If you are following the amazing demos people are making with Midjourney, Stable Diffusion and ChatGPT, you might think the latest wave of AI is all about fun new modes of communication. It’s not. Under the hood, it’s all about work. 

AI is computation; best guesses based on statistics. It is work; joules of energy dissipated through the movement of information. Despite the flashy new veneer, AI is not a revolution in communication but in productivity. It’s not the printing press or telegraph, it’s the assembly line, the jet engine, technologies that produce work rather than transfer information. And for knowledge workers, the latest developments in AI represent a new paradigm for work. 

Although computers are a form of information technology, the ways we use them are still quite mechanical. Much of our work remains trapped in the physics of W = Fs, forcibly pounding our keyboards, displacing our attention across emails, spreadsheets and websites, foraging for the bits of information we need to compose some work product. 

Contemporary AI promises something new altogether: a higher level approach where we work in deep collaboration with software. Computer programming moved from imperative to declarative languages and office work is just catching up. This is the new age of declarative work, made possible by the massive amount of information compressed into today’s text, image and code models. The ease of prompt interfaces and the ability to quickly generate variations will result in higher quality work products per unit of human effort.

The proliferation of AI into everyday office software (finally!) has the potential to create a new kind of hybrid workforce. We may soon enter an era when all of the apps you and every other knowledge worker use every day—your document editor, spreadsheet and presentation maker—will do part of your work for you. The more mundane aspects will be truly automatable—but the real breakthrough will be the teamwork between you and your software. You will have new, virtual teammates to write a first draft, instantly provide research and answer the most arcane questions. 

Will you feel like you have new, trusted digital colleagues, or a bunch of untrained interns to worry about?

This mix of humans and intelligent agents will increase individual productivity and enable leaner teams to do more, but will also be a lot to manage. Not only are generative work outputs subject to problems running from mediocrity to outright hallucination, enthusiastic overuse could make work and the internet itself even noisier. Will you feel like you have new, trusted digital colleagues, or a bunch of untrained interns to worry about?

In the physical world, humans have shared the factory floor with robots for decades. Workers, though, haven’t been programming their own automations, so these robots have been more like furniture than coworkers. This current wave of automation will change that. Companies like the German robot maker Robco now offer easily configurable industrial robot arms designed to be auxiliary workers for small manufacturing teams.

The great monolithic office software suites from Microsoft and Google have also felt like furniture. These are applications we work on rather than with. As AI changes this, how will human workers adapt? And equally important from our perspective, where are the opportunities for disruption in this new way of working?

Making it work

The coming AI revolution will look nothing like the internet explosion of the past 25 years. The internet was a revolution in how humans communicate. AI will be a revolution in how we work. Gaining distribution will be more difficult, and require companies to build passionate communities and unique customer value without becoming overly dependent on an incumbent’s expensive or restrictive platform. 

In the creative new world of generative AI, sharing code and models is the new engine of distribution. What Github is for code, Hugging Face is for models. Developers can use it to easily patch together models and datasets into bespoke pipelines. And Replicate has an easy-to-use API that lets users run open source models in the cloud. Unlike the Imagenet pioneers in 2012, you no longer have to have a PhD to deploy an ML model. 

The elephants in the Zoom are the office suites themselves. Microsoft and Google have their productivity apps installed on nearly every computer on the planet, and many companies use both. The incumbents have deep moats, but also blindspots. The once mighty IBM discounted the threat of the personal computer (and Microsoft) to its peril, and history has a way of repeating itself. 

Each of the pillars of the office suite is a potential vector of disruption. But could the office suite itself, based on the paradigm of human readable documents and messages, be disrupted? Today’s large language models (LLMs) are trained on public information and produce general purpose results. Will tomorrow’s leading companies preserve their advantage by building their own models and custom applications the way we now make spreadsheets or write memos? 

The first beachheads for challengers will likely be vertical apps, but the individual pieces of the office suite make attractive targets given Microsoft 365’s 345M paid users and Google Workspace’s 3B (mostly free) users. The office suite is just one of many software surfaces that startups can attack, but it’s a significant one. Could a 10x better doc editor, spreadsheet or slide maker powered by LLMs make a dent?

Ultimately, the result of all this innovation in AI will lead to something very different than office software on a PC. It will feel like an extension of ourselves; part assistant, part machine. But we won’t get there without a leap of imagination akin to the PC itself. In the meantime, there’s work to do.

The document editor

One of the main forms of work we do on our computers is writing, but the document editor itself has always been more about the formatting of words than the words themselves. Market success for these apps has also been driven more by their modes of distribution than product innovations.

WordPerfect was the early leader but ceded the field to Microsoft by failing to make a Windows version early on. Each new release of Windows was an occasion for the next version of Word, sold separately on a disc.

