT-4o used the agentless passive (“the samples were tested”) at roughly half the human rate (Reinhart et al., 2025). A draft written almost entirely in the active voice can sound mechanical. A mix of the two reads more naturally.
Why some rhythms sound automatic
Lists of three are a familiar example of a rhythm that can start to sound automatic: “clear, concise, and compelling”, or “faster, cheaper, and more reliable”. Three items make a satisfying rhythm. When a draft relies on it repeatedly, though, readers start to hear the pattern more than the words. Count the items you actually have, which might be two or five.
Mirror sentences such as “It’s not about working harder; it’s about working smarter” promise a contrast, but the two halves often amount to the same point. If there’s a real distinction, state both sides plainly; otherwise, keep the half that’s true.
Paragraphs can fall into a pattern as well. A common one opens with a claim, gives two sentences of support, and then restates the claim at the end. The restatement can suggest that the writer didn’t trust the support to make the point on its own. Cut it, and end the paragraph on the evidence or its consequence.
Sentence length matters too. When every sentence in a passage is about the same length, the prose can sound monotonous even to a reader who couldn’t say why. Read a page aloud, and where you hear a run of sentences that are all the same length, split one or join two.
Emphasis without meaning
Intensifiers such as “truly”, “really”, “deeply”, and “incredibly” add nothing to the claim they decorate. “This is a truly significant finding” is weaker than “This is a significant finding”, and both are weaker than saying what the finding actually is.
Some words have been used so often in the same kinds of sentences that they now read as filler: “robust”, “seamless”, “leverage”, “holistic”, “navigate”, “landscape”, and “journey” are familiar examples. Language models have shifted how often some words appear: one study of more than 15 million biomedical abstracts found that certain style words became abruptly more frequent after large language models arrived (Kobak et al., 2025). None of the words on this list is wrong. Wherever one occurs in your draft, ask whether you can find a more precise word to describe what you mean. If you can, use it; if not, consider cutting the word altogether.
Signposts delay the content they announce. “There are three key reasons for this” and “Here’s what you need to know” tell the reader that something is coming when they were about to find out anyway. Cut the announcement and just give the reasons.
Upbeat closing lines are the equivalent at the end of a document. A line like “By embracing these strategies, organizations can unlock their full potential” could close almost any document, so it tells the reader nothing about yours. End on the specific next step, or on what happens if the reader doesn’t take it.
Who is behind this claim?
Phrases like “studies show”, “experts agree”, and “it is widely recognized that” claim support without demonstrating it. Where the studies exist, cite them. If you can’t support a claim, present it as your opinion instead.
An unsourced figure raises the same question. Faced with a sentence like “Organizations lose up to 30% of productivity to poor communication” (an invented example), a careful reader will ask where that number came from. In a document that will be checked, an untraceable figure is a liability. Trace each figure to its source or drop it. If you’re working from a model’s draft, treat every figure and citation as unverified until you’ve found the source yourself. Our guide to finding hallucinated sources in commercial research explains how to fact-check your model’s output.
Made-up examples have a similar effect. A sentence that begins “A company might, for instance, find that…” signals to the reader that the example was invented to fill a gap. Use a real case if you have one, with identifying details changed if necessary, or leave the example out.
Matching certainty to the evidence
A document that states every claim with the same confidence gives the reader no way to tell which findings are solid and which are provisional. Careful writers signal the difference so that readers can decide what to rely on and what to treat as provisional. Make it clear which claims you’d base the recommendation on and which still need testing.
Enthusiasm can stand in for evidence in the same way, as in “We are excited to present”, “a game-changing approach”, or “a powerful new framework”. Show what the framework does, and let the reader judge whether it’s powerful.
How to work through a draft
You don’t need to fix everything at once. Start by reading the first page aloud and marking every place where you hear one of these patterns, since that’s the page every reader sees. Then work through the remaining pages in one pass, first cutting and then rewriting.
Drafts that began with a model tend to contain more of these patterns. Our guide to why AI-drafted reports sound generic shows you how to address them, starting with putting the specifics back and committing to what you know to be true where the draft might hedge instead. If you’d like a second reader, our AI-draft refinement service covers this kind of editing.
References
- Jakesch, M., Hancock, J. T., & Naaman, M. (2023). Human heuristics for AI-generated language are flawed. Proceedings of the National Academy of Sciences, 120(11), Article e2208839120. https://doi.org/10.1073/pnas.2208839120
- Kobak, D., González-Márquez, R., Horvát, E.-Á., & Lause, J. (2025). Delving into LLM-assisted writing in biomedical publications through excess vocabulary. Science Advances, 11(27), Article eadt3813. https://doi.org/10.1126/sciadv.adt3813
- Reinhart, A., Markey, B., Laudenbach, M., Pantusen, K., Yurko, R., Weinberg, G., & Brown, D. W. (2025). Do LLMs write like humans? Variation in grammatical and rhetorical styles. Proceedings of the National Academy of Sciences, 122(8), Article e2422455122. https://doi.org/10.1073/pnas.2422455122
Send us the draft
Send it over in whatever state it’s in, and we’ll agree the scope with you before quoting.
Get a quote