No one wants to go to your URL

Everyone is clamoring to build an AI workspace where all work gets done in some new website / app. This is dumb for a couple of reasons.

For starters, no one wants to learn how to prompt your tool and they shouldn’t have to. Good prompts are structured, well thought-out instructions that give the LLM clear goals, boundaries, examples, and rules of thumb to reason about problems. Everyday people suck at prompting because most people don’t think clearly enough about how they get work done. To be clear, this isn’t their fault. It’s unnatural for people to be able to codify how they reason and think about what they do when their internal prompting is actually just second nature. As developers of LLM applications, you’ve invested in unravelling your own internal prompting into prose. So, you can easily trick yourself into thinking that your new tool on a new url with an empty input box feels like you’re offering endless possibilities, but to real people it just feels foreign.

People are also tired of new tools. Every one of your customers is already up to their eyeballs with old tools on old urls because of the B2B vertical SaaS era. They’ve already felt the pain of the tool and info sprawl that is required to get their work done and they’ve had enough. The pendulum is swinging back around. Making a new workspace on a new URL is just adding to the mental tax of using something new.

The other problem is your new AI workspace doesn’t lean into the fact that AI can do incredibly complex stuff with low mental load on the user. Today, developers are spoiled, they know what it feels like to vaguely point their Claude Code/Codex instance at work that needs to be done and get it done. Real everyday workers are eons behind with current tooling. But making them go to a new site and be in the driver’s seat for instructing how that work should be done for them actually hands them a problem not a solution. They’re not developers; they’re not used to that pattern. This is the reason why, despite their immense value, Claude and ChatGPT have not truly proliferated into solving problems for everyday businesses around the world and why there’s room for a better UX in the app layer.

So if building a new tool on a new url is dumb then does all the value accrue to the incumbents who can build new tools on familiar urls? No. Lots of legacy players are finding themselves creating these kinds of solutions out of convenience rather than intention and there are a couple of problems with their approach that open the door for startups.

Interoperability by far is the biggest pitfall for incumbents. Gemini in Gmail, for example, will be constrained in using Gemini models and integrating with Google tools. Even if they do expand they have an incentive to always make their walled garden better by reducing integrations.

Personalization is another fundamental avenue incumbents cannot pursue. They won’t let you alter the AI the way you want because they can’t risk Gemini drafting a racist email for you and so everything is milquetoast AI behind an iron curtain. They have no incentive to help YOU get YOUR work done proactively, instead they have incentive to make a flat thing for everyone to get marginal value out of so they retain on the Google suite.

Finally, the incumbents fall for the classic horseless carriage problem - their tools will inevitably fail because they mimic the old way of doing things. None of these incumbents are AI native, and some of them are so slow they invented transformers and then didn’t make LLMs because of bureaucracy. Expecting them to be quick and nimble is still an insane bet to make when the underlying tech is changing so fast.

So you have to beat them. You have to shoulder the burden of the prompt from the user by hiding the “prompting” part of LLMs in work that already closely resembles it in familiar urls. Think emails, slack messages, SMS, documents, PRDs, tickets etc. If you do so correctly AI is proactive and magical. No one should have to visit your new URL - all config, setup and capabilities should live in the familiar URL and be improved from within the familiar tool itself.

There are a couple of startups following these principles phenomenally like WisprFlow and Viktor and we’re building Supafax in the same vein. There’s no new email client to go to, all config and rules are stored in your inbox, Supafax works proactively but if u want to delegate tasks just email it (like you would fwd work to a colleague) and it does your work, that’s it. Supafax lives entirely on Gmail and Outlook and uses their existing infra to avoid us storing any of your email content. This is how the next generation of app layer companies should be built - no new URLs.