AI Agents are everywhere! If you followed OpenClaw news this week you can get a feel for what is possible. JuiceAI is a CoPilot agent (using large language models under the hood) trained on data we specified, it is designed to be able to answer the vast majority of our support tickets. This means that instead of relying on a general-purpose AI trained on broad public knowledge, we have pointed our agent to a specific dataset that we outlined and control. See ⁠here for where this started in 2024.
We think all support, including what our clients provide, will go in this direction!

We ended up training 2 models, one for our support team and one for our development team. For support we used ChatGPT 5.2 reasoning but for development we went with Claude Opus 4.5, at least those are the models we are using today. However, the team is meeting weekly as things are moving fast and we're able to change the model easily if something better comes along. ChatGPT was just announced in late 2023 just before I posted the 2024 blog above. It is also important to use low cost models like Kimmy K2.5 where performance is not needed.

Training an LLM to understand your data is very similar to training a new hire. When you onboard a new hire think of what your process looks like and how long it takes a new hire to get up to speed, for us it is typically at least 90 days. You will hear the term Agentic AI or AI agents and these are best described as role specific agents. You need to be able to let them consume all that "new hire" data, while being careful that you don't share anything important with your competitors. You probably will have different training material for each role in your company, but is it digital? This data can't be just in someone's head, it needs to be documented. (See above blog on how to start). We had to use a different approach for our 2 models, Juice essentially just has support and development, no sales:

1. For support, we review when our AI does not answer a question correctly and go back to our documentation to see what is missing. We are looking for you guys (our clients) to help with this feedback! We also had to remove anything that we considered "intellectual property" or sensitive data. We are only answering tickets from clients' domains, so if you email us from a home email you won't get an AI response.

2. For development, we needed to share all of our previous code, and it had to be translated from raw SQL into structured, formatted documentation. We used AI prompt tooling to iterate thousands of SQL files and relevant industry-specific external documentation. While there was some upfront effort to define the process, it is now fully repeatable and can generate thousands of usable training files in minutes.

For those attending the EUG (16th - 19th Feb in Salt Lake City) we will be chatting on tools that can answer the phone and give users a great automated experience, while referencing that LLM that has all the answers. Click here if you're not registered for the EUG yet.

We will make this available in our chat engine soon - you will get one set of limited answers when logged out and access to the full support library once you log in. We are working with various manufacturers to see how we can add manu and model data to our LLM, and we are also looking at how predictive data will play into our agents. Our long-term goal is an LLM that works for the industry and our clients.

We are excited to launch this and will report on what % of our level 1 tickets it can answer successfully. As you implement Agentic AI ensure you are using our tools and reporting to track how effective it is like using our new ID45. (beta testers needed)

For more on OpenClaw I like this post, to really see what it can do, like creating a CRM on the fly, watch this All In Podcast from minute 45.54. (a little scary also)

Look out for Juice ticket responses where JuiceAI answers, coming this week.

You can't automate your business if everything is not integrated. The future is AI, get ready for AGI.