AI-assisted business writing
Where AI helps in a business writing workflow
A model can save you time at some stages of writing a business document and cost you time at others. We look at where a model helps and where it doesn’t, along with an order of work that keeps the important decisions with you.
If you have a report due on Friday and a model open in another window, you’ll want to know where in the process the model will save you time and where it will create extra work to undo later. The answer depends more on the task than on the type of document. Any single piece of work has some stages that reward speed and others that demand knowledge of the subject and the reader.
There’s evidence that the time savings can be real, but that they depend on the task. In one preregistered experiment, professionals who were given ChatGPT for writing tasks from their own occupations took 40% less time on average, and the quality of their work rose by 18% (Noy & Zhang, 2023). A later experiment, with 758 consultants at Boston Consulting Group, found that those given GPT-4 completed 12.2% more of a set of realistic consulting tasks, 25.1% faster on average, and with better results. On a task chosen to fall outside the model’s capabilities, though, they were 19% less likely to reach a correct solution than colleagues working without it (Dell’Acqua et al., 2026). Both sets of tasks were set by the researchers, but a real business document also needs a person to decide what it should argue and what it can leave out. Making those decisions takes knowledge that a model doesn’t have. Among the AI-assisted drafts we edit, the strongest usually had a model handling the tasks that reward speed and breadth, while a person decided what to say and why. The weakest are often the ones in which the model was also left to make those choices.
Stages where a model saves time
Starting from a blank page is often the slowest part of writing. A model can quickly produce a first structure for you to respond to. Asking it for an outline or a rough draft turns the open-ended job of composing into the more manageable job of editing. Read that first draft critically, and expect to keep little of its wording by the end.
A model can also summarize source material. If you have thirty pages of interview notes or a stack of background reports, it can give you a serviceable digest in seconds. Do this only with material you’ve read yourself because a summary can sound plausible whether or not it’s accurate. Only a reader who knows the source can tell the difference.
Reformatting is quicker too, provided the content is settled and only its form needs to change. A model can turn bullet points into paragraphs, a set of figures into a table, a long email into a brief, or a report section into slide notes. Here, nothing has to be invented, so there’s less room for error. Take care to read the result anyway, to check that nothing was omitted.
Before a colleague sees your draft, you can ask a model what a skeptical reader would object to. Many of its objections will be generic, but the list still gives you a useful starting point for revision.
Checking mechanical consistency is another task that suits a model: it can confirm whether a term is used the same way throughout, whether the headings follow one pattern, whether numbers and dates share one format, and whether every figure has a caption. This kind of detail-oriented work is tedious for a person and quick for a model. Review each change it suggests instead of accepting them all at once.
Where relying on a model costs you
Anything that has to be accurate is risky to delegate. A model can draft around facts that you give it, but it doesn’t know whether a figure is correct or whether a cited paper exists, and it might present an incorrect figure with the same confidence as a correct one. Invented sources are a documented problem. When researchers asked GPT-4o to write six literature reviews in 2025, about one in five of its citations were fabricated, and almost half of the real ones contained errors, such as a wrong DOI. Fabrication was more common on less familiar and more specialized topics (Linardon et al., 2025). Trace every factual claim in a model’s draft to a source you can point to, and give the references extra attention. As a standard part of our editing, we confirm that every source exists. Our guide to finding hallucinated sources in commercial research explains how to check them yourself.
Proposals and thought leadership pieces are read for what they tell the reader about you: whether you understand the problem, and what you think should be done about it. Writing with a model’s help can also make different people’s work more alike. In one experiment, short stories written with ideas from a model were rated as better written, but they were also more similar to one another than those written without that help (Doshi & Hauser, 2024). If a document needs to show your own thinking, decide the specifics and the position yourself and limit the model’s involvement to cleaning up mechanical issues rather than content.
The last pass is better left to a person. A model will happily polish a draft, but it can also flatten the whole document until every sentence sounds alike. For a document that matters, the final read should be done by someone who reads it as the audience will.
Take particular care with anything that you may have to account for. If a client or a regulator might query a document, you need to be able to say how it was made and vouch for every line. A model-drafted passage that you didn’t fully understand will be difficult to defend in that conversation. If the model you used watermarks its output, its involvement may also be detectable, whether or not you disclose it.
An order of work that uses both
- Decide what the document is for and who it’s for, and write that down before you open any tool.
- Ask the model for a structure and a rough draft against that brief.
- Read the draft against the brief, and rewrite the passages that make your case in your own words and with your own specifics.
- Let the model check the result for consistency and convert any material that belongs in another format, such as a list that would work better as a table.
- Have someone, ideally not the person who wrote it, read the whole document as the intended reader would before anyone else sees it.
In our experience, that order keeps the model’s work confined to the tasks it handles well and leaves the decisions with you. If you reverse it, the model writes the argument and a person tidies the grammar afterward. The reasoning in the finished document then comes from the model instead of from your organization.
What this means when we edit your draft
When an AI-assisted draft reaches us, the sentences are often perfectly sound. Much of our work is in the structure and the voice, and in flagging claims that you’ll need to check. If the author retained responsibility for every decision that needed real judgment, the draft may need only a light edit. Our guide to what actually changes when we edit an AI first draft describes the work involved when that didn’t happen.
References
- Dell’Acqua, F., McFowland, E., Mollick, E., Lifshitz, H., Kellogg, K. C., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2026). Navigating the jagged technological frontier: Field experimental evidence of the effects of artificial intelligence on knowledge worker productivity and quality. Organization Science, 37(2), 403–423. https://doi.org/10.1287/orsc.2025.21838
- Doshi, A. R., & Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28), Article eadn5290. https://doi.org/10.1126/sciadv.adn5290
- Linardon, J., Jarman, H. K., McClure, Z., Anderson, C., Liu, C., & Messer, M. (2025). Influence of topic familiarity and prompt specificity on citation fabrication in mental health research using large language models: Experimental study. JMIR Mental Health, 12, Article e80371. https://doi.org/10.2196/80371
- Noy, S., & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187–192. https://doi.org/10.1126/science.adh2586
Drafted with a model?
Send it over in whatever state it’s in, and we’ll agree the scope with you before quoting.
Get a quote