How Retool AI Differs from LangChain (Hint: It’s Automation)
Last month, the low-code company Retool launched a new AI product called Retool AI. It’s one of the first low-code platforms to explicitly add AI, so I reached out to founder and CEO David Hsu to talk about the new tool. We discuss how Retool users are making use of AI, what types of products are being built with it, and how it differs from other AI-building tools (like LangChain and Vercel v0).
In the nascent AI engineering industry, user interface tools are still a wide open market. At this month’s AI Engineer Summit which he co-founded, Shawn “swyx” Wang did a presentation that mentioned several UI candidates for AI engineers: Vercel, LangChain and Retool.
Given this context, I began by asking Hsu how Retool AI differs from the likes of Vercel’s v0 (which is clearly a UI tool) and LangChain (more of a prompt engineering tool, but it seems AI engineers are using it as their UI to interface with LLMs). Hsu’s reply was somewhat surprising, because at first, he downplayed the generative AI revolution.
“Our sense of the landscape is that there is a lot of excitement around AI, but actually very few real use cases,” he said.
He went on to say that ChatGPT has been successful because of its convenient, easy-to-use form factor. However, Hsu thinks the real potential of AI, at least when it comes to developers, is to help implement “a continual automation of things.”
How Retool AI Is Used for Automation
Like its parent tool, Retool AI also uses the Lego analogy. The company describes it as a way to “integrate AI into your apps and workflows with pre-built blocks.” These blocks are the key to automation.
“If you look at work that has been done inside of a company, we think that probably 50-to-60% of that work can actually be automated away,” said Hsu.
He then explained how one of its customers, the FinTech company Plaid, is using Retool AI to automate fraud detection. Previously, Plaid had been using humans to do fraud detection. But with Retool AI, they now use an LLM to automatically detect the most common cases of fraud. Now, only the top uncertain cases need to be flagged for human review, which makes the overall process of fraud detection more efficient.
Hsu added that its scale as a platform makes it easier for automation workflows to be created (as “pre-built blocks”), which its users can then deploy.
“Every day probably around 100,000 hours are spent inside of Retool by actual users,” he said. “And these users, every day are going around detecting fraud, maybe they’re refunding orders, maybe they are processing equity grants, etc. And we have so much data that we can use for automation or augmentation of humans.”
As for what type of AI technology is being used in Retool, Hsu said that they vectorize the data their customers want to use [into a vector database], and then offer them a choice of LLMs.
“Because we already have access to all your databases and APIs, what we do is we actually automatically ETL [extract, transform, and load] all that into a vectorized database,” he said. “Then you can choose what LLM you want to use […] and then you can start building automations; and […] all of those automations have full access to all your internal production data.”
Retool vs. Vercel and LangChain
The way David Hsu described Retool AI seemed different to how the likes of Vercel and LangChain are approaching AI tools. Perhaps sensing my confusion, Hsu explained that Retool’s approach is that “AI is more helpful for the applications themselves, not in the building of applications.”
The difference is most obvious with Vercel’s v0 — currently in private beta — which “uses AI models to generate code based on simple text prompts.” So v0 is a tool that uses AI to build applications (that may or may not use AI themselves).
LangChain has similarities to Retool AI, because the applications it helps build are likely to be AI-powered. LangChain is a framework built around LLMs — it’s effectively a developer interface to LLMs and other AI products (like vector databases). But LangChain’s core function is to build better prompts, in order to get better responses from LLMs. So it’s a prompt engineering tool, rather than an application development tool like Retool.
According to Hsu, another core difference between Retool and LangChain is that Retool has a large customer base from which to draw. He estimated Retool’s customer base as around 12,000.
“And these customers have all connected their production databases and APIs to Retool — and they’re building apps on these production databases and APIs, all the time. […] And then, if you want to go build AI applications, for example, they can be immediately plugged in — if you will — to these databases and APIs, [to] all your data immediately.”
Hsu remarked that it’s an exciting time for developers when it comes to AI tooling, since they can use LangChain for prompt engineering and a tool like Humanloop for testing LLMs. He thinks both those products are great at what they do.
However, he claimed “there are very few companies actually focused on: how do I deliver applications on top of my data that will help me automate actions on my data?” This, of course, is what he’s positioning Retool as.
AI + Internal Data
What became clear during our conversation is that Retool sees AI as an additive to what it already offers its customers: a low-code platform for developers. With Retool AI, devs can now add LLM-based automation to their applications. As Hsu put it, “the problem that we uniquely want to solve is: AI does not work for you because it is not connected to your internal data.”
I also clarified with Hsu that the automation he’s referring to is not the same as what the AI engineering crowd call “AI agents.” He said they aren’t trying to make fully autonomous apps — or at least, not currently. But he didn’t rule out that type of functionality for future releases.
“We think we’re probably still […] maybe a year or two away from truly autonomous agents… inside of Retool at least.”