Subscribe to our newsletter to receive business tips, learn about new funding programmes, join upcoming events, take e-learning courses, and more.
Take the first step to successfully starting and growing your business.
Join for freePosted: Mon 15th Jun 2026
As small businesses integrate AI tools and agents into products and processes, many are moving too quickly to fully understand the risks involved.
In this practical session, Olugbenga Latinwo explores how you can build AI products more responsibly in the early stages – without slowing your innovation.
Drawing insights from building Culture Illustro at Craftdash Limited, his session covers practical AI guardrails, iterative experimentation and scalable governance-by-design – all tailored for start-ups and SMEs.
How to integrate practical AI guardrails early without slowing innovation
How iterative experimentation can help reduce the risks that AI products present
Key considerations for safely deploying AI tools and AI agents into products and processes
Olugbenga is a business growth strategist, an AI product builder and the co-founder of Craftdash Limited, where he leads the development of Culture Illustro, a cultural AI illustration platform.
He has led digital innovation and transformation projects across the UK and Africa, including serving as a lead implementing consultant for the UNDP-EU Nigeria Jubilee Fellows Programme (NJFP) Talent Management Services initiative in Nigeria.
Olugbenga recently took part in The Alan Turing Institute's Practitioners Hub as an Expert-in-Residence, contributing practical perspectives on responsible AI and governance-by-design within start-ups.
Watch more expert webinars
Access a growing collection of expert-led webinars covering marketing, sales, finance, growth and more – ready whenever you are.
Lightly edited for clarity.
Ryan: Good afternoon, and welcome to today's Lunch and Learn. My name is Ryan, and I will be your host today. For those of you attending Lunch and Learn for the first time, Enterprise Nation is a vibrant community platform for startups and small businesses. Today, I'm really pleased to introduce Olugbenga Latinwo, who is a business strategist.
In this session, Olugbenga will explore how you can build AI products more responsibly in the early stages without slowing down your innovation. If you've got any questions throughout, just pop them in the chat, and we'll do our best to answer them at the end. The webinar is recorded, and a recording will be sent out later today, so keep an eye out for that in your inbox. On that note, I will hand over to you, Olugbenga.
Olugbenga: Good day to everyone – good afternoon, good evening – wherever you are on the planet. My name is Olugbenga, and I'm the co-founder at Craft Dash Limited, where we build cultural AI tools. Over the next few minutes, I'll be sharing practical lessons I've learned as a founder and product lead in building AI products -- not just from theory, but from experience along the way.
We're not just looking at compliance checklists. We are also looking at things from the practitioner's point of view. So I want to start with a question I'm posing to everyone here: in 100 years, what will AI remember about us, about our businesses, about our culture, about our communities? And most importantly – who gets to decide what?
These questions became very real for us at Craft Dash while we were building Culture Illustro, and they taught us a lot of lessons along the way. Today, we are living through extraordinary moments. Every conversation is around AI. AI has become so accessible that it is being used for different things, and everyone is talking about it. But what we fail to discuss – especially as a founder, a product lead, a product manager, or someone wanting to build an AI tool – is the question of trust. Very few founders I've interacted with over the past few years have deliberately designed for trust. It's just about the AI tool and the output, and that has been our biggest lesson so far.
So, a bit of background: Culture Illustro is a cultural AI tool that builds trust into the images we see online and makes sure underrepresented and unique cultures around the world have a voice within the AI space. When we started building it, we concentrated more on better prompts –we wanted the outcomes and outputs to look good, we wanted generated images that looked great. For a while, that seemed to be working, until we realised a lot of things we were doing quite wrong.
We discovered that an image can be technically correct and still be culturally wrong. One example from our sample datasets involved a prompt to generate a particular cultural dance from the Yoruba ethnic group in Nigeria. Through our evaluations and tests, we discovered a whole lot of things that were wrong with the images we see online – how they are generated and how trust is missing. The challenge became not just quality, but whether trust, representation, and accountability were built into the product from the design stage.
That's when we stopped looking at Culture Illustro as just an AI image generation tool. It went beyond the appeal of clean images. Now, one of the most valuable ideas came from our participation at the Alan Turing Institute – a six-month engagement where we worked on a lot of case studies and discovered what systems thinking is all about. What it made us realise was that we were concentrating too much on training datasets, the AI model, and the outputs. We forgot about the people, the communities, the users, and the trust we were supposed to build in.
I want to take this opportunity to encourage anyone building AI products to look at where you can get help. It might be from platforms like this one, or other institutes running AI acceleration programmes – it saved us a lot of headache going forward. Systems thinking zoomed out a lot of things for us. Outputs affect people, and people influence trust, and trust ultimately determines adoption. Our breakthrough wasn't improving the model -- it was simply understanding the system around the model.
