Cien Solon

I've been training people around the world, but I’ve rarely ventured up north in the UK. Next week, I’ll be in Manchester for my first tour with Lauren Mucklow, Founder of What She Said. We’re hosting a Build and Breakfast event with some fantastic female business owners. This hands-on session will focus on building AI agents tailored for your business. Where real work makes a difference. I can’t wait to connect! Get your tickets today, ofc. The link is in the comments! 🍋

The most important AI decision your business makes is to refuse to marry a single language model. I said a version of this on a panel last week and watched half the room write it down, so let me make the full case here. The "best model" race has stopped being the interesting race. In the past week alone, Moonshot released Kimi K3, which lands just behind the top proprietary models on independent rankings while costing roughly half as much to run, and Mira Murati's Thinking Machines shipped Inkling while openly saying it isn't the strongest model available, because it isn't trying to be. These labs have worked out what most buyers haven't, that "competitive at a fraction of the price" beats "best" for almost every real business task. Nobody needs the smartest model on earth to summarise a contract or draft a client report, and paying frontier prices for routine work. Yet businesses still run AI procurement like a wedding, months of evaluation to pick the one, workflows and contracts all shaped around a single lab, and then a cheaper model that's 95% as good ships eight weeks later and switching would somehow cost more than staying wrong. The companies getting significant value treat models the way they treat supplier, they are swappable, negotiable, and answerable to results. The expensive model takes care of the hard problems, the cost-effective one takes the volume, and when something better or cheaper ships, it gets tested against real work by Friday. That posture also does wonders for pricing conversations, because a lab that knows you can leave behaves very differently from one that knows you can't. Marry your standards. Date your models. All the zest, 🍋 Cien

Last Wednesday, Mira Murati released her first AI model and told everyone it isn't the strongest one available, and I think that admission is the most interesting launch of the year. For anyone who hasn't followed her story, Murati was OpenAI's CTO until she left in 2024, founded Thinking Machines Lab, and raised $2 billion before shipping. The model is called Inkling, and it's open weight, which means you can download it, run it where you like, and finetune it around your own data rather than renting intelligence from a lab. Her bet is that businesses care less about the smartest general model than about a model they can make their own, because a firm with fifteen years of process knowledge gets more from an open model trained on how they actually work than from the cleverest black box on the market. I've been arguing a version of this for two years, so of course I agree with her. The model was never the point, and your data along with everything you wrap around it decides whether AI does real work for you or produces impressive nonsense. We built LaunchLemonade model agnostic for exactly this reason, so Inkling is live on the platform alongside the frontier models, for you to try out! Build an agent, swap the model underneath it, and see for yourself how much of the result comes from the model and how much comes from everything else. The frontier labs are betting you'll keep renting, while Murati is betting you'll eventually want to own. What I don't know yet is whether small firms realise that choice is now theirs to make.

Before ChatGPT and Claude, I helped build an AI chatbot used by millions of people. It took somewhere between six and twelve months, a large cross-functional team, and a significant amount of infrastructure. We built the database and knowledge base, mapped customer queries, then trained the system using natural language processing and machine learning. When it launched, it improved the company’s resolution rate by double digits. Today, one person can build the first version of that chatbot in an afternoon. That does not mean the results are automatic. You still need good data, clear use cases, thoughtful design and proper testing. But the cost and complexity of getting started have collapsed. Capabilities that once required a large budget, a specialist team and months of development are now available to almost anyone. #launchlemonade #ArtificialIntelligence #AILeadership #WomenInTech

