The Realities of AI Investing
How Deciens navigates opportunities and avoids pitfalls in an ever-evolving AI ecosystem.
A decade ago, AI was a sci-fi subplot. Today, it’s the backbone of billion-dollar companies and the silent engine behind thousands of startups.
At Deciens, we’ve been tracking these shifts closely. Epistula #4: The New AI Landscape in VC, published in October 2023, explores the rise of generative AI and large language models (LLMs), along with the challenges and opportunities they bring to the table. Since then, the pace of advancement and excitement has only accelerated.
In this post, we’re taking that initial analysis further, diving deeper into how we think about the AI ecosystem, where we see the best investment opportunities, and how we approach the risks that come with such transformative change. Let’s get into it.
Understanding the AI Ecosystem: A Layer Cake
To understand the AI landscape, we like to think of it as a “layer cake,” with each layer representing a different part of the ecosystem, complete with its own opportunities and challenges.
1. The Base Layer: Hardware and Infrastructure
The foundational layer of AI is the hardware – think GPUs, data centers, and the cloud infrastructure that powers everything. These are the physical systems on which AI runs and forms the industry’s backbone.
Our take: While critical to the functioning of AI, this layer involves highly capital-intensive horizontal businesses and is therefore not investable for Deciens. From a macro perspective, these companies rely on a few hyperscalers as key customers, such as Microsoft and Meta. Any slowdown in their spending could have ripple effects, and investor scrutiny is already questioning ROI in this space.
2. The Middle Layer: LLMs and Foundational Models
Next, we have large language models (LLMs) and other foundational models. These AI systems are trained using the hardware infrastructure mentioned above. Companies like OpenAI are raising billions of dollars to develop these models and continuously iterate to stay ahead in the arms race.
Our take: Though transformative, these ventures require substantial scale and resources, placing them beyond the reach of most early-stage investors like us. Like the base layer, their revenue streams are tied to a small number of hyperscalers, adding vulnerability to market shifts. While undeniably groundbreaking, they are massive-scale ventures that are not typically investable for us at Deciens.
3. The Application Layer: AI-First and AI-Enabled Companies
This layer represents where AI technologies are deployed, and it can be split into two categories:
AI-First Companies
These businesses deploy generative AI to their end users as a core part of their value proposition – for example, AI-powered software solutions that are integral to the service or product being offered.
Our take: These businesses are certainly investable, and we’re seeing a surge of AI-first fintech companies entering the market and generating very significant revenues very quickly. However, many come with sky-high valuations, fueled by bubble-like funding. Combined with (i) accuracy challenges due to hallucinations – especially problematic in financial services, where precision and reliability are paramount, (ii) overpromised capabilities, as replacing human activity in many areas of financial services is still years away, and (iii) unclear moats, these factors lead us to approach the space with caution.
That said, we know that teams who overcome these challenges will have the potential to build transformative companies. We’re particularly interested in those that generate proprietary data as a byproduct of their operations, strengthening their AI use case over time.
AI-Enabled Companies
This will soon encompass nearly every company. These businesses utilize AI internally to build better products but do not necessarily offer AI directly to their end users. For example, an enterprise software company that uses AI to enhance operations, improve product features, or streamline customer service.
Our take: These companies tend to be relatively stable investments, using AI to improve existing business models rather than relying on the unproven potential of generative AI. There’s strong promise in applying AI to make traditional software products faster, cheaper, or more scalable. In today’s boom of AI-first applications – like OCR tools, chatbots, and coding co-pilots – companies have many options from which to choose. While generative AI isn’t fit for every situation, we will be looking in every investment for technologists who are thinking creatively at the cutting edge of how and where it can be applied to their business
4. AI-Adjacent Businesses: The Frosting on the Cake
These are the companies and startups that build enabling technologies or provide the necessary support structures for AI to scale across different industries. Think of this as the "frosting" on the AI layer cake – holding the layers together. These businesses are crucial for the ecosystem’s growth, as they help to enable the deployment of AI solutions on a broader scale. They include everything from data labeling and AI ethics companies to platforms that help integrate AI into existing business models.
Our take: We find these businesses both investable and exciting. They’re critical enablers for the industry and often involve specialized workflows or unique data sets – a sweet spot for creating defensible moats.
A New Paradigm of Growth
The revenue growth in companies – across multiple layers of the cake – has been dramatic over the past 6-12 months. We may be witnessing a new software adoption paradigm. But what’s driving this surge?
