The Decline (Death) of SaaS?

OpenAI’s release of ChatGPT in November 2022 was the harbinger of a massive shift in building software and technology companies for knowledge work. Agentic AI is now widely recognized as the next major technology wave following the birth of the internet and the emergence of cloud computing/ mobile nearly 25 years ago with the launch of Amazon AWS in 2002.

Agentic AI has the potential for a “scorched earth” effect not just on antiquated technologies such as on-prem or mainframe-based solutions but even the SaaS solutions that have emerged over the last 25 years. Rather than continuing the SaaS approach of digitization and augmenting human labor, Agentic AI is bringing the promise of “just getting the work done.” This is a particularly powerful concept in service-heavy industries such as the realms of industrials, supply chain, and mobility that we take interest in. The irony is that many of the investments VCs are making today and in the foreseeable future could seed the potential downfall of legacy investments from the last two decades.

While we’re over two years removed from the pivotal release of OpenAI’s ChatGPT, the words of Buffalo Springfield still ring true amongst founders and VCs alike: “There's something happening here, but what it is ain't exactly clear”. And so, for the last 18 months, we’ve been observing, discussing, and thinking about the implications of the Agentic AI wave as we invest and help build the next generation of technology companies in our sector.

A Brief Review of SaaS

The evolution of enterprise SaaS is best understood by review two key eras:

• The On-Prem Era (1980s-1990s). Companies like Oracle and SAP dominated with expensive, heavily customized installations requiring extensive up front professional services. This was the era where we saw enterprises begin their digitization efforts.
• The SaaS Revolution (2000s-). The launch of AWS in 2002 allowed companies such as Salesforce, Workday, and their contemporaries to standardize software delivery by leveraging cloud computing. Through the use of subscription pricing, they further made enterprise software more accessible and affordable.

The Global SaaS market size was valued at $273B in 2023 and is projected to grow to $1.2T by 2032, exhibiting a CAGR of 18.4% during the period 2024-2032. North America accounts for approximately 50% of the global market value at $131B in 2023.

The reasons why SaaS has been successful include:

• Recurring Revenue. SaaS is based on a subscription model that provides predictable and reliable revenue streams that overtime translate to strong free cash flow generation.
• Profitability. With high gross margins (70-80%+), SaaS scales more efficiently than alternative business models. Once a product has been built, the major costs are attributable to sales and marketing with little cost elsewhere, enabling strong operating leverage as the business grows.
• Sticky. SaaS products integrate deeply into customer operations, leading to high switching costs and strong retention. Many companies also use "land and expand" strategies to grow revenue with existing customers.

The Poor Uptake of SaaS in Supply Chain

Looking through the lens of an industrials, supply chain, and mobility investor, SaaS has several drawbacks that explains why it has never been a good bedfellow in these industries:

• Exceptions. A major reason why SaaS doesn’t work well in the supply chain is the large number of edge cases and exceptions - making standardization impractical and requiring constant human intervention. 
• Legacy Processes/Systems. There is a significant requirement for deep, real-time integrations with existing systems - such as legacy warehouse and transportation management systems that must be constantly synchronized with legacy infrastructure. This explains the reluctance of the industry to move away from existing systems.
• High Fixed Costs. Supply chain stakeholders are exposed to a duality of changing volumes and pricing when considering their revenue forecasts. The fixed cost of the subscription model stresses cash flow because businesses must maintain and pay for SaaS products (think peak season licenses/seats) even when they might not be generating commensurate revenues during off-seasons.

Understanding Agentic AI

So, what is Agentic AI? It is a type of advanced artificial intelligence system that can autonomously make decisions and take actions to achieve a specific set of goals. The key distinction is autonomy – AI Agents are designed to be independent and proactive rather than purely reactive.

We frequently refer to Agentic AI as digital labor because it can complete complex tasks independently and without any human intervention.

It is important to recognize that SaaS is reliant on humans to manage the input of information, its workflow and reacting to its outcomes. SaaS is the picks and shovels that help humans get the work done.

Agentic AI will turn this all on its head, replacing repetitive and mundane tasks, making the need for SaaS redundant. These AI agents (aka digital labor) will not require SaaS to provide structured workflows to ensure that the work is done correctly and consistently.

Moreso, AI agents become even more valuable when they are focused on a specific vertical. These vertical AI agents have significant competitive advantages over standard models:

1. Specialization. The vertical AI agent can be tailored to handle domain-specific tasks, language, rules, and nuances, improving its effectiveness and reducing the number of errors.
2. Data Specificity. Verticals also provide consistency and specialized datasets, making it easier for the vertical AI agent to be fine tuned and adapted for particular use cases. 
3. Tailored Workflows & Goals. AI agents are naturally goal-oriented and a vertical focus provides a much more clear outcome objective. Over time, training a vertical AI agent with a client’s data accelerates improved outcomes and also makes it more sticky.