And when we looked at the system, we discovered a lot of surprising things, which takes me to the next lesson: stakeholders we nearly missed. We discovered hidden stakeholders when we started mapping them across our own AI life cycle. We needed to go back to the drawing board to see what we were missing, which stakeholders we were missing, and then we started mapping them across. We discovered that within our design, we hadn't properly concentrated on the represented communities, communities representing these cultures, which is one of the major sources of truth we could get.
Systems thinking revealed all of this, and we were able to start working on how to build trust into the product. When we talk about governance and accountability, we also have to talk about provenance – being able to trace who is responsible for what, and being able to track things down when something goes wrong. Our next question was: how do we involve the right stakeholders at the right moment?
As a product builder, you need to understand that it's not just about the AI life cycle. We needed to know what happens at every stage, from a cultural briefing down to deployment – which stakeholders are present at every point, who reviews, and who decides what. Trust is easier to build early than to repair later. I always tell founders: don't let trust be an afterthought. Let it be built into the design, and the same goes for inclusivity.
We needed a way to experiment with what we'd learned and build forward, which brings me to lesson four. One of the most valuable concepts we discovered is the Amazon two-way door principle. To be honest, we didn't know that was what we were doing – but we started working on things in smaller increments, making sure we were testing as we went, instead of taking the one-way door where damage had already been done. For example, publishing culturally inaccurate content and putting it online.
We are all familiar with the EU AI Act that is coming into implementation in August, and we know the impact and the advantages of this for every founder and everyone building or integrating AI. The two-way door process has helped us make reversible decisions. For instance, if we're working with a particular dataset and find it isn't culturally accurate enough, we go back to the drawing board and look at our confidence level to assess whether it's plausible and accurately in sync with the culture it's representing.
It is much harder going through the one-way door, because the damage will have been done. Trust and governance should be embedded in the heart of your design, built into your product pipeline from the start. People often think guardrails slow down innovation –I actually disagree. I used to think AI governance, and building trust, accountability, and traceability into the process of building a product, was only for bigger companies, only for enterprises. Later, I realised that even if you are adopting AI into your workflow as a business owner, integrating different tools, guardrails are the lifeline of what you are doing.
Guardrails allow you to move quickly without losing direction. In Culture Illustro, for example, we have cultural briefings, human reviews, and feedback loops from communities and cultural experts. At every point, we have a mechanism for capturing that feedback, and even when we release an output, we still gather reviews so we can improve on what we've already experimented with. Without guardrails, AI scales mistakes. The large images we see online today that we complain about –where we can't trace the source, where they look beautiful but are culturally inaccurate – exist precisely because most of them are trained on synthetic data, and there is no accountability for where they came from. Guardrails do not reduce innovation. This has become very important for us.
So as we move into the agentic era, the world is changing fast, and most AI tools have helped us in one way or another. But we have now reached a stage where it's not only about the outputs – it's about what acts on those outputs. That's where AI agents come in. We can generate a good answer, but we should also be asking: should AI be allowed to take action? Will you allow your AI tool to execute a bank transaction? Will you allow an AI agent to answer customer queries? Are you confident that the tool would do exactly what it's meant to do?
That's why we talk about human-in-the-loop approaches. But AI agents now execute without humans in the loop, and that is precisely why governance matters even more in the agentic age. This is also where many people still misunderstand what Safe AI means. It's not about slowing down innovation or development. Safe AI literally means trusted output. In our case, are the images you see online trusted? And it goes deeper than just saying "we're using the image for this purpose." When AI is trained on unverified images, and those images drift, the output drifts too. That content goes into classrooms, students use it, and we can easily scale what is not true over the next few years until it becomes the accepted truth in society. That's the world we live in, if nothing is done to build pipelines of trust, accountability, traceability, and provenance.
So how do you, as a founder or business owner, remember all of this? I've tried to come up with a simple framework you can see on screen: TRUST: Task, Risk, Users, Supervision, and Training. I use this with my team to make sure we're on track. What problem are we solving? What risk is involved? In our case, we're helping underrepresented cultures be accurately represented on digital platforms and closing the gap between the use of synthetic data for creative purposes. What could go wrong, and who is affected? Referring back to the AI life cycle and stakeholder mapping -- that's why you need to understand who your stakeholders are at every point, so when something goes wrong, you can track both the users and stakeholders affected.
This also covers how the system will improve, which is where evaluation comes in. Feedback loops should be built into products, and most times they should be built with the stakeholders involved at that point. Because what we're building doesn't have right or wrong datasets – we use confidence levels to determine how plausible, how authentic, and how trustworthy an image is. These five questions have really helped to catch many issues before they unfold.