A key driver may be how seamlessly today’s AI tools fit into existing behaviors and systems – reshaping the way we search, the way we create, and the way we work. Let’s take a closer look at each of these dimensions.
From Search to Conclusion
Delivers Instant Answers: AI-driven chatbots provide immediate responses, eliminating the friction of parsing through Google Search results, synthesizing the data, and determining an answer. All else equal, users gravitate toward solutions that minimize time and energy to get an answer, and in that equation, “instant” wins.
Generates Context-Aware Responses: Chatbots can also retain and contextualize user information, enabling them to provide answers specific to the individual or organization making the query. The prior paradigm required users to include all relevant information in each search. In contrast, chatbots begin with more context, and can ingest new information without losing what was shared earlier. For instance, a user can input their medical history to get a personalized interpretation of a test result – and later, the same agent might flag a recipe with a health-related warning based on that context.
From Ideation to Creation
Bootstraps Creative Production: One of the most immediate benefits of generative AI is its ability to help users get started. Whether it’s content, code, or design, it delivers a first draft that, while often gets heavily edited, eliminates considerable time in the creative process spent figuring out where to begin.
Enables a Natural Language Interface: Generative AI is language-driven, enabling users to create or control content, code, and more through modern language. This serves as a bridge between lower-level tasks and human thought. Engaging through natural language is intuitively easier, which is why this interface is likely to dominate over other modes of creation.
From Workflow to Automation
Reduces Adoption Friction: The cost-benefit of utilizing new software is directly related to how quickly it can be implemented. With generative AI solutions, the adoption cost is nearly zero, and the benefits are rapidly improving.
One reason new software often faces an adoption curve is that it typically introduces a new workflow for users – especially in enterprise settings. While the new application and workflow may deliver better outcomes than the legacy systems, it requires education, training, and activation energy. Implementation can take weeks or months, and armies of consultants are regularly engaged for ongoing training.
On the other hand, generative AI-driven solutions automate existing workflows by handling steps a human would otherwise perform, and are language-driven, so there is no new workflow to learn. Adoption often involves removing now-automated steps rather than learning new ones. As a result, these tools can frequently be integrated instantly by plugging into existing processes.
Anticipates Context: Thanks to the intelligence inherent in generative AI solutions, they can now infer context that legacy software would have required users to provide, dramatically increasing the utility of a given solution. For example, where time-tracking software for lawyers and consultants once required users to manually input the client each time they switched tasks, generative AI can now automatically detect these context shifts in the background. Eliminating the need for user input unlocks a fundamentally more efficient experience.
Drives Step Function Increases in Productivity: When the two factors above are present, the resulting productivity gains are hard for end customers to ignore. As a result, companies are racing to trial AI software across their organizations – because if these solutions deliver on their promise, the impact on financial performance could be substantial.
Risks and Realities
While the momentum behind generative AI is undeniable, it’s important to acknowledge the risks that come with such rapid innovation. These are the areas where we think caution is warranted – and where hype can potentially outpace substance:
Operational Risks
The biggest risk to all of this is the potential for hallucinations. The promise of instant answers comes with the danger that those answers may be wrong – and today, there’s often no reliable way to assess the likelihood that is the case for any given query.
Human nature compounds the problem. Because people tend to prefer systems that deliver answers faster and with less effort, many won’t take the extra step to verify the output. This dynamic applies in B2B applications, as well. Although, depending on the work product, there are often human eyes on the output at some point in the process, along with additional business context that can be leveraged to assess the veracity of what was produced.
At the same time, this risk is not fundamentally different from trade-offs many companies already make. For example, traditional human support teams make mistakes – errors that could be reduced through better staffing, training, and systems. But most companies stop short of investing to achieve 100% accuracy because the cost is too high. The same premise could be applied to AI solutions: a more efficient system will still make mistakes. The real question is what level of accuracy companies are willing to accept in exchange for cost savings.
Financial Realities
The tremendous growth seen in some companies has reignited a bubble-like fervor in valuations – extending even to early-stage companies with little or no revenue, simply because of the potential for exponential growth. In addition, even for those companies with rapidly growing revenue, given the nascency of these products and business models, it remains unclear whether the value being captured today will be sustainable over time.
In addition, foundational model companies continue to expand the aperture of their product in a bid to stay ahead. With each product expansion, they subsume entire classes of startups that were trying to solve those same problems as stand-alone companies. In effect, what constitutes true whitespace is always in flux.
