Key takeaways
- Three months after launching Shift AI, our original thesis is holding up: people want AI they can control, not AI that controls their workflow
- We didn't guess at this. Our 2026 AI Insights Report surveyed 1,000 consumers before we built anything, and the findings shaped the entire spec
- A simple on/off toggle sounds basic, but most AI tools still don't give users one
- The Intelligent Omnibox proves AI doesn't need to be everywhere to be useful, it just needs to show up when asked
- Good AI still needs good context, so we built a way to bring your "About You" info in from tools you already use, no re-typing required
- OpenAI retired its Atlas browser this month, less than a year after launch, a signal that speed without fit doesn't last
We've been hearing the same frustrations from Shift users for years. I've sat in on dozens of one-on-one feedback sessions where the same themes kept surfacing: too many tabs, too much switching, tools that promised to help and ended up adding another thing to manage. When AI entered the picture, we didn't assume it would fix that.
So before we pushed a single line of code, we surveyed 1,000 consumers for our 2026 AI Insights Report. Not to justify a feature we'd already decided to build, but to find out if we should build one at all, and what it would need to look like if we did. The findings were clear: 32% of users were already engaging with AI daily, but a growing share were uneasy about it. What happens to their data? Does it act without asking? Can users turn it off?
That tension became our spec. People weren't asking for less AI. They were asking for AI that didn't take the wheel.
Three months after launch, that distinction is the thing I keep coming back to. Not because it makes for a good tagline, but because it's held up against real usage, and against a market that's still figuring this out in public.
What our research told us to build
The Insights Report didn't just tell us people wanted control in the abstract. It told us where to build it in.
51% of hybrid and tech workers told us they wanted customization options for their AI tools, compared to 36% of the general population. That's not a small gap, and it's why Shift AI is shipped as something you opt into and shape, not a fixed feature bolted onto the browser.
58% said AI-generated answers had shaped their opinion without them fully realizing it. This is a transparency problem, and it's why we built context chips into Shift AI's side panel. You can see exactly what the AI knows before it answers, and add or remove context yourself.
And 81% told us they were concerned about how their data was being used. That's the number that sent us to Cloudflare before we ever touched the interface. Every query runs through a privacy layer that separates your identity from your activity, so nothing gets tied back to you. We've written about that architecture in more detail elsewhere. Here, it's enough to say it wasn't an afterthought. It was where we started.
Shift AI is officially live: Built on Cloudflare's privacy-first, secure, and scalable infrastructure
Shift AI is now live, built on Cloudflare's Privacy Proxy to fully anonymize user data, bridging the critical AI trust gap.
Read MoreControl doesn't have to be complicated
Here's what surprised me most: giving people control didn't require anything sophisticated. It required a switch that works.
You can turn Shift AI off entirely from Quick Settings, any time, no follow-up nudges trying to win you back. You can turn off AI Chat but keep the Intelligent Omnibox, or the reverse. Two features, two independent toggles, and no assumption that everyone's ready for AI in their workflow at the same time.
It sounds almost too simple to mention. But most AI tools still don't offer this. They ask you to dig through settings menus, or they don't let you fully disable features at all, just mute the notifications. A toggle that actually does what it says shouldn't be a differentiator. Right now, it still is.
AI that shows up only when asked
The Intelligent Omnibox is probably the clearest example of this philosophy in practice. When you start typing in the address bar, a small on-device model reads whether you're asking a question or looking for a page. If it looks like a question, an "Ask AI" badge appears with a subtle shimmer. Hit Tab, and you're in a conversation. Hit Enter, and you search like you always have.
We actually tested a version where the system decided for you, auto-routing you into AI when it was confident that's what you wanted. It tested well on paper. In practice, it felt wrong. Even when the system guessed correctly, it felt like a decision was being made without you. So we pulled back. The shimmer is a hint, not a choice. You make the call, every time.
That one extra keystroke is the whole point. AI doesn't need to be everywhere in the browser to be useful. It needs to be exactly where you reach for it, and nowhere else.
The context problem, solved in a week
Good AI output depends on good context, and that's usually where AI tools ask the most of you. Re-explaining your role, your preferences, how you like things formatted, every single conversation. It's tedious enough that most people just don't bother, and then wonder why the answers feel generic.
We heard this almost immediately after launch. Within a week of Shift AI going live, we shipped a way to save context once and carry it forward: your "About You" and “AI Chat Memory” details, your preferences, the basics an AI needs to actually be useful to you specifically. And instead of asking people to type all of that out from scratch, we built it so you can import it with a simple copy and paste from whatever AI tool you're already using. No migration, no re-onboarding.
It's a small feature. But it's a good example of what listening looks like in practice: we didn't wait for a roadmap cycle to fix something we heard clearly in week one.
Shift AI just got smarter, here's what's new
Shift AI went live a week ago and we've already improved it, five new features including a first-of-its-kind proactive AI built into your address bar.
Read MoreWhat this week's news confirms
Two weeks ago, OpenAI confirmed it's retiring its Atlas browser, less than a year after launching it, folding its agentic features into ChatGPT Work instead. I want to be careful with how I read this, because it's not a story about one company getting it wrong. It's a signal about the category.
Building a browser around a chatbot is a different bet than building AI into a browser people already use to manage the rest of their lives. Speed to market gets you a launch. It doesn't guarantee fit. And fit is the thing that determines whether a product survives its first year, let alone its third.
We took longer to ship Shift AI than a lot of the market did. That wasn't due to indecision. It was us waiting until we knew what to build, and choosing restraint over velocity. Three months in, with a very public reminder this week of what happens when that order gets reversed, I feel good about the sequence we chose.
The philosophy behind the pace
None of this was inevitable. We could have shipped an AI feature in early 2024 to keep pace with the market. We chose not to, and at the time, that felt like a risk. Waiting always does.
What I keep coming back to is this: we listen first, then we build. Because, it's the only way to build something that actually holds up once people are using it every day, not just trying it once. The research told us what to build, now the people using Shift AI every day are telling us what to build next.
We're not trying to win the AI browser wars by being the loudest or the fastest, but by building the version of browser-native AI that people actually want to keep using. Three months in, that's still the bet, and so far, it's paying off.
If you're curious what shaped this thinking, our full 2026 AI Insights Report is worth a look. And if you want to see what we built because of it, Shift AI is live in the browser today.
