Thoughts and prayers

Watching the 4o demo on the 13th I could hear the sounds of trolls and journalists composing headlines and tweets like these: 

With the release of GPT5 as early as this summer, we’ll have to endure another doom wave very soon.  

As thousands of AI startups have proudly wrapped GPT models to deliver SaaS value, there will be inevitable casualties.  But what have smart founders been doing to prevent their use case from being swallowed up by the mega-brain?  We’ve thought a lot about what creates lasting differentiation for GenAI wave startups. We’ve identified the following keys to differentiation success. They work even in the wake of the 4o hype... 

Differentiator 1: Non-public data and insights

Each successive release of a GPT model from OpenAI and the other foundation titans is going to get better at creating text and code outputs that closely resemble the contents of their vast training sets. Anything that creates outputs with thousands of excellent examples on the public web is already or soon will be swallowed up.

If, however, founders can get access to data that these foundational models won’t have access to, they can create something more defensible. Even seemingly “thin” wrappers sometimes can capture valuable data. If customers give a platform unique, accretive private insights during use, that can make it more defensible.

Take a procurement copilot use case. There are lots of RFP writing apps, but if one of them gets significant traction in an industry, and can store data about win rates, edits users make, etc, it is conceivable that it could create meaningful differentiation—creating RFPs that are more likely to win business (not just look credible to a junior team member).

Many of us are used to thinking about data assets in the Big Data frame. It used to be that valuable data was purchasing patterns, lead lists, or segmentation data—all only valuable at great scale. Now that LLMs are in the picture, many different types of data are now much more valuable. AI systems can integrate Standard Operating Procedures, Internal knowledge bases, transcribed sales calls, business logic in the form of code, and much more. Even application configuration choices can become unique data assets in this new world.

Of course, the data acquisition imperative makes growing a startup’s user base even more important. Apps going for self-service scale need to hurry to compound that advantage. A few development partners might love a new service, but if the product differentiation strategy requires more data sources, it may be hard to create a powerful data flywheel without more customers.

Not only do AI startups have to accelerate data (customer) acquisition, they also need to remove obstacles to training models on that data. This means delivering instant value to users from their data, and finessing any concerns they may have about sharing that data. We counsel founders to educate their partners where possible to think about sharing data the way companies might think about a syndicated research provider. When competitors share data with syndicated research partners, they all benefit from the improved visibility they can all have on their market. In the same way, AI applications that syndicate learnings from multiple user’s data can deliver powerful advantages to their clients only when their model training can comingle client data. While this may complicate some discussions, it also creates a strong case for why the startup in question is highly differentiated from competitors, and GPTX.  

This is one place where new AI entrants likely will have some advantages. Lots of pre-GenAI SaaS applications handle valuable data, but their privacy stance and customer expectations preclude them from comingling data from different customers. It is hard to imagine that established players will invest in the tech to extracting the valuable but non-sensitive subject matter data and comingle it to train models—all while their legal teams fight the initiative tooth and nail.  The generic copilots they integrate with their SaaS platform will seem quite useful until they can’t compete with expert models that deliver agentic quality because they were trained on syndicated subject-matter learnings.  

Differentiator 2: Workflow integration

Increasingly, AI founders have to deeply integrate their solution into workflows to stay ahead of GPT5+. Sure, customers will be able to build directly with the major APIs, but that will take time and investment that will only be reserved for the most high-impact problem spaces (that won’t be likely to be outsourced anyway). And SMBs will not be able to justify custom software projects. That’s why adding Slack integrations, voice, email, beautifully formatted documents and more will be critical to staying viable.

The voice capabilities in GPT 4o are astounding and will democratize the voice modality. Seen another way, adding voice is now even easier for a myriad focused AI startups. When you pair deep and defensible subject matter expertise with a voice interface, the value these startups can deliver is going to be compelling. We recommend all AI startups quickly figuring out how to leverage voice APIs to make human-quality voice a workflow integration point.

Differentiator 3: Quality Assurance

The sticky wicket in building any kind of useful LLM application is making sure that it doesn’t goof up. Most founders have framed their offerings as copilots, but this is dangerous. If GPT5 gets smarter, there won’t be room for a bunch of tools that do things right only 80% of the time. Sooner than later, founders will need to figure out how to deliver high trust outcomes—delivering much more reliably than what ChatGPT/Claude/Gemini offers out of the box.

Agents are coming to every domain. If founders aren’t working on an Agent-level trustworthiness wil find it difficult to thrive in the coming copilot wars. Users aren’t going to be interested in 5 different specialized copilots when ChatGPT gets significantly better. They will however pay for a wide array of tools that deliver productivity outcomes end to end.

Delivering high-trust outcomes is hard. It often comes at a cost of the breadth of what a solution can cover. Many of this wave’s founders think that getting a host of use cases 80% right is their ticket. The problem is that during a free trial, customers will ask for at least five outputs. Getting even one wrong just isn’t going to help beat out other products—especially as GPT5 and other models keep delivering impressive intelligence gains.

Get Special

It should be obvious by now that all three of these strategies point founders in a very specific direction: Focus on specific jobs, tasks, and areas of expertise. There are benefits across the board When founders specialize: 

• They need less data from users to compound and deliver greater value in a given use case than the best LLMs. 

• They can afford to hug workflows because they are much more clear. 

• They can afford to build in guardrails on more use cases—delivering high-trust outcomes that move toward agent-level productivity.   

Framing a solution as specific for a job or job/industry slice is a great way to stand out from the herd of customer service, SDR, developer or marketing copy-writing wrappers.  

We have strong conviction the main way to win for new AI application companies is to carve out and own a specific job niche. It positions a startup for long-term value and captures a greater share of productivity benefits in the form of revenues. All while ensuring you don’t have to live in fear of the next OpenAI demo day.