The misunderstood AI Wrapper Opportunity

For those not familiar with the term “AI Wrapper”, it has become quite popular over the last couple of years in the tech ecosystem. It is a dismissive term that refers to a lightweight application or service that uses existing AI models or APIs to provide specific functionality, typically with minimal effort or complexity involved in its creation.

A popular example of an AI wrapper are apps that enable users to “chat” with a PDF. This type of AI application allows users to upload a PDF document, such as a research paper, and interact with an AI model to quickly analyze and obtain answers about the specific content. In the early days of ChatGPT, uploading documents as part of the prompt or creating a custom GPT was not possible, so these apps became very popular, very fast. These apps could be built in a few days by one or two developers.

As it happens with every hot space (e.g.: crypto), a lot of entrepreneurs looking to make a quick buck, jumped in at the opportunity heads on, flooding the market with AI products that lacked long term defensibility and vision, making them highly susceptible to getting steamrolled by major players like OpenAI with the launch of a single feature. As opportunistic founders quickly moved on to the next hot thing, “AI Wrappers” or better put, startups building in the application layer became second class citizens in the AI space.

AI Wrapper Meme

The AI Wrapper Fallacy

YC partners recently shared some interesting thoughts on AI Wrappers on an episode of their Lightcone Podcast, their logic: calling a startup an AI wrapper for building on top of OpenAI’s API’s is the equivalent of calling a SaaS company a MySQL wrapper for building on top of a SQL database.

I think a better analogy of this would be to dismiss SaaS companies like Aircall or Talkdesk as a Twilio wrapper. These companies have built multi-billion dollar businesses by literally outsourcing their core capability, telephony and telecommunications to Twilio. Why? Because by the time these businesses were born in the early 2010s, VoIP was highly commoditized. Spending millions of dollars in re-building VoIP infrastructure would add no value or differentiation to their business.

Instead they leveraged a new platform like Twilio to offload key VoIP functionality like WebRTC, IVRs, Call Recording, Queuing, and more, and focused their entire strategy on building value on the workflow and application layer. This gave Aircall and Talkdesk a unique advantage: unmatched agility to innovate in product instead of infrastructure, enabling them to focus their R&D investments in adding more value to end users:

1. Developing pre-built integrations that enable customers to quickly, easily and cheaply integrate business applications like their CRM, Ticketing system, etc.
2. Providing tools for key stakeholders like deep analytics and real-time reporting.
3. Freeing employees to work anywhere through desktop and mobile applications that eliminate the need for clunky desk phones.

This is actually how most of the tech ecosystem has worked for years, and I would argue business in general. It’s really hard to integrate vertically and do everything. Instead most businesses rely on partners across their value chain to deliver products and services to their customers. I think the same logic will play out in AI, and startups that understand this dynamic will be able to take advantage of the unique opportunity that comes with this massive platform shift.

Understanding the different Layers in the AI Ecosystem

My 30,000 foot view is that the current AI ecosystem and Tech stack is composed of three major layers (we could break this down into more layers, but we’ll use these three for the sake of simplicity):

1. The infrastructure layer which is dominated by the trillion dollar companies that own the core infrastructure like Datacenters, GPUs, OS, etc. Players like Microsoft, Amazon, NVIDIA, Google, etc. Very few startups will be able to compete in this market (Groq is an interesting case). The upfront costs are so high, that it becomes a massive barrier to entry.
2. The model layer with a mix of large scaleups like OpenAI, Anthropic, Mistral, and tech giants like Meta and Google which are aggressively competing to build better and more capable foundational AI models. It is unclear if one company will create a breakthrough model that vastly outperforms the others or if these models will be very similar in capabilities and will become commoditized. So far, OpenAI is slightly ahead. Competing in this category requires a lot of money, making it unlikely that a large number of new startups are going to appear as contenders in this layer. Niche players have an opportunity to focus on specific or hyper local AI models.
3. The application layer is where the capabilities of AI are made tangible to end users. A great example is ChatGPT, the most successful product that kickstarted the AI race. Through this application, end users can interact with foundational models (GPT4o) that run on core infrastructure (e.g.: Microsoft Azure, Nvidia H100s, etc.).

