LLMs will unlock the next wave of innovation in B2B SaaS

I’m not usually prone to hyperbole, but having spent time with a bunch of startups in the most recent Y Combinator batch, I feel like I’ve had a chance to glimpse the future. As a firm, we diligently meet over 200 YC startups a year which gives us a good sense of the Zeitgeist (embedded fintech —> web3 —> AI). But while flavors of the month typically fade, I have come to believe that LLMs will transform a significant swath of business processes and that we are likely entering a golden age of B2B SaaS investing.

The conclusion above was informed by meeting with various LLM startups, so I want to start by giving some examples. From there, I try and draw some general conclusions about how LLMs can demonstrate value in the enterprise. Finally, I lay out some implications for founders building LLM startups that I see from my vantage point as an investor (still evolving!).

Examples of startups using LLMs to transform business processes

What struck me was how many startups across disparate industries had rapidly come up with 10x better solutions (which is a general benchmark that investors think about when investing in new companies). As the examples below show, many companies used LLMs to save businesses time, headcount, money, automate menial tasks and generate incremental revenue.

Linc (AI copilot for freight brokers) -- freight brokers typically receive 1000+ emails per day which need to be manually entered into ERP systems. Linc’s tool automatically extracts relevant data from emails and syncs it with the ERP, saving brokers several hours per day of data uploads
Casehopper (AI-assisted legal drafting) -- there are 17 million visa applications submitted each year, with the average application requiring 5 hours of an immigration lawyer (or paralegal’s) time. Casehopper reduces this drafting time down to 5 minutes, effectively replacing the paralegal in this process. For a small immigration law firm, this can result in a substantial increase in cases processed and revenue generated
Revv (AI-powered vehicle sensor calibration) (*) -- all modern vehicles have sensors on them which need to be recalibrated when the car needs to be repaired. Sensor calibration requires technicians to review complex 100-page manuals (which differ for every make, model and year) to figure out the correct procedures to undertake. Revv’s software turns this multi-hour research task into seconds, allowing shops to capture substantial revenue without expending any effort
AskLio (AI-copilot for procurement teams) -- internal procurement teams typically have policies and guidelines around software and equipment that their company can buy. By training an LLM on this internal corpus, they can significantly reduce the back-and-forth that happens between internal buyers and procurement teams, resulting in faster approvals and less bottlenecks
Sweetspot (AI-powered search engine for government contracts)(*) -- companies that bid on government contracts do so through an antiquated and dysfunctional online portal called SAM.gov. Sweetspot uses LLMs to help businesses find relevant contracts in seconds, rather than spending hours digging through the government’s complicated online database

A screenshot of Revv’s product that shows what sensor calibration operations a body shop needs to perform. Sourcing this information manually would likely have taken 5+ hours of manual work, which technicians don’t have the time to do.

Generalizations about where LLMs are most valuable

To synthesize across some of the examples above, there are a few core skills that are well-suited to LLMs:

Automating manual, repetitive work that can be clearly defined (e.g. Linc) — doesn’t have to be limited to text-based tasks
Massively simplifying complex, text-based research and summarization tasks (e.g. Revv)
Giving employees creative superpowers
In some cases, replacing employees entirely (e.g. Casehopper)

Sarah Tavel recently wrote a great blog post about how LLMs have the potential to not just augment, but replace, humans for certain tasks

Some implications (for founders and investors)

LLMs can power products that are so clearly better than the existing way of doing things, and can demonstrate rapid time-to-value and clear ROI in doing so. Some implications of this below:

1️⃣

Industries or verticals that were previously hard to penetrate (e.g. due to difficult GTM motions) are now in play. This is because founders are able to build wedge products so compelling (10x better) that they pique customer interest and give them an opening into the market

2️⃣

Relatedly, many previously unsexy industries (e.g. legal tech, ed tech, gov tech) are now fertile grounds to build a startup. When choosing an industry to attack with LLMs, the relevant factors to consider are the suitability of its business processes to automation and the competitive dynamics (see 3)

3️⃣

Although there are many categories where incumbents have been quick to adopt LLMs, prospective founders can offset this by thoughtfully choosing a vertical with favorable competitive dynamics. Specifically: a) choosing a category with slow moving incumbents that are less likely to rapidly adopt LLM functionality (e.g. gov tech), or b) finding an overlooked segment within an industry that incumbents are unlikely to focus on, or c) choosing a category with no software incumbents (e.g. parts of certain service industries like law). As the barriers to adopt LLMs are low, getting air cover (from both other startups and incumbents) is critical. This gives founders a better chance to build an early lead in their sector, which may be critical from a fundraising and industry credibility perspective

4️⃣

Focus on building products where the ROI can be easily tied to increased revenue (e.g. Revv) or decreased cost or headcount (e.g. Casehopper). In contrast, focusing on a use case which broadly makes a worker more efficient may struggle to obtain the necessary budget and staying power required to build a sustainable business

5️⃣

Given concerns around defensibility, the usual lessons about B2B software are even more important. Specifically, founders need to think about: a) developing proprietary distribution strategies b) building switching costs into their product c) incorporating proprietary data into their models d) adding workflow functionality which increases stickiness Said another way, being a wrapper around an LLM may be enough to get initial traction, but won’t result in long-term value. It’s imperative for your tool to have additional functionality outside of the LLM that’s core to solving the customer’s problem

6️⃣

Early traction for an LLM startup can be misleading as buyers are currently overzealous in their willingness to try. Churn metrics may not yet be available, but will be critical to monitor

All in all, it’s as good a time to be a B2B founder as I’ve seen since I first became an investor in 2018. Founders should think creatively about their wedge product and vertical, but can be emboldened in knowing that there are many 10x better solutions across a wide variety of industries to be built using LLM technology.

If you are a founder building business software powered by LLMs and this post resonates, I’d love to hear from you. My email is samit@1984.vc.

* We at 1984.vc are invested in these companies