The rapidly changing reality for applied AI businesses

A cautionary tale about disruption, and why the future is bright

The last 18 months of starting, running, and ultimately shuttering Vectari, an applied AI startup, gave me a front-row seat to the most significant changes B2B SaaS businesses have experienced in the last two decades. 

How did you get disrupted by AI?
Gradually, then suddenly.

-A modern, LLM version of Hemingway

Early-stage founders and VC's talk a lot about disruption. Sometimes it's real, but in many more cases it's just overhyped Silicon Valley shorthand for competition. We saw both happening around us at Vectari. Ultimately, though, we were the ones getting truly disrupted. 

Our story is equal parts a cautionary tale about the risks AI poses to entrepreneurs looking to build something new, and an exhortation of how technology helps businesses succeed. There are three key learnings other founders can take away from our experience:

1. The competitive landscape in applied AI, LLM-adjacent businesses can shift quickly

2. Making hard decisions quickly is a superpower

3. What makes a B2B SaaS product a good venture scale business is rapidly changing

I. Solving a challenging problem using applied AI

The premise of our product rested on two ideas: First, that we could use LLMs - with lots of expert-provided fine-tuning and other techniques - to identify hard-to-spot trends in customer interactions at financial institutions. Second, those trends could be better managed as part of a purpose-built workflow that hid the complexity of the AI from the compliance and customer experience teams that needed to react to those insights.

For the first year of our work, our internal testing and development efforts showed that the commercially available LLMs were powerful but lacked the off-the-shelf ability to generate the insights our customers expected. With our compliance experts doing additional labeling and our engineers fine-tuning and using other NLP techniques, we could significantly improve how the LLMs performed.

The good news for us was that the product worked. 

Suddenly - starting early in the third quarter of 2024 - the latest batch of hosted, proprietary models were released with a bunch of improvements. Much longer context windows, more robust training corpora which seemed to include much more compliance writing, and increasing capacity to use the models were everywhere.

This meant good news for (almost) everyone: the LLMs got much better, much faster than many people expected. Better yet, the ease of LLM adoption made it easy for anyone to benefit from this rapidly improving technology. 

For us, it was ominous. First, it shrunk one key area of our differentiation because our ability to use a fantastic network of experts for labeling and fine-tuning models was no longer very notable. Second, it meant that we were about to face competition not just from scrappy, nimble startups but also from incumbent vendors who already had relationships with our prospective customers. 

Those incumbents historically had lumbering product development capabilities, allowing the scrappy startups to run circles around them because they innovate much faster. The ease of adopting the rapidly improving LLMs meant that the traditionally lethargic incumbents could integrate the technology quickly and deliver "AI-powered" results that were, at least superficially, similar to what our product delivered.

While these large, well-entrenched competitors embedded off-the-shelf models into their tools, we took solace in our product performing better. We were able to generate better insights into potential compliance risks. And, we were able to do it inside of a tool built specifically for that purpose rather than bolting it onto an existing customer support tool or consulting offering.

Unfortunately, those benefits weren't compelling enough for our prospective customers. The hard-to-quantify benefits of better insights and a well-designed UI couldn’t outweigh the simplicity of turning on one more feature from an existing vendor.

II. A difficult decision

Over the course of one fateful week in August, we saw the culmination of these trends: Two of our most advanced sales prospects fell through, selecting other incumbent vendors that were "good enough". Our fund-raising process, where we had made decent progress starting diligence with a couple of funds, stalled: while they were excited about what we were building, they wanted to hear more about the demand from prospects. 

Worse yet, we knew that more incumbent vendors would soon release additional competitive tools. If we had any chance at succeeding, it would require pivoting to a different idea.

That weekend, and the week that followed it, sucked.

Over the weekend, my co-founders and I wrestled with the undeniable reality of our situation. We couldn’t see a credible path forward. We didn’t have enough time. We needed to build something different and generate traction in a matter of months. We’d run out of money too soon. And we knew from our fundraising experience up to that point that raising more capital would only get harder.

Deciding how - or if - to move forward can be more complex when you have multiple founders with different perspectives. We were fortunate in that we all agreed that pulling the plug quickly was the right decision. How each of us came to that conclusion, though, was shaped by our individual experiences. 

I was reminded of a different situation earlier in my entrepreneurial journey. In that case, I delayed making a tough decision to shut down an unprofitable area of our business so we could try to find a better solution. I was wrong. 

I swore I wouldn’t make that mistake again. My earlier experience taught me that without a very high degree of certainty that delaying a decision allows the situation to change for the better, a rapid decision that preserves cash is the best answer.

Making that decision may not be 100% right, but it’s never 100% wrong.