If I were starting as a founder or SME owner adopting AI tomorrow, there are a few things I would do. First, map stakeholders and identify high-risk workflows. Second, run small pilots – working in smaller steps, using the two-way door approach – so you can revert quickly if something isn't working, without burning too many resources. As a founder, you have a limited runway. Third, build feedback loops –I've already mentioned this throughout. Fourth, implement guardrails. Guardrails are not there to limit your innovation – they're there to help you go faster. Once your guardrails are in and properly integrated, you have a framework to work within, and you can innovate more freely within it. And fifth, learn and improve. Without that, once you have a culturally incorrect image and AI scales it, it becomes the accepted truth in public, in content, and it goes on and on.
That brings me back to the first question I asked. In 100 years, AI will remember what we choose to teach it today – especially through your datasets. So let's deal with AI by design, not by accident. Build trust before you scale, through our datasets, our workflows, our products, and our decisions. The future of AI isn't being decided in 100 years' time – it's being decided now, except if we put accountability, trust, provenance, and traceability at the heart of it. The focus should not be only on the model, but on the systems built around the product. I'll stop there, and I hope I haven't overrun my time – Ryan, over to you for questions.
Ryan: Thank you so much for that. You were bang on time, and that was really interesting and helpful. So, a couple of questions. Someone is asking: What is the EU AI Act, and what changes to regulations are coming in August?
Olugbenga: The EU AI Act is a framework – and the part I find most valuable is its risk categories, which classify AI into different risk levels. There is low risk and high risk. For example, certain uses of IP cameras to identify people are not permitted. Part of why these things are put in place is to provide a framework, just as GDPR did for data privacy. Other parts of the world are also bringing in their own frameworks, but the EU AI Act is taking the biggest step forward to make sure we build AI more responsibly.
I encourage each founder to go through it and study it carefully. But the most important thing is to know where your risk lies. If you understand your risk, you'll be able to build better and understand the consequences of your decisions. I'll give some context here: I spoke with a founder in London who had built a fantastic agentic AI platform that could handle customer queries and support project teams. The first question I had to ask was: Do you know the markets that will absorb this product? He couldn't answer that. Was it built for the UK market or the EU market? Was he aware of the EU AI Act? If so, how had he built governance and that framework into his product? That product was dead on arrival – he needed to go back to the drawing board and start over. That's why guardrails are not there to slow you down. Guardrails are there to speed you up and put the right things in perspective from the beginning.
Ryan: So interesting and so important to have those guardrails right from the beginning. Really helpful, thank you, Olugbenga. Another question: a lot of SMEs are feeling pressure at the moment to adopt AI quite quickly. What would you say are the biggest mistakes you see companies making when they're rushing to adopt AI in their business?
Olugbenga: I think one of the greatest mistakes, and this is backed up by research from MIT as well, is that founders and business owners don't understand their own workflow. When you're building a product, you need to understand the need. What are your needs? What's the problem? Define the problem you want to solve, understand it, and then go ahead to identify the risks involved in adopting AI. Even before you get that far, you need to be defining your problem and knowing your gaps – carrying out a gap analysis of where you are. This is where we are; we need x, y, z products to cut costs here or enhance productivity there.
What are the needs of the people in that department? What are their skill sets? You really need to drill down, and that's where your own feedback loops and internal interviews can start. What do we need? Interview your employees. What do we need? What's their skills level? That's how you determine what works for you. What we see most SMEs do today is adopt AI tools – saying "I'm going for ChatGPT, I'm going for Copilot" – and then finding out that different generative AI models and LLMs have different strengths, and you have different needs. You need to match your needs to the solutions you want to adopt. Try as much as possible to understand your workflow: how does your company operate, what are your operations like, what do you do, and how do you do it?
Ryan: Amazing, brilliant, thank you, Olugbenga. That was really helpful and really interesting. Some great comments of appreciation in the chat, too. So, a big thank you to everyone for joining, and a big thank you to Olugbenga -- that was a really great session. I've popped your LinkedIn and Enterprise Nation profiles in the chat, so I'd encourage anyone on the call to connect with you afterwards if they have further questions or want to learn more. Thanks again, everyone.
Olugbenga: Thank you, Ryan. Thank you, everyone, for listening – it's been a pleasure. You can connect with me on LinkedIn, and I do encourage you to check out Enterprise Nation as well – it's a great platform. Thank you very much, and have a wonderful day.
Ryan: Amazing. Thanks, Olugbenga. Thank you, everyone. See you all later.
Unlock more on-demand sessions designed to help you sharpen your skills, grow your business and stay one step ahead. Find more Lunch and Learn webinars
