Here’s a Million-dollar Software Idea

I have an idea for a new type of software company. It is radically different from anything that exists today, and if you read this article closely, you might be able to make millions of dollars. The pitch goes like this:

Software is the greatest business man has ever invented. It is God’s gift to capitalists. The margins? High. The technical difficulty? Surprisingly low. If you follow the strategy proven by hundreds of startups, you can scale a software business to tens of millions of dollars of revenue in less than five years. It’s simple, cash-efficient, and has been the bedrock of tech’s might for decades.

The entire foundation of this mountain of moolah rests on three simple things:

1. A back end that has a table of data,
2. Logic that performs a set of actions based on that data, and
3. A front end that lets a user interact with the results of those actions.

The three together create a virtuous cycle: The user inputs new data or takes some action with the data on the front end, which gets sent to the back end, starting it all over again. 

Annnnd that’s kinda it. 

The multi-trillion-dollar world of software applications is all based on those three things. The technology is so simple that the majority of value over the last two decades has come not from sci-fi innovations in technology, but from innovations in how we distribute it. But I think that is about to change with a new model of design I call ambient software. If I’m correct, the right founder could completely undercut existing SaaS giants. And that’s where the millions of sweet, sweet dollar bills come in.

What comes after Salesforce?

I categorize enterprise software into two general periods: BS and ASS (Before Salesforce and After Salesforce’s Success). 

During the BS era—from the 1960s through 2000—software was installed on premises by professional services teams and sold via license. In 2000, the ASS period began. Salesforce proved that you could deliver software over the internet and create new business models with it. You didn’t have to install the software onto someone’s servers and charge service fees. You could just charge a monthly subscription fee and update the software remotely, generating superior long-term economics. The ASS period, which continues to this day, has birthed many tech winners, especially those that have a good product and a novel distribution strategy—like Slack (which had product-led growth) and Snowflake (which implemented usage-based pricing). Software didn’t need technical innovation to win as much as it did unique distribution.

Accordingly, software adjusted to fit these requirements. Focusing on distribution meant that companies needed to specialize. There was software for salespeople (Salesforce), for surveys (Qualtrics), for restaurants (Toast)—that sort of thing.

If these companies want to grow, they can either expand horizontally (use the same tools to serve a new set of customers with similar problems) or vertically (create new tools that serve other problems its existing customers are facing). For example, a piece of software that helps you tip your barista could expand horizontally by helping other types of retailers implement tipping or vertically by selling other coffee shop software.

Once again, note that the majority of the value comes from little technical innovation and more of a focus on distribution. Now, remember my original pitch: ambient software, in which large language models automate everything but the database away.

A million-dollar idea

Large language models have two net-new technical capabilities: taking in large amounts of context, and using that context to inform a probabilistic generation of outputs. Basically, an LLM can read a bunch of shit and then make stuff on the basis of that shit.

The end result will be ambient software. Ambient software is different from what exists today because it simplifies what an end user needs to see, reducing the amount of time and the number of applications needed. Context windows—the amount of stuff that LLMs can ingest—are growing ever larger, and an LLM can record every single user interaction, drop it into a table, create new actions, and then display only what an end user would need.

That…was kind of abstract. Let me illustrate. An ASS customer relationship manager (CRM) like Salesforce works like this:

1. A back end holds data about a customer and how the sales team has interacted with them,
2. Logic automatically triggers marketing actions depending on the data that appears, and then
3. A front end displays what actions have been taken and allows employees to enter in new data from the sales calls they had with customers.

An ambient CRM would work like this: 

1. A back end holds data about a customer and recordings of every single interaction sales staff has had with them (calls, emails, etc.).
2. Then, an LLM pulls out relevant data points and creates logical work alerts for what matters to company stakeholders—for example, the customer may have mentioned a problem they’re trying to solve during a sales call.
3. Finally, a front end displays a dashboard, alert, or analysis targeted only to the person who needs to make a decision or take a new action—so the LLM could automatically write the follow-up email from the sales staff and attach custom-made marketing materials that show how the product they’re trying to sell solves that problem.

That last nuance around “only to the person who needs to make a decision” is very important because it allows for a total disruption of ASS company business models. A good indicator a company is building ambient software is that it reduces the number of end users. There isn’t a salesperson alive who will say, “I enjoy using my CRM.” Because an ambient CRM records and analyzes all the calls and interactions with a potential customer, a salesperson would get only the most crucial information they need about a customer, exactly when they need it. There is no “CRM” for them to interact with—there is just a helpful email that appears with the info that they need, when they need it.

