Rudy Lai

Automating knowledge work has long been a goal of the tech industry.

In 2020, I launched a startup called Tactic to tackle this exact problem. The vision was to generate documents by automating the finding and summarising of information. When I say documents, it can eventually include spreadsheets, presentations, and more.

We couldn’t really build it back then based on technology constraints. Every year since Tactic, I have seen new startups popping up to tackle this problem.

You'll remember this company called Tome.ai, where they gathered a huge amount of funding, traction, and users—1 million users in 134 days, 10 million in 10 months, and 20 million in 18 months, to be precise. But despite this traction, Tome announced it would shut down Tome Slides by the end of April 2025 and pivot to the sales space with their tagline "Make deals, not decks."

It's clear that there's a big need for this, but because the technology is still a bit early, the timing is very tricky. When I am writing this in October 2025, Microsoft just launched “Vibe Working”, an “Office Agent”, powered by Anthropic. Gamma AI’s presentation maker is the new hypergrowth darling. I am very keen to see if these will succeed.

Decks for reading vs decks for presenting

Now, I think it's quite interesting to dissect why automating the making of slides is quite a challenging problem. To do that, let's first break down what it takes to build great presentations.

It helps to split presentations into two categories:

Read-first decks (send-and-read). These are decks you distribute — investor updates, executive summaries, reports you email to people to read on their own.
Present-first decks (talk-and-show). These are decks designed to be accompanied by a live presentation. Visual cues, timing, animations and the presenter’s voice do a lot of the work.

When a knowledge worker builds a deck (say, a marketing strategy for next year), the workflow typically falls into three stages: finding information**,** analyzing it**,** and presenting it. Each step has its own hard problems. All three of these steps present unique challenges that I think current generation AI cannot handle.

Challenge #1: Finding offline information

When it comes to finding information, a lot of valuable information simply isn’t sitting in a single API or a neat database. Some of it is in different internal systems; some of it lives only in people's heads; some is analog (whiteboard photos, slides handed out at a meeting) or fragmented across emails, spreadsheets and image scans. Gathering the right material, in the right format, from the right people, and turning it into something you can feed into an automated system is a heavy human task.

This is not a small engineering problem where you just call an API — it often requires social work: knowing who to ask, how to get access, how to interpret partially structured data, how to trust a source. That layer of human effort is a big reason automated slide-makers struggle to produce meaningful, high-value decks without human input.

Challenge #2: Emotional and social context

But then, the second step, analysing data, is an even harder challenge, an even more often overlooked challenge.

Collecting data is necessary but not sufficient. Turning that data into insight — the actual storytelling — is often the most human part of the job. Analysts look for trends, call out anomalies, weigh emotional and social context, and judge what will resonate with decision-makers. These judgments aren’t purely mathematical; they’re interpretive.

Pulling together heterogeneous data sources and synthesizing them into a single narrative requires domain knowledge, context about incentives and audiences, and subtle judgments about what to highlight and why. Current generative models can help draft narratives and surface patterns, but replacing a skilled analyst is its own huge startup problem — and it’s not the same as “generate slides” work

Challenge #3: A diagram is worth a thousand words

This is the stage where automation looks most promising: once you have clean data and a clear insight, turning it into charts, a slide layout, and copy is something many modern tools can help accelerate. Good visualization converts raw numbers into a story and — when done right — makes the insight instantly graspable.

Still, there are important nuances: which chart form most clearly communicates the point? Should a rise be shown as one stacked bar or multiple bars? What animation, if any, will help a live audience follow a complex idea? And then there are the tiny but critical things that make slides sing: concise, punchy slide headings (the “one-line takeaway”), clean contrast, and layout choices that guide the eye. Models tend to be verbose and can miss these micro-optimizations that a human with an eye for presentation will iterate toward.

Automating decks: a three miracle problem

Elad Gil has a very eloquent observation about startups: You know a startup is likely to fail when its product strategy or business plan includes multiple miracles.

Each of the three areas of building a great deck — collection, analysis, and visual storytelling — is a miracle. For example, you can probably build an entire company by automating data analysis. That would disrupt Looker, Tableau, and the BI space in general. An automated deck-maker only becomes useful when you ‘solve all three miracles’ at a useful level of fidelity.

That’s why you see some tools that do the early draft well (helpful for brainstorming) but don’t produce decks you’d send to an investor or to the board without meaningful human edits.

You can get a lot of value from current tools for quick ideation or broad summaries. I know people who use these systems daily for first drafts. But in many professional contexts, the output is still “too generic” — it lacks the depth and crispness you get from deliberate research, rigorous analysis, and careful visual design.

Why present-first decks are even harder

When you’re designing slides that will accompany a live presentation, visual storytelling becomes paramount. You have to compress complex ideas into a sequence of visuals that support spoken narrative. You may rely on animations, pacing, or diagrams that only make sense in a live context. Automating those choices — and ensuring the visuals reliably support a human presenter — is a level up in difficulty from producing a send-and-read deck.