After almost 3 years of running AI workshops, I can confirm every office contains the same five AI users, and I say this with love because I've been at least three of them. 1. There's the Secret Cyborg, who swears they don't really use AI while producing suspiciously polished reports at suspicious speed, and whose browser history would end the charade in seconds. 2. There's the Doomsday Correspondent, who has never actually opened ChatGPT but forwards every article about AI taking jobs, adding "worrying!!" as analysis. Knows more about what AI will do in 2030 than what it can do on a Tuesday. 3. There's the One Prompt Wonder, who tried it once in 2023, got a mediocre paragraph, declared the whole technology overrated, and has been dining out on that verdict ever since. 4. There's the Maximalist, who runs everything through AI including two-line emails to their own team, and whose messages now read like they were written by a Victorian butler with a marketing qualification. Enthusiasm ten, judgement pending. 5. And there's the one nobody notices, the Delegator, who doesn't talk about AI at all. They just worked out which three tasks it does well, briefed it like they'd brief a person, check everything before it ships, and leave at five. Everyone secretly wants to be come the delegator. The good news I spend most of my time delivering, is that the gap between the other four and the Delegator isn't talent or a technical brain. But it's treating AI like a hire instead of a magic trick, and that's a skill anyone who's ever managed a person already owns. Tag your Secret Cyborg. They'll deny it, which proves it. All the zest, 🍋 Cien

Last year, I got THE mother of all bills from one of our AI providers. We run governed AI agents as a business and governance is what we sell, but we made errors along the way that helped us get here. This is because we prioritised the job to be done before making sure the structure was robust enough to hold it. We see this in many companies trying to become AI-first, where agents are already working but the organisation around it still has to catch up. "The catch-up" is a transformation challenge, and whilst businesses have run digital transformations for decades, one new variable changes the game, the AI agent. That one variable breaks your existing transformation framewoks in three ways: - First, agents fail by continuing. An employee who hits confusion stops and asks, while an agent keeps going - Second, agents remember out dated instructions unless regularly pruned. - Third, agent costs scale like consumption. So how and who should be managing your teams' AI agents? Read the rest of this piece in my newsletter, the link is in the comments. 🍋

This is what my recently used AI models look like today. Two weeks ago, the list looked very different. The GPT-5.6 series had not yet become part of my daily workflow, and I was reaching for a different mix of models. Since then, I have been testing what each one is best at and changing my defaults around three things: speed, output quality and cost. That is increasingly how I think about models. Not in terms of which one is “best”, but which one is right for the task in front of me. A strategic piece of work may justify a slower, more expensive model if the reasoning is stronger but a high-volume workflow may be better served by something faster and cheaper. Image generation, coding and long-document analysis may each need different again. The market is moving so quickly that model preference is becoming less useful than model judgement. Two weeks can be enough to change the stack. I suspect this is where more businesses will end up too. The question will not be, “Which AI model should we standardise on?” It will be, “How do we route each task to the model that gives us the best balance of quality, speed and cost?” That is a more operational way of thinking about AI, and a much more useful one. Has your model mix changed in the past few weeks?

Building a company changes the questions you ask yourself. When we first started LaunchLemonade🍋 , we spent most of our time asking whether we could build something. Now we spend more time asking whether we should. Every feature, partnership, and customer request has a cost that includes engineering time and focus. One of the hardest lessons we've learned is that growth often comes from saying no to the ones that don't move the company closer to its purpose. The last few months have been some of the busiest we've had. We've onboarded new team pilots, signed new partnerships, and spent a lot of time with customers. What's surprised me is that these milestones have made our direction clearer. And it is something we are still learning every week.

I’ve connected Fireflies to LaunchLemonade and built a few agents around the transcript that have saved me 3 hours of work every week. - One sends me a summary every Monday of the meetings I had the week before. - Another looks through the conversations and pulls out ideas that could become LinkedIn posts. - For client meetings, I get a clear action list afterwards. - For stand-ups, the actions are separated out so the team has something practical to work from rather than another transcript sitting in a folder. None of this is particularly complicated, but it saves me from going back through every meeting and trying to remember what was said, what mattered and what needs to happen next. So what I've been finding more useful with AI is not generating more information, but helping me do something with the information I already have. All the zest,

People ask where I work. Sometimes it's my desk, sometimes it's a client's office and sometimes it's a conference. Over the past few months I've spent more time in boardrooms than behind my laptop. Why? Because AI adoption is still a human problem. Every organisation has different systems, different constraints and different ways of working and you don't discover those from behind a screen. You discover them by sitting with people, asking questions and watching how work actually gets done. That's still my favourite part of this job.