I believe layers 1 and 2 are going to be mostly dominated by large companies. They are going to invest trillions of dollars collectively into these two layers to compete against each other, which will in turn create massive waves of innovation for startups that can leverage things like new foundation models, APIs, and more, to disrupt specific categories downstream at the application layer (both B2B and B2C).

The massive opportunity for startups in the Application Layer

The launch of ChatGPT was fascinating. From a product perspective, it had just one core feature that worked exceptionally well: seamless back-and-forth chat between a human and an AI system. It lacked all the bells and whistles modern assistants like Siri might offer: voice interaction, task execution (e.g., setting an alarm), or internet connectivity (check the weather). It was a bare-bones text chat app. Why weren’t these features needed for its success? Because they introduced a great core product in an entirely new category. There was no competition, no benchmark — it was a vast blue ocean.

ChatGPT User Adoption

AI adoption is growing rapidly and will accelerate even more. With the introduction of Apple Intelligence, and AI being deeply integrated into Android, Google Workspace, Microsoft Office, and other platforms, millions of people will become daily AI users.

Pair these two things together, and you get blue ocean opportunities in a lightning-fast growing market. New startups building in the application layer can leverage new innovations in layer 1 and 2 to leapfrog incumbents that have been building for years in the space using legacy technologies. Additionally, incumbents need to deal with the maintenance of existing products and platforms, active customer bases, and potential cannibalization. This makes them slower and susceptible to leaner and more agile startups that don’t need to carry that baggage. Think of a cargo ship which might not react instantly to captain input versus a speed boat that is highly maneuverable.

What might not seem obvious at first, but to me, is a hidden opportunity is that this new type of software introduces an entirely new paradigm. Moving from traditional deterministic, rule-based software to AI first products requires a fundamental change on how we approach software, both as users and vendors. So an interesting product that might seem at first as a simple feature, like ChatGPT, might actually represent the foundation of something much deeper, that requires a grounds up rebuild to succeed in the era of AI.

Consider a new startup entering the communications space. They might compete with Aircall or Talkdesk. Will they start by building legacy infrastructure like IVR (Interactive Voice Response) Menus and softphones for end users? Or will they focus on creating specialized AI Agents that can handle inbound calls in every language, execute complex workflows like setting up appointments, and transfer calls to humans when necessary? The new startup will likely concentrate on the latter, building on the AI application layer to deliver a robust platform that enables companies to securely deploy these AI Agents in production. This shift requires a new approach and platform to support:

1. AI agent design: Using instructions and goals versus pre-configured phone menus.
2. Tooling and workflows: Adding integrations, intention-based routing, and sub-agents to precisely execute business workflows.
3. Quality assurance: Testing an AI Agent requires powerful simulation tools that benchmark performance and security versus traditional manual or automated testing strategies.
4. Deploying to production: AI Agents need to be monitored and audited for continuous improvements, just like a team of human agents would. Customers will need call recording, transcription, real-time alerts, and conversation analytics.
5. Upgrading: Changing a foundation model, like GPT-3.5 to GPT-4, can have a massive impact on agent output. Having the right platform to test and deploy these changes is key.

Add vertical-specific integrations to enable customers to purchase phone numbers globally in seconds, SIP integrations to connect their existing phone system, etc., and the startup has built a significant moat that safeguards them from new rollouts from OpenAI, Anthropic, and major players. Actually, innovations upstream and launches are highly anticipated. OpenAI will likely empower their company. A great example is the new voice feature in GPT-4o.

I remember this article about the Salesforce ecosystem which showed that for every $1 Salesforce made, their ecosystem made $6. My gut feeling is that something similar will happen in AI, with the application layer capturing massive amounts of spending and value, and creating space for new generational companies to be built on top of layer 1 and layer 2 innovations.

Just make sure you choose the right strategy for your startup. Sam Altman explains this very clearly in this 2 minute video below.