I went through 500 of the 1500 companies that got accelerated by Neo, YC, and Miracle+ (YC's Chinese spin off) in the last three year to find interesting AI use cases, and found these: Many dusty industries are stuck with old software that's abysmal but too dangerous to replace, since any downtime would halt the entire business. Computer use has the potential of changing this by allowing startups to build on top of old software, whether is is to extend it, or gradually migrate away from it. An example is Lance.live - they use computer use to extend traditional hotel management software. AI working with unstructured data enables new interfaces to meet users where they are. For example, when factory workers bypassed software by texting each other, textsidekick.com gave its AI a phone number. And when construction field inspectors spent hours putting together reports that rely on precise coordinates and building blueprints, opusense.com created a dictation + picture with live-coordinates + blueprints app that allows them to easily capture data while on site, and later generate reports automatically. AI allows companies to model human behaviors. One use case for this is social engineering attacks. Most phishing attacks used to be one-offs: attackers send an email, if it gets clicked, they win, if not they lose until they try again with a new email. With AI, it's becoming trivial to launch attacks that are a) hyper personalized down to exact writing style of the target provided they have an online presence, b) use many touchpoints across several channels (email, text, and even linkedin) and c) span multiple weeks of relationship building before malware gets sent. Ghosteye.ai uses this approach to test companies' defenses. Counterintuitively, having AIs run voice interviews, data gatherings, or calls of various sorts can results in a better experience than having underpaid, tired, and disengaged staff do it. Two companies betting on this to lead customer interviews are ListenLabs.ai and Miravoice.com. These two also exemplify how stark of a difference in customer perception a website can create. There are more govtech startups than in any batch before, which is not surprising when you consider the government is an administration machine, and that AI is good at automating admin work. Or is it? Plenty of government admin tasks are deceptively difficult to automate reliably. A good example is anonymization of citizens' names for things like Freedom of Information Act requests. ~95% can be caught with regex and dictionaries of names, but I know from working on similar datasets at Doctrine.com that the remaining long tail of edge cases require a lot of NLP work, or burning a lot of tokens. This is what led Govguard.com to focus on automating it for. Another benefit of FOIA is they require grunt work from whichever administration is concerned, meaning Govguard can close small administrative entities one by one instead of having to struggle for decades with the federal government. AI enables "outbound marketplaces", like with Lanesurf.com that instantiates swarms of agents to call, text, or email logistics carriers to buy and sell empty load. There are also many examples of good use cases for AI that don't translate in good businesses. For example, Serifis mines petabytes of unstructured data to identify narrative inflections for traders. I'd wager it's a bad business because you can't split searching for signal and turning it into strategies without heavily slowing down iterations, and therefore getting beaten by teams that bundle both internally. Flying drones will soon be everywhere. When I read about Grazemate.com using quadcopter drones to herd cattle, I intuitively qualified it an overkill use of quadcopters because flying is much less energy efficient than rolling on the floor. But as we learn in Ukraine, there's a long-tail of terrain features that result in land-drones getting stuck, while flying drones almost never do. As an engineer, if you have to choose between a long tail of unknown problems and a single known one, in this case energy density, picking the later makes a lot of sense. Curious to hear if you found something else interesting.