Finally, vertical AI agents are better at exception management because they are naturally designed to handle dynamic, context-specific scenarios and they can adapt their responses based on learned patterns or programmed logic. We are also seeing early evidence that vertical AI agents are able to demonstrate results sooner, supporting a faster time-to-ROI, and therefore a potentially faster growth pattern. During our 2023 AGM, we shared that verticalized solutions achieved Product/Market fit ~2x faster vs horizontal players. We expect this to be true in the agentic world.

SaaS typically operates on predefined workflows and rules that make them less flexible in unexpected or edge-case situations - which is more common than not when shipping goods from Monterrey to Montreal or fabricating that precision component inside of an electrolyzer.

The Market Opportunity: Digital Labor > Software

The potential for digital labor is a massive market relative to that previously seen with SaaS. Consider that for the Fortune 500, labor costs are around 50-60% of revenues (or >$5T of annualized costs ), whereas their average SaaS expenditure is only 1-3%.

We see the market opportunity for Agentic AI unfolding in three ways:

1. Displacement of the existing $131B US SaaS market. As is described above, there are many reasons why SaaS should be concerned about being overtaken by vertical AI agents. Furthermore, AI requires meaningfully different architecture to SaaS and so transforming into an AI business will be very challenging. The SaaS paradigm that humans control the input and output through forms, views, and integrations is not relevant in an agentic world. 
2. Eroding the future SaaS market. The Global SaaS market size is projected to grow to $1.2T by 2032. We think that vertical AI agents will quickly eat into this opportunity because it “just does the work” rather than being a tool for a human to leverage. If existing SaaS vendors are displaced, we expect the erosion of future SaaS markets to be inevitable.
3. The potential of digital labor. The limitations to SaaS (as described above) means it has not always been a good fit for every industry. Vertical AI agents can address use cases that have never been within the scope of SaaS. The realization of the potential for digital labor will only accelerate the decline of SaaS solutions.

The New Business Models of AI

The advent of vertical AI agents has also spurred a rethink of business models to better align the value created for a customer and that captured by a startup. While we are still early, there are several different business models being tested both in our portfolio as well as amongst startups we encounter.

“Charging for the work” or outcome based pricing is being tested by various different startups based on measurable outcomes, for example Raft (Fund I) charges by load rather than documents processed. For “cost per load” and similar business models it is important to ensure that the fee for work done covers not just the costs of standard processing but also exception handling.

An alternative model is not to sell the software but go “full stack” by building a business with the vertical AI agents embedded into it. This approach means the full value captured by the technology is realized at the bottom line. For example, Importal (Fund III) has taken this approach to customs brokerage driving a gross margin 60% which is substantially greater than the industry average of 20%.

Similar to the above, we are also seeing a lot of discussion around the potential for “roll ups” where long tail, service-heavy businesses with limited SaaS uptake are being considered as the next frontier for vertical AI agents. This allows for the automation of labor intensive legacy industries such as accounting, legal, or brokerage/agency models. Logrock (Fund II) has adopted this approach for its insurance super-agency ambitions.

It is worth noting, that the “full stack” and “roll up” approaches described above represent a massive opportunity for private equity (provided they have access to the right technologists) to acquire businesses and radically downsize its labor expenditure by implementing AI agents outside of the public gaze. Such radical labor downsizing may be more challenging for publicly traded companies.

Buyer Momentum Behind Agentic AI

In speaking with decision makers in our corporate network, portfolio companies, and VC counterparts, anecdotal evidence suggests that buyers are willing to adopt AI faster than existing SaaS solutions. It is common for us to hear from a corporate buyer statements such as “our board has told us that we should be spending 50% of our time looking at how to use AI.” This is supported by Gartner who estimates that by 2028 over 30% of enterprise software purchases will include AI agents.

We believe that part of the momentum is attributable to two “hard” factors:

• AI agents are generally simpler to roll-out than SaaS because it requires limited change in habits and training. We note that implementation can still be meaningful in instances where the solution sits inside the workflow and might need read/write access to key systems.
• ROI is easier to prove because of the digital labor paradigm that is now being unlocked. Many business leaders are questioning why they have a $75k per year analyst to reconcile invoices, a $1M per year trade compliance group, or the role of a $100k per year junior software engineer when an AI agent could do a substantial amount of the work for a fraction of the cost and also be more consistent in its quality of work output.

AI is Sucking All the Air (and $$$) Out of the Room

For the last 15 months, conversations in VC circles have generally been centered around funding AI versus non-AI businesses. Looking at Q4 data from Pitchbook/NVCA and speaking to our network of Series A investors, it’s clear that AI is sucking all the air (and also funding) out of the room.

As can be seen below, the value and number of deals during 2024 remains at a level comparable to 2019. However, there does appear to be a surge in Q4 2024, caused by a material increase in outsized deals pushing up the total capital deployed (which is explained further below).

