Valet Blog

Everyone can build now: what we learned from 60 conversations with AI enablement leaders

By Michael Ducker·

People tending raised garden beds of app windows inside an office atrium. Some plant seedlings while others prune, water, and stake a tall tree of apps.

Over the last two months, my cofounder Jamu and I spoke with more than 60 engineering leaders who run AI enablement at their companies: CTOs, CISOs, heads of IT, VPs of Engineering. They work at some of the most prestigious companies you know of, public and private, in tech, fintech, law, and even hardware. We talked to big companies and small ones, and to startups in biotech, insurance, and healthcare.

We went in with one question: what happens when people outside your engineering team start building software with AI?

It turns out these people are building plenty. Scientists at one biotech built their own lab notebook software, wired into the rest of their toolchain. A sales team at a fast-growing software company built its own sales pipeline and shipped it to its own cloud account. A finance team at a large fintech wrote its own comp tool. One AI enablement leader led us to the term we’ve now adopted, “citizen developer,” because it’s “a more respectful term than vibe coder.”

It turns out that making a software tool isn’t the hard part anymore. Everything after that is.

Here’s what we heard

  • More than half of the people we met said security and data access is holding them back. The tools work and the models are fine. Nobody has figured out how to let a finance analyst read the data warehouse without handing over the keys. One CISO said it plainly: “Shadow IT automatically extends to Shadow AI.”
  • Half of the teams we met were passing AI-built work around in one-off links that aren’t behind their identity provider: Slack, Google Drive, screenshots, or a shared laptop screen running localhost. Leaders told us they can’t keep up with the rise of one-off artifacts, all potentially containing sensitive data without the proper controls.
  • Half of the leaders came to us worried about quality, and about as many asked who maintains these tools after the builder moves on. One AI enablement lead told us, “The industry is heavily focused on the creation and the 0 to 1. The playbook for the ongoing maintenance has not been democratized yet.” Others shared a concern: how will SOC 2 and other compliance standards cover these apps?
  • The same tool keeps getting built over and over. One leader found four managers who had each built nearly the same thing. A platform lead said insecure apps were one problem, but just as concerning was having “so many duplicative solutions to the same business problems that there was no leverage.”
  • Claude was by far the most popular AI tool, both the Claude desktop app and Claude Code. Where it wasn’t sanctioned, people used it on their personal accounts anyway.

Three stages of adoption

Across our interviews, we saw three stages of adoption.

Laggards still treat AI as a chatbot. People at these companies work in the ChatGPT or Claude web app, copying inputs in and pasting outputs out. Their adoption of new tools is stuck behind slow IT and InfoSec processes. One leader spent seven months getting InfoSec to approve a single cloud environment!

The middle bought into enterprise AI. Engineering teams took off with Cursor and Claude Code. The rest of the company didn’t follow. In this middle group, self-directed people are thriving with AI, while the remaining 70–80% are left with AI directives that are “very generic.” At these companies, whatever people build on their laptops gets built once and then forgotten.

Leading companies treat AI-built software as something with a lifecycle. These companies have deliberately built infrastructure that solves for where AI output lives, who can see it, what data it can touch, and what happens when it breaks. Two-thirds of the leaders we met had already built their own internal app stores, MCP routers, knowledge bases, and other infrastructure on top, so their teams could get the most out of AI.

What the leaders do differently

1. They mandate from the top and teach in person

At the companies furthest ahead, leadership wrote AI into the culture through top-down conversations and performance reviews. The mandate alone wasn’t enough, because, as one engineering leader put it, people “don’t have time.”

What changed things was building together. One fintech leader ran build days where an engineer sat next to a business user and built with them. An e-commerce company runs a monthly AI guild, and about 70 people show up. People come in overwhelmed and leave with new tools and a sense of what to do next.

2. They give data access one governed path

The leading companies don’t make every builder work out data access alone. One of the world’s largest companies opened its data lake to every PM through the AI tools they already use, governed by the table controls it already had in place.

3. They build a consistent path for sharing

The most advanced company we met with was a public fintech company with plenty of capital to spend. Half of its employees (not half of its engineers) have deployed at least one app to its internal app store, which has passed 400 apps! Every app runs behind SSO inside the company’s cloud account and gets a security score. A platform team supports the store and its SLAs, such as two weeks to fix whatever security issues are flagged.

Smaller companies have built their own versions. A 10-person startup we met has an app store with 35 apps, sorted by how much each one got used in the past week. Someone on the engineering team reviews every app before it goes up, so what’s there is useful to the whole company.

4. They separate citizen development from production engineering

One data company keeps its vibe-coded tools in a separate GitHub org, with deliberate friction before anything moves to production. A utility company lets its engineers build freely in a “soft tech” repo, without the product constraints of the “hard tech” repo and its full review. These companies let a thousand flowers bloom, and only the ones that take root move into the supported part of the company.

5. They curate after launch

They maintain what gets built, retire the tools nobody uses, and promote the ones people do. In one engineering leader’s words: “We want to celebrate landings, not launches.”

What it costs to build it yourself

Every leading company developed its own infrastructure with its own engineers. None of them had found a product that met their needs when they needed it. Some platform teams had two people on it, while others had dozens of people whose only job was to support it. A CTO asked us, “Do I need to build an AI hosting platform? Is that our core competency? Probably not.”

Most teams want to deal with this soon. Everyone who gave us a timeframe said it needs solving within the next six months, and most said it was part of their Q1 or 2027 planning. One leader told us, “In Q1 we need to start working on this whole internal hosting. We need to solve this problem for reals.”

What we’re building

Valet is the one-stop way to secure your internal apps and keep them off the public internet. It runs inside your own cloud account as a VPC, behind your VPN, behind your SSO, and hosts and secures whatever your teams build with AI: websites, internal apps, agents, and skills. We built Valet so the safe, governed path is also the easiest one. To everyone who gave us an hour of their time for this research, thank you.

If this looks like your 2027 plan, try Valet or get in touch.

Prior posts