Google changed little about Word when it released Docs on the web in 2006, but it had a distribution advantage…it was on the web. In 2010 Google acquired the technology to let multiple people edit Microsoft Office documents on the web simultaneously. Online distribution and collaboration completely changed the way teams worked together. All Microsoft could do was follow suit with what became Office 365.

There have been many challenger doc editors over the years, Hemingway, Ulysses, Dropbox Paper, Salesforce Quip and most prominently today, Notion. Many of these apps aimed to simplify the writing experience, but Notion bypassed the old paradigm all together. There are no “pages” or skeuomorphisms, just documents fit for the internet age, designed for a screen and built to connect. Unsurprisingly, Notion has also been quick to incorporate generative AI into its product in a seamless way.

Coming at the problem sideways, Canva has continued to expand its offerings beyond social media and marketing graphics and recently released a document editing application that includes a generative “magic write” feature. There may be an upcoming generation of social-first entrepreneurs for whom Canva represents their office suite.

The big emerging opportunities for doc editors in the age of generative text is to innovate on the writing experience itself. As AI researcher Katy Gero recently wrote in Wired, AI can intervene in three distinct parts of the writing process, planning, drafting and revising. Current generative systems focus on the first two. Most applications will draft sentences and paragraphs for you as a completion of your prompt. More sophisticated approaches might return an outline for a blog post based on a headline. 

The big emerging opportunities for doc editors in the age of generative text is to innovate on the writing experience itself.

Just as the writing process is iterative, generative writing assistants also work best as recursive loops that the writer uses to converge on their desired outcome. It is not hard to imagine generative agents trained on constructive criticism that can offer feedback from different perspectives as you revise your piece. Eventually we may come to take this virtual writers’ room for granted as we do spell check today.

The first breakout text generation apps, like Jasper and Copy.ai, specialize in marketing copy. Their early success may be an artifact of the preponderance of marketing language in the training data for LLMs themselves. Nonetheless Jasper is a good example of a challenger playing the SaaS marketing game well. Following the HubSpot playbook it has cultivated a very enthusiastic community and created an engaging product. By relying on GPT-3 for its model infrastructure, it has been able to focus its efforts on the hard work of building its own distribution channel.

The biggest market for AI enhanced writing apps may turn out to be people who find writing difficult with current tools. This will take a more holistic approach than the first crop of vertical writing tools and the imagination to remake the writing process. Lex is an early stage generative writing product that pairs a fully featured doc editor with both generative prompts and an adjacent chatbot that can gather facts and answer questions to supplement the writer’s efforts. AI affords new ways of working which means novel and delightful product thinking can really make a difference.

Future applications will go much further: they will guide brands and individuals towards creating qualitatively exceptional content.

Applications can also combine generative feedback with traditional analytics and ML pattern matching. Jasper, for instance, provides SEO metrics on generated text. Future applications will go much further: they will guide brands and individuals towards creating qualitatively exceptional content.

Exceptionalism is one of the most promising areas for startups because it falls outside the scope of the large office suites. Brands and eventually individuals will want to fine-tune underlying models to capture their own unique voice. Right now, generative text sounds plausible yet generic. In Jasper’s premium Boss Mode you can add “tone of voice” keywords to your prompt (e.g. “Elon Musk”) much as you can add style suggestions (e.g. “digital art”) in image generation apps. But organizations and individuals want their writing to sound differentiated. We value a high-quality writer for their voice, not just their word count. There will be demand for easy ways to take your proprietary corpus of text and fine-tune a model on it.

Finally, looping back around to the planning process, generative models trained on code, like OpenAI’s Codex, have demonstrated emergent abilities for “chain of thought” and complex reasoning. It is possible that the doc editor of the future will be able to poke holes in your argument and not just give you a series of bullets based on what millions of other people have already written.

The spreadsheet

Visicalc, the original spreadsheet for the Apple II, was the first “killer app,” a program so valuable to business and academic users that it gave them a pressing reason to buy the whole computer. Microsoft actually launched the first version of Excel for the Mac in 1985 before following on for Windows two years later. 

Even more than the document editor, spreadsheets flourished in the graphical user interface. Where Word over-indexed on formatting, Excel delivered increasingly sophisticated functionality to the point that Excel formulas are now the world’s most popular programming language. For geeks, Microsoft’s introduction of LAMBDA functions in 2021 makes Excel Turing complete.

Google Sheets, launched in 2006, has kept rough feature parity with Excel, but as with Docs its main differentiator has been online distribution which eased collaboration. Excel appeals to power users with macros, larger data capacity and more visualization options than Sheets. Both spreadsheets have ventured into automated data analysis and formula suggestions in recent years potentially blocking new AI-powered challengers. And Google’s new Simple ML add-on for Google Sheets puts TensorFlow within the grasp of no-coders.

Bringing computational abilities to words and images is a new thing, but spreadsheets are computational to their core.

