A path forward for AI startups
Startups promise full AI automation but frequently fall short. Here's a more realistic path forward
I see a lot of startups claiming to have automated various business processes with AI. Think AI-led sales, to give one easy example1.
The problem is, AI technology is not nearly sophisticated enough (yet!) to cut people out of these processes. When AI technology is sophisticated enough to cut people out of these processes, those AI systems will be controlled by the hyperscalers & OpenAI/Anthropic. This means that the startups selling this layer of AI tooling will be cut out.
This is, suffice it to say, a hard row to hoe for these startups. And, yes, there are some business processes which today can be AI-led effectively, but most can’t. And, yes, there are some AI startups which have access to proprietary industry data which the hyperscalers will never have access to. As ever there are always exceptions.
Today’s AI systems require human oversight and intervention to be effective. Fully autonomous AI systems are largely aspirational outside of very controlled use cases. When these systems become truly autonomous, the hyperscalers will control them, given their advantages in infrastructure, compute, and data access.
This tension suggests a few strategic paths for AI startups that don’t rely on unrealistic claims of full automation:
Specialization and Proprietary Data: These are startups with deep domain expertise, access to proprietary datasets, and solutions tailored to niche but high-value processes. Industries like industrial sales, healthcare, and finance have data sources and processes that are difficult for hyperscalers to penetrate or scale. These startups’ markets are less likely to be commoditized, as their AI systems will be fine-tuned on highly specific data that general-purpose models, even ones that can be said to be AGI, can’t match.
Human-in-the-loop Systems: Position the AI as a tool that augments, rather than replaces, humans. Create AI systems that recommend optimal next steps, surface high-priority leads, or identify patterns, but leave decision-making to human users. This approach makes AI less threatening to potential customers, while also sidestepping the technical limitations that pure automation faces. Unfortunately, as AGI-capable systems come on line, this may no longer be a sufficient competitive differentiator.
Own Proprietary Data and Models: As foundational models become increasingly powerful and centralized by the hyperscalers, owning or licensing proprietary data will be a competitive advantage. Startups can focus on creating proprietary knowledge bases, ontologies, or fine-tuned models that leverage unique data or domain-specific processes. Proprietary data and customization act as barriers that make it difficult for hyperscalers or general-purpose AI providers to replicate.
Integration and Workflow Optimization: Integrate AI into existing workflows in ways that hyperscalers can’t. For example, focus on API-driven solutions that seamless connect with legacy software.
AI startups will face tough competition, but those that leverage unique datasets, focus on domain-specific needs, and offer value-added integration within existing workflows will have a stronger chance of fending off AGI-empowered hyperscalers. These startups would do well to view themselves as providing specialized, augmented tools rather than the more dubious proposition of full automation. The former is likely less valuable than the latter, but a thriving company is better than a dead one.
I am not naming specific startups in this post, for the very simple reason that entrepreneurship is hard, and me publicly calling out startups for overpromising, or for making fraudulent claims, doesn’t help anyone. Those who pay close attention to the AI startup space will no doubt be able to read between the lines and deduce some startups for which this post’s suggestions are relevant.