2024 Investment Outlook
Data → AI Assembly Line, AI-Native Productivity Apps, Leapfrogging the Adoption Curve in Industry AI
As 2023 came to a close, it felt like the startup / VC community took one collective sigh of relief. 2023 was a very challenging year to navigate, as the ground was shifting beneath us across multiple fronts simultaneously — a venture market downturn, a banking crisis, growing geopolitical conflicts, and of course, an emerging platform shift to AI that is grabbing all the attention (and hype!). Any one of these events in isolation would create dislocation, but the fact that they all converged at once has made it an especially dynamic environment for founders and investors alike.
I believe there will be more rough waters ahead in 2024, but I am optimistic for those who are weathering the storm. Those founders and investors who have been self-reflective, resilient, and continuous learners will not just survive, but be stronger moving forward. They have grown new muscles, deepened their conviction into what they’re building, and will benefit from a less competitive environment as others who have failed to adapt quickly enough will disappear.
Dislocations in markets also create opportunities, particularly for intrepid startups and investors that are paying close attention and move swiftly and aggressively to capitalize on the opportunities opening up. And that is what is most exciting to me in 2024 — partnering with founders that have unwavering conviction on building something that isn’t obvious to the crowd and has enormous upside if successful.
On the AI hype cycle…
You can hardly get out of bed before reading another article from a VC or tech pundit claiming something that resembles “AI is going to be bigger than the Internet.” It’s become the opposite of a contrarian view at this point. I believe we’re already reaching peak hype cycle in generative AI, and it’s only just over a year since the launch of ChatGPT!
I personally don’t view AI as its own theme or sector, but rather a technology wave that has been building for many years and is now primed to reach mass adoption due to its accessibility. In what used to be a field reserved for teams with top ML researchers and engineering talent hand cranking their own models, the barriers to entry have been dropped with the release of foundation models that provide access to highly intelligent AI models for all who want it via API. Despite being overheated in certain areas, it doesn’t change the fact that this current wave of AI is transformational, and its implications are wide reaching across the entire tech ecosystem.
Many investors have claimed that the value from generative AI / foundation models will only accrue to the incumbents that have large valuable datasets and established distribution advantages. It’s a compelling argument given it’s relatively easy to bolt on these models into existing applications, and there is plenty of investment coming from large tech incumbents building their own AI capabilities and product extensions. While incumbents will flex their data / distribution advantages to win in many cases, there are also benefits to being a startup at the start of a technology wave. Startups have a blank canvas, do not suffer from innovator’s dilemma, have no technical / architectural debt, and can iterate quickly as a small, nimble team. I refuse to believe that all of the most interesting opportunities that emerge from this rapidly evolving platform shift in technology will be met by large, notoriously slow-moving incumbents…
Which brings me to my top 3 investment themes for 2024… 1) At the infrastructure layer, there are hair-on-fire problems to solve for companies building, deploying, and maintaining performant, safe, and reliable AI applications at scale — the Data Assembly Line becomes the AI Assembly Line. 2) There is a budding generation of AI-native application startups that will break out by building brand new user experiences that lift productivity to heights never seen before in their markets. 3) AI combined with industry vertical-specific datasets and technologies in both the physical and digital world will help startups leapfrog the adoption curve in Industry AI.
Theme #1: Data Assembly Line → AI Assembly Line
With the emergence of AI foundation models and generative AI, there is a lot of energy from companies both large and small in figuring out how to bring AI capabilities into existing and new applications. I’ve covered what I call the Data Assembly Line for the past couple of years, which is about how you take raw data, refine it, structure it, secure it, and take action on it for analytics and operational workflows. This current wave of AI adoption is a forcing function to streamline the data assembly line and incorporate both proprietary and open-source AI foundation models into a self-reinforcing data flywheel. There is a platform shift underway that leads companies to build their own AI Assembly Line to establish self-reinforcing data pipelines for an effective AI product strategy. Given the recent surge in accessibility and excitement around building AI applications, there is also a new class of customers that have previously been on the sidelines in participating in building novel data / AI applications that will bring new sources of budget and market opportunity to those innovating in this area.
