New Gen Enterprise AI Interfaces

vansh

2025/09/20

Lately I have been heads down on imagining and developing the LLM interface should look like for enterprises. From talking to friends across industries and interacting with executives, I have gathered fair bit of feedback.

First, the observations:

  1. Given how easy it is to write software, enterprises are looking for custom solutions. Interestingly the non-technical teams are the most demanding in terms of UI, and sometimes under estimate what the model can accomplish.
  2. A custom URL for your customer on your platform is quickly becoming the bare minimum, but we maybe headed towards custom UIs; if not for individual companies, at least for individual industries.
  3. Chat is good, but it feels like a small ripple in time that will be quickly superseded with more natural interfaces that allow for muscle memory and to build up expertise.

Even if we assume that custom interfaces will eventually replace all UI, there will be a bridge moment that will be the default in the absence of a specialized solution. Esp. because the current nomenclature and design is guided by a mix of hype, old school ML lingo and a promise of AGI.

Examples:

  1. I know people using the term “agent” in at least 4-5 different ways, and even more in enterprise.
  2. For now, most LLMs are still only doing reads, but once they start doing deterministic write operations, things will rapidly change. Grok attempted to tackle this with “scheduled tasks” but it didn’t really stick.

My proposal: Bring back the old school windows file explorer and map everything to that, for now.

  1. Chats: These become some form of .txt or .doc files. You can open a chat, which in turn will have different styles - nested, forked, brain map, etc - but the idea is that you wouldn’t default to a chat.
  2. Processes/Tasks: If an LLM is accomplishing a task for you, you now have an .exe. Build it, write it, test it, but once it’s ready, save it as a .exe type. The trick is to distinguish between adhoc conversations with an LLM vs defined tasks.

The final UI should look like the classic windows file explorer. It’s something that’s been around forever, and until we figure out all the new emergent capabilities of the models, it works pretty well.