Harnessing AI: A Guide for Startups

This post summarizes a Y Combinator Office Hours episode, How To Use AI In Your Startup (January 2025), in which four Y Combinator partners talk about how startups can use AI. The views and examples are theirs, from a venture investor’s seat. They are advice, not research.

Key Takeaways #

• Don’t pivot to AI just because it sounds good. The partners say no to that, and yes to building something with large language models at its core.
• Use AI inside the company too. One partner says it makes little sense for a new company not to use AI internally, as with the cloud.
• Watch someone do their job. A few hours at a friend’s screen, one partner says, shows which repetitive tasks could be done better.
• Place matters, in their view. The partners recommend spending time in the Bay Area, where the people working on this are.

The AI Landscape for Startups #

The question put to the group is whether a startup should pivot to AI. One partner gives two answers. The strong one: no, do not pivot because it seems like a good idea. But yes, you should almost certainly be working on something that uses LLMs at its heart. Their summary at the end is that you probably should pivot to AI “with some caveats”.

The fundamentals still decide whether it works, they add: switching your idea to something that calls an API will not change a startup’s fate if it does not deliver value to customers.

Why AI Matters, According to the Partners #

• Efficiency: one partner describes a company that provides homeowners-association (HOA) management as a service and uses LLMs to make its internal work more efficient.
• A repeat of earlier shifts: in the early 2010s there were many good ideas lying around in replacing on-premise software with cloud versions, and one partner sees the same now with software rebuilt AI-native.
• Big companies are slow: a partner says not to fear them, since ChatGPT was about two years old and, he says, Alexa is still not very smart.

When to Consider AI #

From the discussion, three questions to ask:

1. Does your idea work better with LLMs at the core? If you are only adding a call to a model API, the partners say it will not change your fate.
2. Which repetitive tasks could a model do? Sit next to someone doing the job and watch their screen.
3. What will be possible when models are twice as good as today? One partner says this is the question he would ask as a founder.

The Importance of Community #

The partners make the case for being in the room where it is happening. The mobile wave is their example: one partner recalls arriving at YC in 2007, the week Steve Jobs presented the iPhone, and says the interesting mobile companies were mostly started over the following five years.

They recommend that founders:

• Come to the Bay Area, even for three or four weeks, if not ready to move.
• Go to hackathons and meet people working in the field.
• Embed in the community, which they say makes the opportunity easier to see.

One partner gives a story of a YC company that moved to San Francisco and switched products after seeing the LLM wave. That is one story the partner knows well, not evidence that moving works.

Real-World Applications of AI #

Examples the partners give from recent YC batches:

• Replexica (now Lingo.dev): translates a software interface from one language into another with AI. See Lingo.dev on Y Combinator.
• Gecko Security: described as an AI security engineer. See Gecko Security on Y Combinator.
• A Medicare Advantage copilot: a team that came from an insurance-technology company built an AI copilot for agents who sell Medicare Advantage plans, a workflow few engineers know.
• Healthcare administration: one partner says, “I believe”, that US healthcare spend is around $4 trillion, of which $1.3 or $1.4 trillion is administration. He describes much of that as legacy software plus a person moving data between systems, and describes a startup that uses LLMs to automate pre-authorization requests. Treat these totals as a partner’s recollection, not a checked figure.

The partners’ common thread: a specialized skill or obscure workflow, automated behind an interface that people without that skill can use.

Conclusion #

The advice from this episode: do not chase AI for its own sake. Build on LLMs when they do the core work better, use them inside the company, learn the repetitive work of one group of people, and spend time around other people building in the field. It is advice from investors who profit when founders build, so weigh it with that in mind.