- Data Acquisition & Aggregation: At the core of any successful AI strategy is acquiring and aggregating unique and useful data to feed and refine models. Data is a modern company’s crown jewels, and building an effective data moat is becoming even more valuable than it already was when investing in AI to automate internal processes or create a better customer experience for digital products. Synthetic data generation is an interesting emerging field for startups, but there is often no replacement for authentic real-world datasets to serve as ground truth to train or fine tune existing AI models to customize and form fit AI applications for a business or use case. Startups that capture a unique and valuable dataset will become even more powerful moving forward in the AI assembly line.
- Data Unification and Quality Assurance: There is no quality AI without quality data. Most enterprises have massive quantities of data, but a majority of it still sits in data silos or is in raw / unstructured form, not finding its way into the company’s models, workflows, or products. Enterprises have continually struggled with getting data out of silos, unified, cleaned, structured, and ready for use. Cloud data lakes, data warehouses, and data lakehouses such as what is offered by Snowflake or Databricks help with serving as a place to dump data, but that data is typically replicated (i.e. expensive) and requires tools / engineering effort to pull the data back out to be useful. Startups such as Syncari (*Crosslink portfolio co) seek to solve these problems for business users to circumvent laborious IT / data engineering work to synch, unify, and improve the quality of business data across an organization. Data unification and quality is a critical area for investment moving forward.
- Data Governance & Security: AI safety and security is a big hot button item for any enterprise, particularly in regulated industries. Based on conversations with executives at several Fortune 500 companies, I know that some organizations currently have upwards of hundreds or even thousands of AI projects actively being worked on within a single company. However, most of these projects remain in the prototype phase before being launched into production due to concerns over what sensitive data is being fed into the models, what vulnerabilities are available for attackers to infiltrate them, or how they can defend the accuracy of the models. These are critical areas to get right for getting AI projects out of prototype and into production, and there will be a boon for companies filling this need of governing and securing AI models and applications.
- AI Modeling: Data scientists and developers are living in exciting times with the access available to world class AI models off the shelf. The emergence and rapid growth in performance of powerful foundation models from both proprietary sources such as OpenAI or Anthropic, and open-source models such as Llama 2 (out of Meta) is a key reason we’re seeing such a boom in AI development. However, these foundation models are generalized, and often don’t solve a target problem well enough by themselves. Organizations need to fine tune the models with proprietary, targeted datasets and/or combine large generalized models with smaller, focused models to accomplish tasks end-to-end. Tooling such as LangChain or its enterprise-grade alternative, Griptape (*Crosslink portfolio co), help developers at enterprises orchestrate models and connect them with defined datasets. There is an exciting opportunity forming for developer tools that aid in the connection of data to models and fine tuning of models to build and prepare sophisticated AI applications for production.
Theme #2: AI-Native Productivity Apps
At Crosslink, we’ve been investing in both enterprise and vertically applied AI/ML applications for the past decade. The field used to require a deeply technical team of ML research and engineering practitioners to ship AI/ML products, as you had to build and train your own models, which typically took years to have something ready for production. Since the launch of ChatGPT in November 2022, there has been an explosion of activity from both startups and tech incumbents building capabilities on offer from foundation models into both consumer and enterprise applications. With what is now on offer from both proprietary and open-source foundation models, any software company can relatively quickly and easily ship AI capabilities in their products, rapidly speeding up the pace of innovation and development of new use cases.
Many of the initial use cases and applications have been bolt-on AI capabilities such as chat, search, or text summarization into existing product experiences. I’m more excited about application startups that are starting from scratch and reimagining what a completely new product experience could look like with the capabilities that these large, powerful models bring. We’re only starting to scratch the surface of what is possible by linking together various models (open source, proprietary, large, and small models) and datasets to complete complex, sophisticated workflows. A new life changing product experience coupled with supreme GTM execution can propel startups to build their own data flywheels from user / human-in-the-loop interaction and produce moats for a new class of AI-native application winners.
- Natural Language Interfaces: One of the most exciting unlocks from large language models are the ability to work with unstructured data, or data sitting in documents, emails, web pages, or knowledge bases. Being able to take action on unstructured data without needing to gather, structure, or segment it before querying it unlocks new datasets previously untapped by software and new automations that were previously thought impossible. With a new interface and tailored product experience, knowledge workers within sales, marketing, CS, finance, HR, etc. will have contextual information at their fingertips as they work without needing to searching through files, emails, SaaS tools, databases, etc. Startups thinking outside of the confines of traditional software will deliver on a meaningfully more productive user experience that will be disruptive to the status quo.