VC Deal Activity by Quarter

Source: Pitchbook/NVCA.

As you dig further into the data, it becomes apparent that this bump is from an acceleration of AI deals that are disproportionately larger than the average. While AI represents almost 30% of all the deals being done, they account for almost 50% of the capital deployed.

AI & ML VC Deal Activity As a Share of All Deal Activity

Source: Pitchbook/NVCA.

When reviewing the data related to SaaS there are several takeaways. While the total number of SaaS deals has dropped to that comparable in 2017/18, the total capital deployed has increased to a non-ZIRP high. This suggests there has been a concentration in fewer but larger later stage rounds - those SaaS startups that have achieved “break-out” status.

SaaS VC Deal Activity

Source: Pitchbook/NVCA.

This assertion is further substantiated by the following graph that shows the proportion of later stage deals has increased relative to pre-seed/seed rounds. To double down on this, there are both fewer SaaS deals being done and fewer pre-seed/seed rounds being funded as a proportion.

It might be purely coincidentally (or not) that this decline started in 2022, the year ChatGPT was launched.

Share of SaaS VC Deal Count by Stage

Source: Pitchbook/NVCA.

It appears investors are doubling down on later stage SaaS startups as they consolidate their market position, whereas there is a decline in early-stage SaaS startups being founded/funded. This might be a function of

1. There may be simply fewer early-stage SaaS startups being founded by talented entrepreneurs. There is a finite number of ambitious founders, who appear to be more excited about the potential of AI rather than the opportunities in SaaS.
2. Similarly, early-stage investors may be opting to deploy more capital into AI and less into SaaS. This might be exacerbated by the decline in capital being raised by early stage investors - with the bulk of LP funds being diverted to the larger VC franchises. Numerous investors in our network have stated that they are unsure of the impact AI will have on SaaS (and the inability of SaaS to attract follow-on capital), further exacerbating the precarious position of SaaS.

The Death of SaaS?

We don’t mean to come off as trend chasing, or if you’re SaaS-inclined, to sound like it’s doomsday, but there are multiple reasons to believe that SaaS will go into long term decline. In summary:

• AI agents may simply outcompete SaaS in the market and be a better fit for the market opportunity - focusing on digital labor substitution where ROI is greater and faster to realize.
• In a battle for entrepreneur mindshare, fewer SaaS businesses are being founded by ambitious entrepreneurs.
• VCs (inadvertently) choking off the funding supply to SaaS in favor of AI businesses, thereby accelerating their decline.

Does the decline of SaaS become a self-fulfilling prophecy? We strongly believe the answer is YES.

Defensibility & The Importance of Data

As the competition around foundational LLMs remains fierce, we believe that there has never been a better time to build an AI-first startup. We have seen across our portfolio how substitutable these LLMs can be - with many startups running LLMs in parallel for redundancy purposes/maximizing uptime and also to allow them to cross reference their outcomes to help mitigate the risks associated with hallucinations.

Competition between LLMs will also drive greater functionality and reduce the effort required to build on the application layer. This will increase the importance of specialization and access to proprietary datasets to ensure competitiveness and defensibility.

One of the major risks associated with AI agents is its ability to be displaced by a comparable AI (similar to that in LLMs). For example, there have been a huge number of transcription services that can easily be substituted for an alternative. These “point solutions” are excessively exposed to displacement and price compression in a “race to the bottom.”

When considering an investment, we believe there are a number of ways to build long term defensibility:

• End-to-end vs Point Solutions. Providing a full service enabled by a series of AI agents that interact with each other to “just get the job done” are compelling and sought after. For example, Freight Hero (Fund III) provides all the load management workflows inside of a freight brokerage from scheduling, pickup/drop off confirmation, track/trace, driver ETA checks, and exception handling. In the world of VC. This embedded approach is harder to displace and captures all the data generated from “success” AND “failure” cases alike - thereby creating a data moat that becomes increasingly more valuable.
• Vertical proprietary data. As was previously highlighted, the use of historic and non-public datasets not only allows for the fine tuning of vertical AI agents, but also creates an unfair advantage (or data flywheel) over potential competitors over time. This delivers immediate value to new customers as they are onboarded from these fine tuned models and realizing a benefit much faster than the customer before them. Note, AI that is reliant only on public data will always struggle to differentiate itself and be subject to price compression. 
• Elimination of integrations. The underlying architecture of AI agents is a departure from the age-old integration philosophy of Extract, Transform, and Load (“ETL”). The concept of integrations are less relevant as AI agents are effectively interconnected between themselves. The greater the number of AI agents, the greater the stickiness of the product (similar to “End to End Solution” described above).

This is clearly an emerging sector that brings with it never-before-seen opportunities. For us in the industrials, supply chain, and mobility sectors, we feel like this is the productivity unlock that’s been sought for decades. That said, we remain vigilant to how this new technology and approach to startup building will impact both our investment approach and the startups in which we invest.