[taaron]

Valid point. At its core, our AI is indeed a LLM that is prompted to provide an output. A lot of the work however is in (1) prompting it in a way that allows it to actually understand the user instructions and current state of the workflow and (2) allow it to reliably output a response that acts as a set of instructions on what needs to be done within the platform e.g. add this, remove this, change this question text to this, write this code etc.

One example of what we had to do to achieve this was to develop an "intermediary language" defines how the current state of the workflow is represented to the AI and how the AI responds back - this needed to capture enough detail about the workflow without overwhelming it with too much context. We also developed techniques for structuring the prompting, with the process of building a workflow actually split into 3 stages: a pre-build planning stage, a build stage where the overall structure of the workflow is set, and then a build node stage where each individual node its configured. There is a bunch of other techniques we developed to get LLMs to be able to do what they current do, but these are just some examples of how it's a bit more than just a "You're a business consultant" prompt.

One thing I'd encourage people to do is test these co-pilots head-to-head on the same prompt. If you were to ask Zapier or Make to "build me a process for triaging customer complaints", I'd expect them to not get very far, perhaps an outline of some apps you could connect together to achieve it. If you asked our AI this same request, it would be able to deliver a complete workflow with fully configured forms, tables, branching logic, tasks etc