- Copilots / Agents: As we’ve seen from the amazing release and rapid adoption of GitHub Copilot, I can imagine a world moving forward where nearly every employee has their own AI copilot or agent to assist and augment their work. When combining the right datasets with models fine-tuned for a vertical or functional application, these AI agents will be able to conduct research, automate repetitive tasks, and even generate content or recommendations to take action on. Examples: Yotascale (Crosslink portfolio co*) launched Yota AI Assist to serve as an engineering team’s copilot for optimizing cloud costs. Qualiti (Crosslink portfolio co*) serves as a copilot that can test software products for engineering / QA teams.
- No-code Automations: Given the proven ability of these models to auto-generate code and understand natural language, I am excited about the potential for combining natural language UIs with code generation to enable business users to build their own automations or even small applications end-to-end without needing to write a line of code. This will enable a new class of creators and builders and a new class of software to work alongside them.
Theme #3: Leapfrogging the Adoption Curve in Industry AI
As long-time investors in vertical technology businesses at Crosslink (Weave, ServiceMax, BuildingConnected, Overjet, etc.), we’ve learned that startups can be very successful when they go deep in an industry and focus on solving critical problems not addressed by broad / horizontal solutions. However, many verticals are still far behind the status quo in technology adoption, and it isn’t due to the availability of technology or the potential impact it can have on their businesses. The biggest challenge is adoption, particularly in some of our oldest industries, as you’re more likely to run into resistance or fear of change, deeply entrenched incumbents, and other industry-specific nuances. The vertical-focused companies that have succeeded deeply understood the workflows of their industry, usually born out of a lived experience by the founders, and have incorporated technology that introduces a 10x or better experience for critical aspects of the value chain.
For many industries that don’t rely on workers sitting in front of screens (i.e. manufacturing, construction, logistics, field service, etc.), traditional software tools often fail to meet the customer / user where they are, do not fit within their existing workflow, and fail to deliver on their ROI potential in result. It can require too much change in behavior to sell a solution with a high degree of implementation/customization requirements, user training, and requires a high degree of on-screen user interaction to capture enough value from many products on offer today.
By combining the latest AI developments with industry vertical-specific datasets and technologies (software, sensors, robotics, additive manufacturing), there is great potential to remove frictions and massively improve the user experience. I predict that startups with this approach will leapfrog the tech adoption curve in certain industry verticals that to-date have been behind in digital transformation. I’m particularly excited about bringing AI-powered automations, no-code analytics, and agents into the physical world to level up productivity / insights and create new solutions never seen before in industry. These are mostly all untapped opportunities that won’t be addressed by the horizontal incumbents or generalized foundation models.
- Natural Language / Voice-Based Interaction: Solutions that incorporate natural language or voice-based interaction have the opportunity to unlock productivity for workers in verticals that have previously been unaddressed by recent technology advancements. For workers that don’t spend their time in front of screens (i.e. field service, construction, manufacturing), taking actions hands-free will better fit their workflows and serve to maximize productivity.
- Robotics / Sensors: Most of the initial applications of the current wave in AI has been applied in software, but there are early upstarts working on interesting problems in applying sophisticated AI models into the physical world. I’m excited about the potential leaps that are possible here in decreasing the amount of implementation or configuration requirements for robotics, drones, AVs, and sensors by capitalizing on the rapidly growing intelligence of the latest models. A key challenge for these companies will be in gathering enough quality data from the physical world to bring model performance to where it needs to be — I predict that startups innovating at the data and infrastructure layer will see a growing opportunity.
- Unstructured Data Automation: Documents, emails, audio, and images, are the lifeblood for many industries, such as legal, financial services, healthcare, and insurance. Leveraging the latest in AI, companies can incorporate game changing data automation flows into new product experiences that deliver meaningful productivity gains to constituents in these industries and others. By tapping into imaging and other unstructured datasets, Overjet (Crosslink portfolio co*) has built a highly effective AI platform for the dental industry to enable clinics and insurance providers to enhance clinical care and administrative efficiency. Multi-billion dollar companies will emerge when combining advanced AI models with novel, untapped datasets and deep industry domain expertise.