What it actually takes to bring AI to PowerPoint

A lot has been written about what AI can do. Less has been written about where it currently falls short, and why the gap between what users need and what the technology offers them matters more than the hype suggests.

PowerPoint slides are one of those instances where the technology doesn’t match the needs of the power user. The people who rely most on PowerPoint aren't casual users looking for quick and easy solutions. They're walking into rooms where the quality of their slides directly affects the decisions that get made.

For them, good enough isn't good enough.

The problem with using generative AI to create best-in-class presentations is that it’s genuinely harder than it looks and the bar for what "solved" actually means is much higher than most people writing about this space seem to realize.

In this blog article, I want to look at the technical hurdles that make slides a hard problem for AI, what today's most demanding PowerPoint users actually need, and where I think the real opportunity lies.

Technical hurdles

Slides are multi-modal

Large Language Models are trained on text and excel at processing it. But a slide isn't just text. A slide is a multi-modal artifact with text, images, and data arranged in a two-dimensional space where both the individual components and their layout carry meaning.

The position of a callout box relative to a chart matters. The size of a label in relation to the data it describes matters. Whether the eye moves naturally from the slide title to chart matters.

LLMs haven't yet developed a reliable, consistent feel for this spatial and visual reasoning. They can process text descriptions of a layout and generate images from text prompts, but understanding how visual elements work together remains a significant challenge.

Probabilistic vs. deterministic

LLMs are probabilistic. They predict the most likely output based on patterns in their training data. So, results can be inconsistent, and outputs are often mostly right most of the time, but not always exactly right every time.

Sure, for some tasks, mostly right is good enough. But not for professional slides.

Slide work is deterministic, so it demands consistency and repeatability. A first draft should be developed enough that a user is refining slides, not rebuilding them.

A single formatting inconsistency is an edit. Inconsistencies that keep reappearing are more work for the user.

When you're presenting to a board or a senior client, the outcome depends on the clarity, accuracy, and credibility of your slides. If an LLM can’t maintain consistent standards across a presentation, then it isn’t ready for that room.

Lack of training data

Human knowledge exists as text, which LLMs can access to train on. For images, there's also an enormous pool of training data.

There is nothing comparable for slides. High-quality slide data simply isn't available online. Nobody is systematically labelling good slides from bad ones and feeding that into a training model. Text descriptions of what a good slide should look like won't cut it either because text doesn’t provide an LLM the same level of input as learning from the slides themselves.

Unless this changes, AI outputs for slide creation that rely on the current breadth and quality of training data will keep missing the mark in ways that are immediately obvious to anyone who works with slides professionally.

Demands of today’s slide users

Accuracy and reliability

The top tier of PowerPoint users, such as management consultants, investment bankers, and strategy teams, can’t afford unreliable outputs. You can't walk into a board meeting with the wrong numbers. It might cost you your career, and it will definitely cost you your career if you blame AI.

Whatever AI does in this space needs to be accurate and reliable every single time, not just most of the time.

Readability

Professional slides exist to convey a message, share information, and support decision-making, often in high-stakes meetings where time is short and attention is limited.

They need to meet high visual standards, not just contain correct content.

This sounds so simple, but in practice, it's where current AI tools fall short. Power users work in situations where the quality of a slide directly affects the decisions made in the room. The bar is high, and the current generation of AI tools doesn't clear it.

Right now, even the best AI-generated slides tend to fall short of what an experienced consultant would consider acceptable. And it will take more than a ‘pls fix’ prompt to clean up.

Productivity

An AI solution for PowerPoint must fit into users’ existing workflows, not sit alongside them. Copy-and-pasting from ChatGPT into PowerPoint and back again is a workaround, not a solution.

This is about more than saving time. It means reducing cognitive overhead from switching between tools, maintaining a professional level of accuracy and consistency, and freeing headspace for the strategic thinking that no AI can do for you

A solution that adds steps, allows errors, or breaks concentration isn’t productive, no matter how impressive the underlying technology is.

The opportunity

Despite the challenges, generative AI undoubtedly has potential in supporting the presentation workflow. It's effective at generating insights, building out storyboards, drafting and refining text, and helping users think through the structure of an argument. These are real capabilities that can make a real impact.

The opportunity is there for whoever can combine those strengths with what AI still struggles with: accuracy, visual precision, and deep integration into professional tools.

This will require a rare combination: serious AI capability and a deep, established understanding of what the most demanding PowerPoint users need.

Conclusion

AI will change how we work with PowerPoint. That's not a prediction worth debating. The more interesting question I come back to with is whether the tools being built will meet the standards of the people who value slide quality the most: top tier PowerPoint users.

For them, a slide deck is a professional statement as much as a deliverable. And as AI makes it faster and easier for anyone to put a presentation together, it’s the attention that will help power users continue to stand out. But they won’t stand a chance without similar gains in productivity.

At think-cell, we've spent 25 years focused on quality and productivity for users who refuse to compromise on either. What we're building next is grounded in that same foundation. And I think it will be worth the wait.

Alexander von Fritsch is the CEO of think-cell.

Explore this topic further with AI: