Cien Solon

Cien SolonSep 14, 2026

For five months, I have been trying to get an AI to make money from cat horoscopes. 🙀 Before you judge me, I do have a legitimate reason. Most AI benchmarks tell us how well a model performs at things like maths, coding, reasoning and knowledge and all useful obviously, but I wanted to know what happens when you give an AI a real-world goal where the instructions are messy. So I gave it a business - a cat horoscope company and make money from it. A completely normal thing to spend five months testing, right? I started with Claude 4.6, upgraded to 4.8 and eventually tried Fable 5.1. The models could do a surprising amount including the planning that was generally sensible, the output looked convincing and, from a distance, it could almost look like the AI knew exactly what it was doing. And five months later? Revenue: £0. The cats remain financially irresponsible. I have now started a similar experiment with GPT-6 Astra and, so far, the planning and collateral are materially better, but I've yet to find out if it'll make money. So, for entirely scientific reasons, the cat horoscope business continues.

Cien SolonSep 16, 2026

What does “approved” mean with AI tooling? Approved what exactly? Approved for employees to use? Approved to access company data? Approved to connect to SharePoint, HubSpot or Salesforce and to do what? Update a record, send an email or take an action without someone checking first? Those are very different things, but I still see them bundled together under the same label. Approving ChatGPT or Claude for employees to use is not the same as approving a workflow that lets it read client data, update a CRM record or send something externally. There is a big difference between an AI that can read information and one that can change it. There is another big difference between drafting something and actually sending it, and again between making a recommendation and making the decision itself. So when someone tells me an AI tool has been approved, my next question is usually....approved to do what? Because the important bit is not just whether the business allows the tool. It is what the business has allowed the tool to do.

Cien SolonSep 15, 2026

One of the questions I get a lot is, "If companies already have ChatGPT, Claude or Copilot, why do they need another AI platform?" And I ask myself this daily. Companies that I speak to are not short of AI tools anymore, and in fact quite the opposite. They might already have Copilot across Microsoft, people using ChatGPT and Claude individually, a few teams experimenting with agents and another AI product sitting somewhere inside Salesforce. Some also want to run their own private models, to have better data controls or they simply do not want their AI strategy tied to one provider. That is one of the reasons why we built LaunchLemonade as a harness rather than around a single model. A business can use the frontier models it wants, but it can also connect its own model into the same environment and build workflows around it. When you let AI sit inside business processes, managing what those models can access and how you keep visibility over what is happening needs to be a priority. Businesses should not have to rebuild their AI infrastructure every time they want to change models and they should not have to give up control of their own model in order to use the rest of the stack either. 😉