By centralizing all sales data into a data lake—an industry term for a very large back end storing data—you can undercut everyone else on margin while making a product everyone prefers. Salespeople don’t have to enter data, management can perform better analysis, and customers have fewer hoops to jump through.

Everyone wins. Well, everyone but Salesforce. This would change a SaaS product from one for which you try and sell more and more seats into a database with a logic engine on top. You can accomplish everything a CRM does with dramatically less investment in training, UI, and end points. You can undercut ASS companies on price while improving the utility. While some software companies today have the data necessary to make this work, it will be extremely challenging for them to make the leap.

Why ambient software will work

This type of product would present a classic innovator's dilemma: The UI and workflows would be so different from that of existing software companies that the incumbents wouldn't be able to replicate it easily. Ambient software is a radical way of reimagining knowledge work. Unlike ASS software, where humans are involved in every step, all data entry, analysis, and manipulation will just be done for us. Humans get to focus on human-to-human connection and more creative tasks.

In practice this means that management gets database access with an LLM to help them query it. Everyone else just has an email inbox. An LLM will cover the details, all the way up to figuring out what sort of actions to provide a user with. 

This would make an ambient AI app dramatically cheaper, more powerful, and easier to run than any SaaS app of today.

But wait—I think there is even more money to be made (lucky you!). Over the last few months LLMs have gotten really good at producing mini-applications or mini-websites in response to user queries. In March 2023, software researcher Geoffrey Litt saw this coming when he wrote about “malleable software,” where LLMs temporarily make new, unique end points that users interact with. Once the user is gone, the end point disintegrates into digital dust. Ambient software could enjoy the revenue expansion of traditional software by increasing the complexity or volume of quick-use applications customers could build on top. Instead of charging for adding more users, you could charge for adding more mini-apps or workflows, like an always-referenced view of data.

The best evidence I have that this idea will work is that some of the best software startups today are already building versions of this into their applications.

• Notion AI: The note-taking app was well-positioned for this idea because it was already acting as a database for users. People dump everything onto the company’s platform. Early ambient software implementations at Notion are simple instances of “find this info for me” or “summarize this dataset.” So end users need to put in less work to dig through their search results, and Notion should eventually remove the many UX buttons they have for this. Additionally, the company has been acquiring email and calendar startups—you can see that they have a vision to hold all of a customer’s data. Millions of customers have tried Notion’s AI offering, according to its cofounder, and I’ve heard through the grapevine that it is a revenue driver. 
• Airtable Cobuilder: Airtable started out ideally positioned for ambient software because its product was a widely popular database and spreadsheet software. People were already accustomed to plugging their data into the app and organizing it. The company launched an AI app builder in July to allow users to build quick applications on top of the data they were already inputting.
• Clay: Perhaps the most compelling vision I’ve seen so far of ambient software is Clay. The company operates a table software that allows go-to-market teams to use AI to automate sales activities—like scraping websites and sending emails on the basis of that data. The product is growing like a weed and currently has 100,000 users at more than 2,500 companies. At the moment, it utilizes the data stored on a CRM, and many of its actions are reliant on third-party email services, meaning that it's not fully ambient computing. However, when I profiled the company in early July, it seemed fairly obvious to me that the CRM would eventually become relegated to just a database if Clay had its way.

To be fair, none of these are full-fledged ambient software companies—at least not yet. Instead, think of them as tasting notes on a plate full of startups. If you take each of their elements—Notion with all the data, Airtable building micro-apps, and Clay relegating the CRM to the background—and combine them, you get a mouthful of ambient software. They’re LLMs that interact freely with users, supported by a concrete foundation of data.

What software business should I build?

In ambient software, whoever controls the back end controls the AI. It works best when the startup has all of the data centralized in one location. This is challenging because most software incumbents have already realized the problems that AI represents and have started doing all they can to lock up data. Getting ambient software live with a customer would require a large migration process—and a number of cultural changes. Convincing employees to allow every phone call, keystroke, and conversation to be recorded will be challenging, to say the least.

Most glaringly, models aren’t yet good enough to do this at scale. The big bet of ambient software is that you are hoping the models will improve. That is far from a sure thing! LLM progress has felt relatively slow compared to the gains that came in the year between the GPT-3 and GPT-4 release.

I actually find the model problem encouraging. The differentiation doesn’t come from model strength—something most startups can’t compete on because it’s expensive—but from implementation and understanding customer problems, which is where startups excel. This means that there is an actual startup to be built! If you’re building a company like this, reach out—I would love to chat.

Evan Armstrong is the lead writer for Every, where he writes the Napkin Math column. You can follow him on X at @itsurboyevan and on LinkedIn, and Every on X at @every and on LinkedIn.

Was this newsletter forwarded to you? Sign up to get it in your inbox.