The Truth About Building AI Startups Today

The most promising AI startups automate mundane, information-processing tasks (e.g., government contract bidding) where humans currently read, summarize, and re-enter data—these are perfect fits for LLMs and face little competition.

Concepts like generic AI co-pilots or vague ‘throw your data in and we’ll automate everything’ tools are easy to sell and pre-sell, but hard to make truly useful because customers don’t know concrete use cases.

Relying solely on chat forces users to know how to ‘talk to a computer’; instead, use LLMs behind the scenes in traditional web or mobile interfaces that align with existing workflows and jobs-to-be-done.

Fine-tuning open-source models only for cost savings is fragile because foundation model prices keep dropping; lasting value comes from domain customization, proprietary data, and superior performance on specific tasks.

For domains like coding, hardware, or SQL parsing, older or smaller models fine-tuned on narrow vocabularies can be “good enough” or better, cheaper, and faster than state-of-the-art general models.