Finally, given the massive amount of capital trying to participate in the AI boom, investment discipline has seemingly become an afterthought. This dynamic creates the potential for many minefields across all stages of investment.
The Agents Are Coming
Agentic solutions have rapidly emerged as the current vehicle for deploying generative AI at the application layer. At their simplest, AI agents are autonomous bots designed to fulfill specific functions – for example, completing KYC processes for a financial institution. These agents can operate internally, on a company’s internal data and systems, or externally, in interfacing with systems beyond the company’s walls. As AI companies unlock autonomous engagement with the broader internet, external use cases are going to come online rapidly.
While this is a large topic with many implications, two use cases in financial services stand out – and more are sure to follow:
Agentic Payments: As agents operate externally, a likely use case is enabling them to make transactions on behalf of users or companies. For instance, a shopping agent tasked with finding a particular item could also complete the purchase. This could meaningfully accelerate the level of automation, raising important questions about whether it will reshape the existing payments ecosystem, redefine the roles of various players, and either enhance or erode the value that certain players provide today.
Financial Institution Interactions: It could be equally transformative for agents to interface with a person or company’s financial institutions – retrieving data, answering questions, or conducting transactions. But this, too, raises questions similar to those posed by agentic payments. A further consideration is how institutions choose to grant access, particularly in light of the recent rollback of Section 1033 of the Dodd-Frank Act (the “open banking” rule). The CFPB’s revised stance suggests that 1033 did not encompass requiring data access or portability to other applications or providers, as opposed to just the account owner. Whether agents qualify under that definition remains unclear but critical, and will likely be a source of significant regulatory debate.
Unresolved System Questions
As these agentic use cases evolve, they bring with them a set of complex questions across regulatory, logistical, and technological perspectives. Among the most critical:
How do financial credentials get managed and verified?
How can you differentiate a "good” bot from a “bad” one, especially in systems historically designed to screen out bots entirely?
Who bears liability when an agent initiates an improper or unauthorized transaction?
How do we ensure control and accountability in systems where chains of agents may act autonomously and in coordination?
These are not edge cases – they sit at the center of how agent-based systems will integrate into core financial infrastructure. It remains unclear who will be best positioned to provide these solutions: incumbent financial institutions, scaled fintech startups, or new entrants. This uncertainty could create both significant challenges and meaningful opportunities for the fintech ecosystem.
Our Approach to AI Investing
At Deciens, we believe successful AI investing requires a disciplined and focused approach. The dynamic nature of this space demands careful evaluation of companies, technologies, and market trends. With that in mind, we focus on:
Value Creation: Backing businesses with a clear impact, not speculative explorations. These businesses demonstrate not just potential but a measurable ability to create value for their customers and stakeholders.
Near-Term ROI: Prioritizing solutions that can deliver results in the near to medium term. This approach ensures that the companies we back can sustain their growth without relying solely on long-term, uncertain developments.
Top-Tier Teams: Seeking out and supporting founders and teams with proven expertise, visionary thinking, and the resilience to navigate the challenges of building in a fast-changing market. For us, the strength of the team is often the clearest indicator of a company’s ability to succeed.
Defensible Moats: Investing in companies with unique advantages, such as proprietary data or specialized workflows. These factors not only differentiate them from competitors but also make them more resilient to market shifts.
Transformational Impact: When being deployed directly to end users, situations where generative AI can not just replace existing workflows, but do something that simply wasn’t possible with the legacy solution.
AI-Adjacent Solutions: Betting on the enabling technologies and infrastructure – the “picks and shovels.” These businesses are well-positioned to succeed regardless of which specific applications or trends dominate, offering a more stable and enduring opportunity within the ecosystem.
Striking a Balance
Although much focus is on AI right now, it's crucial to remember that traditional machine learning (ML) hasn't become obsolete. In fact, its deterministic nature makes it better suited for certain applications, particularly in financial services where predictability is essential. We're increasingly seeing successful combinations of traditional ML and generative AI approaches, such as retrieval-augmented generation.
The AI revolution is transforming our portfolio companies, whether through traditional machine learning, generative AI, or hybrid approaches, and we expect more of it. As investors, our role is to carefully navigate this landscape, seeking opportunities that combine innovative technology with sustainable business models and clear paths to value creation.
As we move deeper into this era of technological transformation, it's clear that the most successful ventures will strike a balance between innovation and practicality, leveraging AI to solve meaningful problems while maintaining sustainable business models.