While the more recent decision to wind down Vectari was undoubtedly the right one, that didn’t make it easy. Employees, investors and customers were surprised by what felt like a rapid, irreversible decision. They were right, but that didn’t mean we were wrong. 

The reality is that the hardest decisions leave what-ifs hanging in the air. Only founders know what it’s like to sit in the seat, making decisions and knowing that, no matter what, you will always be second-guessed. Some folks will always have lingering “what-ifs” in these situations. 

Courageous decisions rarely generate unanimous support. 

III. The changing calculus for applied AI startups

This experience has led me to believe there is a new calculus for starting applied AI companies going forward.

Since the advent of public cloud technology - basically, since 2006 - the core of a SaaS business could “simply” be doing something better. The introduction of public cloud is an important demarcation point because before public cloud, software companies had to raise money for both building the product and buying, building and managing the infrastructure. The availability of fractional, on-demand, managed infrastructure allowed SaaS businesses to focus on making better products.

While rarely simple in practice, “better” usually meant having some unique insight about the problem, focusing on a more narrow niche than others, or executing better. The act of building the product was so challenging that it, in and of itself, could be the differentiator.

Now, many SaaS businesses are becoming de facto applied AI companies. The business model is still SaaS, but products are quickly morphing into ways of applying AI to a specific problem.

Building these types of products is still hard, but the bar is lowering. In part, AI itself has made it easier, with tools like Github Copilot, Cursor, Replit and others continuously advancing how products can be built. 

We saw the effects of this first-hand. Our own developers initially had limited success with GitHub Copilot. Over time, tools like Copilot and other chat-based models like Claude began working as a “second-line” for developers, helping when they had questions about code.

By July of 2024, AI was the “first-line”: developers would ask a tool like Claude to write a feature, fix a bug, or refactor existing code. If the AI didn’t do it properly, the developer would do it by hand.

It was a complete evolution in software development practices in less than a year.

Other tools used in the creation of products have also gotten better. And there are simply more people with more talent than ever before doing this sort of work.

The availability of these AI tools levels the playing field greatly. The proverbial one-person developer in a garage can access the same powerful tools as a multi-thousand person software development team at a big tech company.

Suddenly, companies applying AI to business problems are finding it harder to compete on product, and simultaneously finding that every imaginable competitor has access to the same AI models they do. 

In this new world, SaaS businesses that are applying AI, or are somehow LLM-adjacent, will ultimately succeed because of better data, better distribution, or both. Of course, better data and better distribution always helped. But lots of SaaS businesses were able to thrive - despite having mediocre data and distribution to start - because of a great product. I believe that will happen far less frequently going forward.

Great products certainly help, but the advantage held by simply being great will become more and more elusive as the barriers to building new products continue to get lower. 

Instead, durable advantages will be created through proprietary data that can be used to generate unique insights (e.g., because the data isn’t generally available) or value (e.g., because the volume of the data is predictive). 

Proprietary data in this context could be first-party, meaning it is data that is sourced by the same company that is using it. It could be third-party, meaning that it comes from another source. For third-party data to convey a durable advantage, I believe it has to be proprietary, meaning other companies would have significant difficulty getting the same data on the same terms. 

Durable advantages can also be created by unique distribution arrangements. This could take the form of being able to “piggyback” on an existing product that a company already has, or through partnership with another company that brings a new product to market through an existing sales channel and set of customer relationships. Again, like third-party data, this is only a proprietary advantage if the distribution arrangement is uniquely available to the company leveraging it.

The key takeaway is this: SaaS companies have to assess if they are really an applied AI company with a subscription business model and, if so, prioritize creating durable distribution or data advantages very early.

IV. This is all, actually, good news

Well, okay, not all of it. Vectari’s demise is an example of the worst conclusion that a founder ever imagines. While the learnings from a failure can be useful, it pales in comparison to the thrill of winning.

But in a much larger sense, this disruptive evolution of how great companies and products are built is good for everyone. 

Software, and specifically SaaS, was an unusual breakthrough because of its economics - there are few businesses that have near zero marginal cost for scaling. Another inflection point that further helped software as a business model was public cloud computing. While we take these dramatic changes for granted now, they fueled a fundamental improvement in how value could be created through digital products.

Now, we find ourselves on the precipice of another dramatic period of change.

The widespread availability of high-quality and ever-improving AI models represents the next wave of real - not Silicon Valley buzzword - disruption. I genuinely believe that individuals, businesses, and society will be better off because of these technological advancements.

Along the way, though, a lot of what we took for granted is going to change. Entrepreneurs aren’t immune to this.

It’s going to be a wild decade, and I’m excited about what the future holds.