Generative AI: Context and Future for existing Companies

The digital transformation within the business world has undoubtedly been accelerated by artificial intelligence (AI) and one of the most influential and promising subcategories of AI is generative AI. The three key developments I see for existing companies applying generative AI:

• The "moat" for companies lies in internal knowledge and data, with which they can best apply their 'in-house AI' to business activities. Consider a specialized notary office that has been creating shareholder agreements for 3 generations: based on their internal database, they can better train their 'in-house AI' and thus generate better agreements for each unique company structure.
• Service becomes instant and 24/7: all 'first-line' customer service work will be replaced by AI. Only for more complex matters will a second-line employee (supported by AI) take on a customer case.
• The revolutionary application of GenAI mainly lies in combining "multiple dimensions" such as a combination of text and audio, where AI-generated speech is based on AI-generated texts. Imagine a tireless sales employee who can conduct dozens of personalized conversations simultaneously over the phone.

But before we cover each in turn, let's kick off with a brief overview of different AI approaches and how they differ from generative AI like ChatGPT, after which we'll delve into its application and impact on existing business in particular.

Artificial Intelligence: An Overview of Approaches

AI is a broad term that refers to machines or programs capable of performing tasks that typically require human intelligence. Any computer program can be classified as AI (spreadsheet software helps us calculate more easily), but when we talk about AI today, we roughly mean the following categories: machine learning (ML), deep learning (DL) with neural networks, and generative AI (GenAI).

Machine Learning (ML)

Machine Learning is a form of AI where a system learns and improves without being explicitly programmed. By feeding the algorithm with a lot of existing data (pictures of peppers at different ripeness levels), it eventually learns to recognize new inputs (an image of a pepper in a greenhouse) and classify it (ripe / unripe / rotten).

Deep Learning (DL) and Neural Networks

Deep learning is an evolution of ML and uses artificial neural networks to recognize and interpret complex patterns. It simulates the structure of neurons in the brain (hence the name), made possible by better algorithms and cheaper computing power. Training with multiple layers of deep structures makes the (maximum) reliability of the models much higher than 'normal' machine learning. The operation (first training on existing data, then applying to new data) is the same.

Generative AI (GenAI)

Generative AI, however, goes a step further. It is a form of AI that can create output from scratch: from texts and images to music and speech. Examples of generative AI models are OpenAI's GPT-3 (text) and Midjourney (image). It uses the techniques of ML and DL (deep neural networks trained on a lot of data), but the result is an AI model that generates new text and image by itself.

How Does Generative AI Work?

It's good to understand in broad terms how GenAI models produce output. In short, three components come together:

1. The model (both commercial models like ChatGPT & Claude and open-source models like Llama, Alpaca, and Mosaic). The "base model" is trained on many billions of sources (websites, textbooks, photos, videos, podcasts). Each model has subtle differences, but the foundation boils down to "the worldwide web". You can retrain parts of the model, but is a topic for another day.
2. The context: a model can fed with specific context, such as internal documents (manuals, reports, emails). The AI then generates an answer based on the broad knowledge of the "base model" while taking into account specific knowledge from the internal documents.
3. The prompt (or 'question'): the AI generates an answer based on the question you ask ("What steps should I follow to return a product?"), and variations in the instructions greatly influence the answer. The instructions affect the tone and length, but can also contain concrete descriptions about the role of the AI and what to do if it cannot generate a good answer ("You are a helpful customer service employee. Can you tell me how to return for free? If you don't know the answer, respond with "Sorry, I don't know"). This is just the tip of the ice-berg: prompt-engineering is another topic we will revisit.

As expected, a whole new generation of GenAI startups has exploded on the scene. But even more surprisingly: existing tech-giants are rushing out their GenAI-applications as well: Adobe launched Firefly, Salesforce launched Einstein, Microsoft is integrating OpenAI in every tool they offer... This is not a revolution that will see them disrupted, at least not right away.

Where is the moat for existing companies?

Maybe you are not a tech-giant, but you are exploring AI technologies and tools. Where should you focus your efforts when it comes to exploring the vast world of generative AI? Picking the right (open source) model? Training it? Crafting the perfect prompt?

For existing companies, I believe the most significant added value is not so much in the choice or application of the model (there are already tools to easily exchange them for each other), but in:

• Internal knowledge and data: on which the standard models are not trained as they had no access to it. This internal data can be provided as context to create an internal AI-tooling that provides much better answers.
• Prompt-tuning: testing and 'tuning' different (rounds of) questions to the model to get the best answer. Less defensible and less valuable than internal knowledge and data, but certainly a competitive advantage.

We will now focus on the application of AI and leveraging that internal knowledge as context in customer contact via chat and phone.

Applying Context: Generative AI in Customer Service

The combination of a broad base model and the ability to provide it with a specific context makes generative AI models ideal for customer service applications. They can answer simple questions, interpret complex requests, and even conduct chat conversations with customers without having to define entire conversation trees in advance (as with most traditional chatbots).

A live example of "applied AI" in customer service is the Hauster chatbot. Based on ChatGPT-3.5 and trained on the context of our "Frequently Asked Questions", I was able to create a tailor-made, multi-lingual bot in <30 minutes using Chatbase. Feel free to try it out:

Hauster Chatbot

Although this first iteration is already able to answer simple customer questions (specific to the Hauster context), in future generations it's possible to train the chatbot so that more complex questions on home improvement projects can be answered ("I have existing insulation, can I apply a normal vapor barrier when adding insulation?"). That's a matter of months, not years.

Revolutionary Impact: Combination of GenAI "Dimensions"

All the "Big Tech" companies are launching and implementing their own models in their tooling. An AI co-pilot in many software packages (from Outlook to CRM software) will be the standard in the coming year. I don't see AI startups emerging that suddenly overthrow one of the big players by applying AI to an existing business model (AI for e-mail, AI for your CRM, AI for e-commerce product descriptions): the application of AI in existing software is too simple for that.

I'm much more optimistic about the revolutionary combination of GenAI "dimensions", such as text and audio. With AI-generated text and an AI-generated voice, you can have conversations run completely 'naturally' over the phone by a non-existent person, for example in this example from Air.ai/Tesla:

Again: 'we've seen this before' (e.g., Google Duplex), but broad application of this Google technique is not yet there (and only in English). Or take my own experimentation using Stable Diffusion (for the images), ElevenLabs (for the voice of Joe Rogan) and HeyGen (to tie it all together):

Two Dimensions (AI image + AI Animation):

Three Dimensions (AI image + AI Voice + AI Animation)

Actually, the last video was based on ChatGPT text, so make that 3.5 dimensions....

Conclusion

Generative AI is transforming the way companies operate and communicate with their customers. The implementation of this technology, whether it's commercial or open-source models, offers significant benefits including improved customer service, operational efficiency, and cost savings. The application of GenAI in all possible software will quickly become the new standard (months, not years), and existing tech companies are taking the lead.

Nevertheless, correctly applying internal data to make the best AI model and the combination of "AI dimensions" is an important route to success, for both existing organizations and start-ups alike.