YC’s latest batch sure was a lot of ‘maybe AI can do… this?’

Sitting through hundreds of startups on toptechtrends.com/2023/04/05/y-combinator-demo-day-favorites-part-one/”>YC Demo Days, you’re not always sure whether you are actually perceiving patterns or if your brain, as coffee battles with monotony, is inventing them in a kind of pareidolia for business plans. This year, though, the theme was pretty obvious: “AI can do that, probably! Maybe.”

Certainly today’s AI models are more capable than yesterday’s, and yesteryear’s. But we’ve seen over and over how these systems demo well but fall down under systematic requirements or as tools with reliable and repeatable results.

It’s hard not to see this batch as the  precursors of a coming wave of AI-powered shovelware. Pick a use case, do a little fine tuning of an available model (no one actually builds their own), cherry pick some good examples for screenshots, and bolt on a prefab UI. Congratulations, you’re now the very first AI social media content generation platform for independent bars and restaurants in the Middle East and North Africa. Buy a couple hundred 5-star reviews and you’re on your way!

Now, it’s not that restaurants in Cairo and Beirut couldn’t use a helpful tool to gain some traction online and attract new customers. It’s that having AI, as it currently exists, do something for you is kind of like admitting that it doesn’t matter.

Creating an AI-powered conversation agent that answers the phone at your business sounds good when you frame it as a way to never lose a customer. But what does the customer think when the business they call decides AI is the reception they deserve? Personally, I would hang up and try someone else. What about a trade worker who gets an AI calling to make an appointment? Same thing.

Realizing an email to you has been trivially “personalized” by AI is like being told, we can’t be bothered to personalize our emails, but we want you to think we do. Wouldn’t you feel tricked? It’s a systematic imposture upon the customers.

If your first interview with a company is with a conversation agent or a person obviously reading generated cues from the knowledge base or whatever, do you feel like a person joining a team or a part being sized up for installation? You’re not even worth the full attention of a qualified human.

That’s not necessarily the vibe I got from every AI startup in this YC batch, but I sure got it from a few of them. Here’s a partial (!) list of the “AI can do that, probably” companies I jotted down.

• Type – AI-first document editor.
• Iliad – Generate game art assets.
• Layup – Build workflows across apps with one line command, like onboarding a hire.
• Nucleus – AI-powered onboarding orchestration that understands “the true nature of a business.”
• Hadrius – SEC-compliance robo-advisor.
• Speedybrand – Generated marketing content for SMBs.
• Quazel – Language learning with an AI tutor.
• Booth.ai – Generative AI “photographer” for e-commerce.
• Squack – Natural language accountant tools.
• Berri.ai – Creating ChatGPT apps as a service.
• Semantic – Financial news insights “enriched” by AI.
• Credal.ai – ChatGPT-like interface for employees that references company docs but protects business secrets
• Defog – Add AI data assistant to your app.
• Linkgrep – Suggests things from knowledge base and adds to chat or notes live in browser.
• Sail – Automated sales emails.
• Aiflow – Automate market research based on reviews and feedback
• Tennr – Turn knowledge base into a custom LLM.
• Truewind – AI powered bookkeeping and finance processes.
• Flair labs – Collect insights from customer service call data and emails.
• Justpaid – Automate bill pay, catch over-payments to vendors.
• Kyber – Automate insurance industry tasks like answering questions and underwriting.
• Meru – Platform for training your own LLMs.
• Sameday – AI that calls workers like plumbers and roofers to make appointments
• Zenfetch – Analyze customer calls live and surface talking points.
• Syncly – AI to analyze customer emails.
• Pair AI – Video courses generated using AI.
• Latent – Automating electronic health records.
• Avoca – AI receptionist to answer missed calls at SMBs.

Until about 30 seconds ago, I actually had appended thoughts about the companies to these brief and likely insufficient descriptions. But I realized the list was in danger of becoming a litany of complaints (not to mention way too long). No one likes to read someone just shooting down ideas left and right, especially when many of those ideas are being worked hard on by people for whom they are important. It’s easy to criticize. So easy someone in the summer batch may try to automate it!

But I challenge you to look at that list and not wonder about some of the entries: Is that really what’s needed? Won’t that need lots of oversight? Doesn’t this introduce liability, or decrease transparency? Did anyone ask customers if they want this? Who verifies and audits the results – another AI? Who is displaced by these tools? Who trains people on them?

Practically every company that presented said they’d gone live a few weeks earlier and miraculously were already at some healthy ARR. But a few weeks is hardly enough time for a major automation tool to be even installed and the documentation read, let alone evaluate its performance and whether it’s worth the price tag. I can’t imagine even half of these have been used, really used, by a potential customer.

One example I can’t help but share: a generative marketing imagery company in its slide had the following prompt for the system to work with: Our classic ketchup is made only from sweet, juicy, red ripe tomatoes for the signature thick and rich taste of America’s Favorite Ketchup. The AI’s copy: SWEET & JUICY KETCHUP FOR ALL! If I was a marketer at Heinz and that was in the demo I was given, I would stand up, thank them for their time, and open the door.

Some of the companies admitted they’d pivoted halfway through the program and wrote their first line of code for this new application just recently. Of course we must allow for the adventurous and freewheeling nature of early stage startups, that’s part of the fun and excitement of the space. But do these companies really feel “innovative” to you? They seem rather to be big fans of innovation, sneaking into its room and trying on its clothes. (“Cute… here, you try it on, fintech.”)

I know I’m underestimating the amount of work it takes even to build the most perfunctory AI-powered B2B SaaS service, but a lot of these feel like our old hackathons where someone would make an API available and everyone would try to shoehorn it in to the most realistic-sounding application, hoping to get that $1,000 gift card from SAP or whatever. There’s joy in the process of creation but the results don’t really stand on their own.

Probably I’ll be proven wrong when one of these companies goes unicorn and everyone laughs at the TechCrunch writer who doubted them. But I can’t shake the worry I felt in hearing founder after founder say with such conviction that their AI could do something better, when I suspect that conviction is has been cultivated upon false pretenses.

toptechtrends.com/2023/04/08/ycs-latest-batch-sure-was-a-lot-of-maybe-ai-can-do-this/”>YC’s latest batch sure was a lot of ‘maybe AI can do… this?’ by toptechtrends.com/author/devin-coldewey/”>Devin Coldewey originally published on toptechtrends.com/”>TechCrunch