At a high level, our experiments underscored that you can’t outsource deep thinking. And to get any benefits from gen AI in content creation, the relevant expert—whether writer, editor, or designer—needs to be involved to filter results and apply learnings. When done right, gen AI can not only educate, automate, accelerate, and elevate content and processes but also support your efforts to deliver distinctive, compelling insights that maintain your credibility and authority.

Figure: Gen AI use cases for content development can educate, automate, accelerate, and elevate.  

Example use cases:

1. Improve awkward sentences or phrases

2. Identify a data narrative

// Help translate raw numbers or break down complex charts.
// Sample prompt: Review the attached dataset/chart and write a one-paragraph narrative highlighting the most important takeaways.

3. Analyze content structure

// Test and validate the logical flow of an argument and the strength of transitions throughout a given article or report, as well as note unnecessary repetition.
// Sample prompt: Attached is a draft of an article for a professional-services firm on X topic. The goal of the article is Y. Critique the text’s flow and logic.
// Sample follow-up prompt: Suggest an alternative structure.

Example use cases:

1. Conduct market and campaign research

// Understand the current conversation on a given topic and potentially identify white space that you can own—or test the distinctiveness of your ideas.

2. Play devil’s advocate

//  Test the strength of an argument and identify potential areas for deeper research or concession in early development.
// Sample prompt: I’m writing an article on X for a professional-services firm. Below is the article’s thesis. Generate five counterarguments or potential challenges to the approach, citing common evidence for each.
// While some of the potential counterarguments may not be insightful, in our experience this exercise always generated one or two that were worth considering.

3. Support video concepting

4. Outline a campaign strategy

5. Check designs against compliance standards for accessibility and readability

// Provide a second opinion on how design elements flow together. Is the hierarchy intact? Are there any disruptions?
// Check that designs are accessible, including reviewing color contrast ratios, font size and spacing, and alt text quality.

6. Validate data visualization

// Sample prompts: What kind of chart would best show a change in composition over time? or I have data showing X. What are three chart types that could effectively visualize this, and what are the pros and cons of each?

Example use cases:

1. Draft emails to help experts communicate more effectively with nonexperts (specifically non-designers)

2. Support fact checking

// Create a list of facts in a document (for fact check) and company mentions (for reputation/risk review).
// Check for factual inconsistencies within a document (does the statistic given for enrollment rates on page 5 match what is given on page 25?) or repetition (did we already cite this stat?).
// Confirm math in exhibits.

3. Support copy editing

// Scan documents for errors and consistency of terms, acronyms, and capitalization. (If a company has its own style guide, models may need to be trained on it.)
// Search multiple style guides simultaneously.
// Format citations.
// Alphabetize names in lists of acknowledgments, etc.

4. Create reusable templates for content development timelines and checklists

5. Process images and videos

// Generate video captions.
// Resize images and videos.
// Edit images.

6. Build technical design frameworks

// Create responsive systems and automate repetitive tasks. For example, prompt an LLM to write After Effects code for an adaptive animation that can seamlessly scale and sequence multimedia assets.

Many people are tempted to use LLMs to help them write drafts, fill in outlines, and generate images, but our experiments confirmed that these tools fall woefully short when it comes to wholesale content creation. Four failures stood out.

1. Rewriting technical language: When we asked LLMs to rewrite highly technical language for a general business audience, they had a tendency to make things up and pull in additional sources from the web—even though we specifically asked them not to use outside sources. So, while the exercise may help explain or break down complicated language for a writer who is not a specialist in a particular field, the actual output isn’t usable.

2. Drafting from outline: In a drafting experiment, we fed various LLMs an early version of an outline for a published article. We provided context on the article’s audience, purpose, and ultimate publication destination. In every instance, we were disappointed (but also relieved?) by how little the LLMs expanded on the content provided; for the most part, they just reformatted what they had been fed. And the writing. Oh my. As one colleague put it: “This is the most boring thing I’ve ever read.” Across experiments, LLMs tended to stuff content with jargon and meaningless mush, particularly when asked to rewrite prose that was already jargon heavy. When we asked the LLMs to analyze their own drafts to identify gaps and weaknesses, they did. But they were still unable to fix the draft by addressing the issues. Writing is thinking, and there are no shortcuts for that.

3. Discerning nuance in tone: Our team asked LLMs to create two new versions of a short paragraph: one in a diplomatic and cautious voice, and one that was energetic and confident. For both versions, the LLMs spat out results that were exaggerated and cartoonish. They also padded the paragraph with useless fluff and phrases that are out of place in thought leadership—for example, “R&D pipelines are packed like never before.” Even when prompted to tweak the voice to fit that of a subject matter expert, the models struggled to soften their exaggerated rewrite.

4. Creating brand-compliant imagery from scratch: “Firefly is testing my patience,” a senior designer wrote one morning in our design group chat. She was using Firefly, an Adobe design tool, to create imagery for an AI-hesitant client along well-defined brand guidelines. Her prompts specified content, mood, style, even the exact hex codes. But no luck. The design team took turns pitching prompts, but the AI engine couldn’t balance the overall brand guidelines with the design objectives of the project.

Thought leadership is particularly challenging for LLMs to generate, because the whole point of this type of content is to offer original insights, not synthesized knowledge. Using gen AI to make better content more efficiently is thus a premium service, not a shortcut—and it requires expertise from a team of professionals. Only expert content writers, editors, and developers can comb suggestions for white space and potential topics and quickly determine which are original and worth pursuing. Only graphic designers can weigh a client’s brand guidelines against real usability needs. Only production editors can review citations to ensure they’re accurate and consistent with house style.

Writing is thinking, and there are no shortcuts for that.

As our long list of use cases indicates, there are plenty of ways gen AI can add value during content development. But there’s a clear ceiling on what it can do. The use cases we’re excited about are quite narrow, and much of the added value comes from explaining terms or concepts or poking holes in existing materials. As for actually generating new content? Not so much—at least not yet. You won’t hear us complaining about that.