The push to integrate AI technologies into IT departments is intense, but it must be even more so for companies within the tech industry — especially if their own product offering has been stuffed with AI.
So I was intrigued when I was approached by Webflow, one of the largest website builder platforms, to talk about how it has fully integrated AI into its own engineering department. I spoke with Webflow CTO Allan Leinwand about the company’s “all-in” philosophy on AI development.
Webflow has about 300 engineering employees and in a recent blog post, Leinwand explained that “we’ve made a company-wide commitment to bring AI into every engineer’s daily workflow.” Webflow provides each of its engineers with “a complete AI toolkit,” which includes a ChatGPT Enterprise license, access to Cursor and Augment Code, AI workflows in GitHub, and more.
The coding tools in the AI toolkit were selected based on which ones Webflow’s engineers used the most. There is flexibility within that, but what about when small teams are working on a single project — does Webflow prefer that each team member use the same AI coding tool?
“I don’t think it really matters that much,” Leinwand replied. “From our perspective, we don’t really view AI as something that has to be prescriptive. We want to let the developers find the tools that work best for them. And you know, to me, the output is really what matters, right? So I’m looking at things like, did our PR rates increase? Did our change failure rates increase or decrease? Did we see cycle time, from the time the first Jira ticket is written all the way through deployment?”
Using AI Across the Software Dev Cycle
On the DevOps side, as well as Jira, Webflow uses GitHub — although not Copilot.
“We don’t use GitHub Copilot on the AI side,” said Leinwand, “but we use a lot of the other [AI] features embedded inside of GitHub.”
In his blog post, Leinwand explained that they use “an opt-in AI-powered PR [pull request] linting tool that engineers can trigger by tagging a PR with ai-linter on GitHub.” (Linting is when you run a program to analyse code for potential errors.)
“What that does,” he explained, “is it basically runs a webhook that tells Claude Code to go in and actually read the PR, read the code changes, and make a PR description and put it right into the PR — so that’s pretty cool.”
Leinwand added that they spend a lot of time thinking about how to take the developer workflow (from coding to deployment) and “making sure we can use AI across that entire cycle as best we can.”
“That testing infrastructure, through the CI/CD process, is being augmented a lot with AI.”
– Allan Leinwand, Webflow CTO
That said, he noted that the build and deployment parts of the cycle don’t use much AI. The company uses Buildkite as its build tool and they deploy on AWS.
“So we’re basically deploying with Docker containers on top of Kubernetes,” he added. “Not a whole lot of AI in there. It’s more like automation that picks up the container and then deploys it across our global infrastructure.”
For software testing, on the other hand, Webflow uses AI extensively.
“Engineers can quickly validate assumptions [with AI],” said Leinwand, “they can write unit tests, they can write functional tests, they can generate end-to-end and smoke tests. […] That testing infrastructure, through the CI/CD process, is being augmented a lot with AI.”
Humans in the Loop
In his blog post, Leinwand wrote that although AI is increasingly a part of their development process, they “continue to rely on human judgment where it matters.” He mentioned code reviews as a specific example: “Every change, whether written by a human or generated by AI, is reviewed by an engineer.” I asked him to elaborate on how Webflow does this.
“Yeah, we actually have augmented that [process] since I wrote that post,” he replied. “We actually have a tool right now which goes through and looks at every PR that’s going in, trying to merge in the main [branch], and we evaluate it for risk: low, medium and high.”
“…we have humans in the loop to determine how many people up the stack need to review that code before it gets pushed into prod.”
– Leinwand
How this works in practice is that the AI tool will read the code, determine its impact across the code base, and then rate its risk level.
“And based upon that low, medium, high [risk], we have humans in the loop to determine how many people up the stack need to review that code before it gets pushed into prod,” he explained.
While human engineers are still very much core to Webflow’s IT department, the company is using AI agents. Leinwand noted that it has enabled agents both in Cursor and in Augment Code. But he’s careful to add that it’s very specific work that these agents are tasked with. He talked about having “a series of remote agents look at a particular line of code and go off and do their own work.”
“We don’t quite have the swarms of agents that you’re starting to hear about in the industry yet,” he continued, “but we definitely have the ability to take a piece of code, put it up into Augment Code, and say: Go work on this particular issue, and come back […] and tell me the end result.”
The agent will usually come back with suggestions for PR changes, he said. The human engineer then evaluates that, and if the change is accepted, they will “merge it into their diff, put that diff into the build system and run it through tests, and see if they want to push that to prod.”
MCP Usage in Webflow
Other than agents, Model Context Protocol (MCP) is perhaps the hottest topic in AI right now. Webflow is using MCP both internally and as part of its customer offering.
“Webflow as a product offers an MCP server, in partnership with Cloudflare. So Cloudflare hosts and authenticates our MCP server, and that allows our customers to talk to their websites, talk to their CMS, talk to their design systems, using an LLM.”
Internally, Webflow engineers use MCP to talk to products like Jira and GitHub (via the MCP servers that those products offer).
MCP is “the stitching for the workflows inside of engineering.”
– Leinwand
“If I’m in Cursor and I’m in Augment Code, I use MCP to talk to Jira, I use MCP to talk to GitHub, I use MCP to talk to various parts of our code base.”
Leinwand describes MCP usage internally as being like “the stitching for the workflows inside of engineering.”
AI Resistance
It does seem that Webflow’s engineering department is getting a lot of value from AI — but we all know that a proportion of developers are anti-AI. I asked Leinwand what the response to its AI toolkit has been from that archetypal cranky old-school engineer (of which there are always at least a few present in any large IT department)?
“One of the bigger challenges we had was the natural resistance [towards AI] and, you know, ‘Who Moved My Cheese’ sort of thing,” he admitted.
But according to its internal metrics, about 89% of its engineering staff now use AI tools on a daily basis.
“That’s 17% up from what we even saw in Q1, so we’re seeing a pretty good jump in terms of adopting the functionality and the tools,” said Leinwand.
He added that as the AI models have evolved and the tools “have become easier to put in people’s workflow,” the adoption has increased.
AI in the Webflow Platform
I mentioned at the beginning of this article that AI appears to be integrated throughout Webflow’s platform — which includes design tools, CMS, build tools, hosting, security and much more. The company now describes itself as a “website experience platform” and the homepage states that AI is “embedded across the Webflow platform.”
I asked Leinwand how its internal engineering use of AI has fed into the platform it offers customers?
“So the product that our customers use has AI sort of threaded throughout,” he replied. “We don’t see AI as a feature, but we see it as a thread that works throughout the entire product.”
“We don’t see AI as a feature, but we see it as a thread that works throughout the entire [Webflow] product.”
– Leinwand
As a couple of examples, he says they offer an AI site builder — “where you can describe what you want your site to look like, and we’ll generate the pages on the site for you” — and that you can also generate items in the CMS product using AI.
Webflow began as a simple website builder back in 2012, but these days it’s a full-on platform with lots of bells and whistles. More recent additions include optimization functionality (e.g., A/B testing) and localization.
“There’s some AI in there too,” Leinwand added, regarding optimization. “So you can automatically generate various variants of a landing page, or buttons, or call to actions — using AI.”
In short, AI is everywhere in Webflow’s products. But based on what Leinwand has told me, the company also practices what it preaches by extensively using AI tools internally.
Originally published at The New Stack: https://thenewstack.io/how-webflow-got-89-of-its-engineers-to-use-ai-daily/