How To Build Your AI and More From the Generative AI Summit
SPECIAL CHAPTER: Notes from the Generative AI Summit in London
As promised, here I am with a special feature on the Generative AI Summit in London, where I shared my thoughts on how to create and implement an AI-assisted chatbot within a highly regulated corporate environment. I won't bore you with the whole presentation - I'll just focus on the DIY checklist. I will also report on the mood of the room at the end of these engaging two days, as consensus formed around key topics.
Featured Insight: How To Build Your AI
1. Goal Setting. The first step is to define what the assistant is expected to achieve. Are you trying to automate customer service, generate leads, offer personalized product recommendations, or a mix of these? A clear objective will guide the overall development process and provide a metric of success. My advice is to start small from a palpable business pain point before scaling up to additional use cases.
2. Stakeholder Engagement. Gather requirements from all relevant stakeholders such as sales, marketing, IT, and customer service teams. They can provide important insights about the needs and pain points of users that the bot should address. Also, engage corporate leadership and ensure their support, as their buy-in will be critical for resource allocation and overall project success. Make sure to set the expectations clearly. The best way to put it is that even in the case of failure, you'll have gathered valuable customer insights (e.g. what they ask the assistant to do) which you can use in future iterations. If leadership is sceptical, suggest starting with exposing the assistant only to a sample of your visitors as opposed to all of them.
3. User Research. Understand your audience and their preferences. Are they comfortable with AI interactions? What platforms do they use the most? What are their expectations? This step is crucial to ensure the bot can cater to user needs effectively. Make sure to be upfront with your audience. Tell them it's a digital assistant. Don't try to fool them. It's unethical and it doesn't work.
4. Technology and Vendor Selection. Decide on the technology stack to be used. This could involve deciding between building from scratch or using an existing AI platform. Plenty of great providers out there (e.g. Drift) but keep reading till the end to find out what was the mood of the room at the end of the two days of debate.
5. Data Collection and Analysis. AI systems learn from data, so collect and prepare relevant data that the AI can train on. This could include previous customer interactions, sales data, or publicly available datasets. Needless to say, this is, together with the technology and the model selection is the most important phase. Make sure the data you provide is of high quality, in the right tone of voice and clearly represents what best in class looks like so the model will learn accordingly. Also obvious, but repetita iuvant - ensure data privacy and security while handling this data.
6. Design Conversation Flow. Create the conversation flow, i.e., the potential paths a conversation can take. The flow should reflect your brand's voice and tone, while also making it easy for users to reach their goals. Once you had to do this manually with trees etc. Recent technologies apply AI to this phase as well, hence you can have the technology itself unearth paths based on goal-driven prompts you can later refine.
7. AI Training and Development. Use your prepared data to train your AI model. This involves choosing an appropriate AI model, feeding it your data, and adjusting parameters until it performs well. I mentioned Hugging Face in my last article - a comprehensive aggregator of open-source models you can test and train with your own data.
8. Integration with MarTech. Resist the temptation to over-integrate. Focus on the MarTech and ServiceTech you really need to fuel live data and enable workflows - you can always add more connectors as you scale the assistant.
9. Testing and Human QA. Before launching, allow a broad cohort of colleagues to stress test the assistant. This should be done following both scripted and improvised interactions, covering as many use cases and variations as possible. The human feedback will be critical in fine-tuning the model and ensuring better outcomes once live.
10. Deployment. Once the assistant is ready and tested, you can deploy it to the chosen platforms. As hinted when talking about stakeholder management, a prudent approach would suggest launching it only to a sample of your total audience and gradually expanding from there, closely monitoring real-time feedback from internal users and customers.
11. Continuous Learning and Improvement. AI models will continue learning over time, so regularly retrain your model with new data. Also, regularly reassess and improve conversation flows, integrations, etc., based on user feedback and new business needs. This should go hand in hand with ongoing monitoring, as you evaluate the assistant performance against your set success metrics.
As with most MarTech tools, developing an AI-powered assistant is not a one-time event but an ongoing process of learning and improvement. To be ready to adapt your strategy based on evolving business requirements, technological advancements, and user feedback you need to onboard flexible technology solutions which will allow you to pivot and morph as required.
The Mood of the Room
The Generative AI Summit in London featured a jam-packed agenda of industry-leading speakers, with Day 1 focussed on IT and Day 2 on sales and marketing. I'll try to cut through the noise and hype and distil the one key takeaway I brought home: the (near) future belongs to vertical applications.
While the conversation around artificial general intelligence (AGI) is fascinating, as it explores the possibility we might be close to an AI capable to undertake any task just as we humans can, the reality is that we are not quite there yet.
Instead of trying to build AGI models - which are very complicated, require vast amounts of data, computing processing power and generate tons of CO2 - corporates should look at finely trained, vertical AI models that can be exceptional at specific tasks. I think this is what we will be witnessing in the coming months: niche solution providers will flourish as brands explore the possibilities offered by building in-house models, trained on quality data in a secure, private environment. I will keep an eye on this space and circulate case studies as soon as they emerge.