Over the course of the past 4 years, AI has become the most popular industry or vertical among investors, both for the early stages of investments, such as pre-seed or seed, and later stages. The hype, unlike that of the mid-2010s crypto hype, is based on a real, useful, and impactful technology. AI can truly solve problems that we previously thought to be impossible or economically unviable.
The fact that most investors are gravitating toward AI is not a speculation but our assessment by analyzing our sizable investor directory at Funding Banker, which clearly supports the idea.
During our analysis of our directory, we could not find a single instance where an investor openly states that they would not invest in AI startups, while many investors have a list of verticals/industries that they stay away from (regular findings include Crypto, Gaming, Gambling, etc). The vast majority of them have invested in AI or AI-related startups, and 33.1% of them have explicitly mentioned AI as their vertical focus.
I'm sure that there are investors and VCs out there that, for one reason or another, actively stay away from AI startups, but we have not yet encountered any in our directory.
Analyzing our directory has led me to a few conclusions about the state of the investments in AI startups.
AI is becoming the new software
We have reached a point where we all know that any piece of technology will have a software aspect, whether as the main product, embedded in the product, or as an important tool during the process of manufacturing/creation. As a result, no investor would object to the presence or usage of software in a project. In fact, the absence of software would be extremely unusual.
Back in the 80s and 90s, many of the tasks, both casual day-to-day and critical, were handled manually or through analog means. The idea that we can automate a very large portion of these tasks with software led to the boom in the software industry and, interestingly, the late-90s dot-com boom, a period that is constantly referenced during any funding hype cycle.
AI is becoming the new software, and we are now facing the second wave of major automation in the world, where many of the tasks that still need human intervention could, to an acceptable degree of accuracy, be automated by AI.
Every investor (and soon every consumer) will implicitly expect the presence of AI in a project, either as the main product, embedded in the product, or as an important tool during the process of the manufacturing/creation.
However, we are still in the early phases of AI, in the same way that software was in its early phases in the 80s and 90s. We even have parallel conventions in naming where the old software companies included terms such as “sof”, “software”, or “technologies” in their names, and current AI startups often include the term “AI”.
AI as the product or AI as a tool
Using the previous anecdote of software, there was a time when owning the underlying programming language, operating system, or other software infrastructure was essential for a software company. If a company needed a tool, they most likely had to (or wanted to) recreate it in-house. They would still need to provide value with their in-house tools and technologies to be able to sell their product, but they needed to own most of the tools they were using.
In the modern software ecosystem, a startup (or even a large company, for that matter) owns only the very top layer that they have created. No one expects a company to only use what they have built in-house, as it would be, quite literally, impossible. We are all using programming languages, frameworks, libraries, packages, operating systems, cloud infrastructure, and many other tools that are created and offered by other companies. There is a thorough understanding, among every section of the market, that complete ownership is not possible, and each company owns a small portion of the stack.
Back in the early 2010s, developing an AI product meant owning a dataset, a robust pipeline of data transformation, proprietary engines, and a very talented (and expensive) team of data scientists and machine learning specialists. Of course, over the years, many aspects of this development process have changed with the introduction of open-source libraries, open datasets, and free pre-trained models. But the commercial aspect of AI remained, for the most part, the same. You own the data, pipeline, and the eventual product. If you were raising capital for your AI product, you needed to own most of the process.
After the introduction of ChatGPT, a new category of products emerged: products that would use or wrap the power of the AI models to extend the usability of their products. This created a new dynamic where a product (the wrapper) had virtually no control or ownership over the underlying technology (the AI model), had no guarantee that they'd still have access to the technology in the future, and could not protect the usability of their product, as thin wrappers could easily be replicated.
Initially, the wrappers were the best the investors could get. Startups such as Jasper (previously Jarvis) are remembered as cautionary tales as they were thin wrappers with no tangible added value. Other companies, such as Cursor or Lovable, despite their initial complete use of external models as their main selling point, provided a tangible augmentation layer on top to survive long enough to roll out more in-house solutions and own a larger part of their stack.
The current attitude of the investors toward AI startups is neither similar to the 2010s nor similar to 2021-2023.
What the investors are looking for
Based on what we have seen from the investors on the Funding Banker directory, the fundable AI startups can be placed in the following categories:
1. AI pipeline
The first category of AI startups does not offer any AI models but rather provides tools, infrastructure, and workflows for companies training their AI models or using pre-trained models.
Virtually every medium to large company in the world is rushing to utilize AI to improve their workflows and products, and we are witnessing companies struggling to create a robust pipeline of data for the AI models they are trying to create or adapt. Developing a data pipeline, infrastructure, and execution layer for AI models is different from the conventional deterministic paths, and many of the tools and technologies used in the current setups do not fit the AI workflows well.
The problem is real and tangible, but the startup needs to have a clear picture of who their customers are (developers, enterprise, cloud providers, etc), whether those in their target demographic are willing to pay, and what their options are to market their tools.
A considerable portion of the startups in this category choose the open-core model, where the core of their offering is free and open-source with a paid layer on top.
2. Proprietary augmentations of the pre-trained models
When ChatGPT API was introduced, we saw a blizzard of thin wrappers receiving substantial funding. Almost all of the main features of these thin wrappers came from ChatGPT or other models they did not control, merely providing UI, a workflow direction to the user, and appending context to each prompt, reminding the AI models they were an "expert" in the field. Needless to say, thin wrappers rarely raise any capital nowadays, unless of course, they gain a meaningful traction.