There is however no true Notion-like competitor in the spreadsheet space for the hard core spreadsheet user. If you’re embarrassed to touch a mouse when in a spreadsheet (guilty as charged!), Airtable or Rows won’t do it for you. But automating quantitative work in spreadsheets should be a slam dunk application for this new era. Most likely this will not come primarily through generative approaches but from more rigorous formal methods. Math is still math.

GPT3 turns out to integrate quite easily into existing spreadsheets via extensions that add =GPT() functions. It’s also possible to use language models to transform text prompts into working spreadsheet formulas. And perhaps a prompt interface could streamline tedious data cleaning and validating? Another application we would love to see would be the ability to specify synthetic data to fill out spreadsheets for modeling purposes.

The biggest moat for incumbent spreadsheets is inertia and user familiarity. Yet it’s easy to forget that spreadsheets themselves already represent a tremendous amount of automation. Bringing computational abilities to words and images is a new thing, but spreadsheets are computational to their core. The biggest disruption may be the merger of spreadsheets into document environments (see Notion and Coda) to create code notebooks for office workers. 

The presentation maker

Powerpoint is an island of creativity in office life and unsurprisingly presentations are the only part of the office suite that Apple has had traction. Slideshows seem primed for their generative AI moment, but decks themselves are unlikely to overturn the reign of the big office suites. But if presentations peel away from the suite—if Powerpoint is no longer good enough—the result could weaken the bundle.

Powerpoint and Slides have become design software for non-designers, but challengers like Canva are now applying the social media toolset to decks. Figma has made inroads with designers for making slides and offers FigJam for non-designers. Notion, a suite unto itself, has become a favorite of startup founders for their pitch decks.

One thing the main challengers in the presentations space—Figma and Notion—have in common is their emphasis on community as a way of building out functionality through templates and plugins. Figma’s plugin ecosystem provided a fast ramp to incorporate AI features like prompt-driven prototypes and generative icons without any changes to the core product. These communities are not technically open source, but they share the enthusiasm and good will associated with it. As we’ve seen with Hugging Face, open source represents an important distribution medium for AI technologies as do energized user communities. Any challenger who wants to enter these markets will have to build a productive fan base.

As we’ve seen with Hugging Face, open source represents an important distribution medium for AI technologies as do energized user communities.

Figma has had a big effect on the way design software has encroached on the bastion of the office suite in many companies. As products like Retool and Github Co-pilot make it 10x easier to build and code internal tools, a much wider swath of office workers will become engaged in making products which inevitably need to be wireframed and designed. We will address the AI disruptions to the worlds of design and code in a forthcoming post.

As with generative text, companies and individual designers will want to put their stamp on the visuals, so there may be a multi-modal platform play for a product that can ingest brand assets and spit out fine-tuned models for use across generative apps.

A real presentation standout at the moment is Tome, which is generative AI native, not a bolt on. It is currently in beta but has the ability to generate entire slide decks from a prompt, make variations of individual slides or add new ones and rewrite lines of text. This is the kind of magical work experience that people expect when they see Midjourney or ChatGPT and it may be that slide decks will be its first natural home.

A new kind of work

What does an enduring AI-native company look like? Will it follow the playbook of the PC era and be the next version of the word processor or spreadsheet? Will it follow the AWS model and provide metered infrastructure for personalized AI workloads? Even better, will it be the next Microsoft, Google or Apple, integrating human and AI workers into new operating systems and hardware?

Obviously, big tech companies with massive investments in AI are not going to let their incumbent distribution advantage slip away easily. Yet their business models and scale place them in an innovator’s dilemma that impedes their embrace of the new products and services that could undermine them.

Disruption is inevitable. The world of highly-distributed AI applications capable of producing human-level work products is very different from the winner-take-all mass distribution of the internet era. Ironically, it was this extreme centralization that pushed the internet corpus to the critical mass that made the scale of large language models possible. 

The world of highly-distributed AI applications capable of producing human-level work products is very different from the winner-take-all mass distribution of the internet era.

There are clear dangers, though, in accepting machine outputs as “good enough.” For individual workers there is the risk of producing large amounts of mediocre work. Knowledge work, whether marketing or scientific research, is valuable to the extent that it is particular and excellent. Not all knowledge work needs to be “true” in a factual sense, but it all needs to be “correct” for its intended purpose. The raw ability of AI systems to produce content places an even larger premium on our judgment and critical thinking skills.

For organizations, the profusion of generative outputs will create obvious risks of inappropriate content but also less obvious ones like cultural drift. If they rely on third-party models supplied through applications they risk losing control of their own message. Top brands will want to fine-tune their models for their voice and style while also hiring and retaining workers that can tell the difference. After all, successful businesses are different from their competitors.

If you’re building a new AI-powered product we’re excited to see what you’re hard at work on. Consider applying to Arc, our seed stage catalyst, or email me, konstantine@sequoiacap.com.