However, a large portion of the AI startups that raise capital still use AI models they do not control or own. The difference is that these startups can prove that they have a proprietary augmentation (whether through data nobody owns or other parts of their products) that:
- a. They provide a meaningful solution to real-world problems that could not be previously fixed with existing software/AI combo, and only their augmentation can lead to a solution
- b. Their augmentation is so unique that it can neither be easily replicated nor replaced by “just asking ChatGPT.”
The most visible mistake that I commonly encounter is founders presenting their AI startup as simply "AI-based HR platform" or "AI for HR" with the mere presence of AI being the only selling point of their product. If you find yourself using the presence of AI as the main selling point of your product, you need to find a set of tangible problems affecting a sizable portion of the market that your AI-based solution solves. You should lean into the AI aspect of your product, but do not present it as the only unique angle.
Proprietary models
Even though this is the most important category, it is also the trickiest of the categories to raise funds.
Many real-world problems require a new model with a new approach, and it cannot be solved by simply using the models that are already available in the market, paid or free.
Imagine an AI model that is used in the detection of pre-cancerous cells in histological slides. Not only can you not simply use an off-the-shelf model for that, but you also need various regulatory certifications (such as FDA approval), which forces you to control most of the lifecycle of the data, model design, and training. This was largely the case with PathAI, which needed to design and train its own models.
Even though AI models are more accessible than ever, training an AI model from scratch is still a very lengthy and expensive process with very little guarantee of a usable outcome and/or commercial viability, and that's all based on the condition that you have access to the necessary large datasets in the first place.
Such startups have a few problems:
- a. They need to prove that there is no existing equivalent model
- b. Believe it or not, they have to prove that ChatGPT/Claude/Gemini cannot do what their model does with “a few prompts.”
- c. There is a commercial demand for their model (or its outcome) as a product
I am going to elaborate more on the last problem, as many founders confuse the usability of their custom AI model and whether their model could be sold as a product.
Embark Studios developed a popular game named Arc Raiders. The studio created a series of technically impressive AI models for their NPCs, which created a very unique and interesting way that robots move, navigate, and interact with the environment. Now, the AI models that were created for this game are extremely useful and valuable for the game, but cannot be sold as a standalone product. The AI models are so tailored for the game that their usefulness and practicality are limited to the game's NPCs. Embark might, in the future, find a way to monetize these AI models as a standalone product, but that's a completely different story.
If you are developing a custom model, you need to look past its usability and adjust your strategies to ensure that the model can be sold as a product.
AI control
The last category of AI startups works toward the interface between humans and AI models.
Software, in general, is used in the most critical aspects of our lives and society, where the stakes could not be higher. Sometimes, the software itself needs to be extremely reliable (such as in the case of the aviation industry), and sometimes the data needs to be in an airtight and secure environment (such as in national security).
Over the past decades, we have created elaborate, complex, and effective workflows and processes to ensure that everything would be in order, people's information would not leak, airplanes would not crash, or people would not get hurt while using a medical device, and for the most part, we have succeeded. Things do go wrong, processes do fail, and workflows sometimes prove to be ineffective, but overall, we see fewer incidents as we go along.
The introduction of AI changed everything and a whole new frontline was opened as almost every workflow, compliance, audit, verification, implementation, and monitoring we had designed for the past 50 years, have to be rebuilt for the AI models, which unlike traditional software, are often viewed as a blackbox.
As a result of these changes, we are seeing a new category of startups that work toward better usage, after-effect, and blast radius of the AI models rather than developing AI models or offering AI-based solutions.
The problem that companies, governments, and semi-government organizations face is very real and tangible and this category of startups are looking to solve them.
The biggest challenge that these startups face during fundraising is proving their ability into entering enterprise or government industries which are notorious for difficult and long sales-cycles.
If you operate in this category, your biggest hurdle in convincing investors is your path to commercial viability and need to address investors with existing enterprise/government networks. Fortunately, many investors openly prefer B2B and enterprise projects over consumer-facing ones.
What you should do as a founder
The first thing you need to do, if you are a founder of an AI startup, is to find the right category for your startup. Not only the strategies that you need to adapt are different for each category but also you'd often find investors prefer one category over the others. Pitching the wrong investors could be a fatal waste of precious time and resources. This is a place where Funding Banker could be of great help.
Second, be clear about the classic problem/solution pair that every startup has to state. What problems the customers face, how you can solve them, and what is your plan to get paid for it. Stay away from the mistake of using the mere presence of AI as your main selling point.
Third, chase traction ruthlessly. Investors always considered early traction as the ultimate indicator for long-term success but the bar has been raised significantly. Even though creating tools and products, even to the point of MVP, can be a very costly endeavor (which many founders may not be able to afford), there has been a growing sentiment that anything could be built with AI in a matter of days using little to no resources and if a product does not have initial traction, injecting more resources would not make a difference. Even though the sentiment is partially true, it doesn't change the fact that any business activity requires capital. I know, it's a chicken and egg problem.
Investors are focusing on AI startups more than ever. But they've matured in their approaches and evaluations. Instead of a magic word, they see AI as a tool to solve large problems, a powerful tool, but still a tool. Founders who treat it that way will find investors ready to listen.