There are a lot of marketing tools, and it seems like every new tool is powered by AI.
With the increased availability of commoditized AI platforms over the last two years, it is easier than ever to build a product powered by AI. However, there is still a big variance in the sophistication of the products popping up in the market.
Since we are an AI-focused company, we are going to focus on marketing tools built on top of the new AI Language Generation platforms like GPT-3 and Cohere. See our previous post for a broader overview on AI-powered Marketing Technologies moving towards 2022.
We look forward to Scott Brinker’s Marketing Technology Landscape Infographic every year because it feels absurd that there could be a viable market for thousands of marketing tools and hundreds of new ones each year. Here is the first one from 2011 and the most recent one from 2020.
Some tools are built for you: the sophisticated, highly educated, enterprise, specialized marketing expert who can rattle off messaging around software defined security architecture or the latest products from your financial derivatives firm, and some are not.
Price doesn’t always tell the whole story, because some inexpensive products can solve very real problems in your marketing stack. However, there are some applications where the out-of-the-box solutions are far too general and won’t cut it for your specialized messaging.
AI relies heavily on examples
AI designed to help you chat more efficiently with your customers relies on examples of how a successful customer service agent interacts with customers.
An algorithm that tries to predict your ideal customer profile likely needs examples or information about your current customers.
The quality and specificity of the examples fed into whatever AI you are using will affect the relevance of the output and the product experience as a whole. Most AI language generation platforms (like GPT-3 or Cohere) try to predict the most likely word to come next over and over to generate a block of text. The best way to interact with these platforms is to speak to it like a human, because it was trained on examples of text generated by humans.
Example 1: No Examples
Let’s say we are writing a blog article about password managers for a cybersecurity company and we wanted to use an AI tool to help us get started. We would send a “prompt” to the AI platform like
Prompt: Write a list of introduction sentences for a cybersecurity blog about password managers.Running this through GPT-3 a few times, we get
AI: The number of cyber attacks is growing constantly
AI: Penetration testers are often asked to test password managers.
AI: A password manager is a tool that can be used to store and manage multiple passwords.Overall, these outputs are pretty good. They are topically consistent with the what we are looking for and 2/3 have the tone of an introductory sentence. They might be a good start.
Example 2: Industry Specific Examples
Let’s see what happens when we use some industry specific examples of introductory sentences from other cybersecurity companies and articles about password managers.
The new prompt looks something like
Prompt: Write a list of introduction sentences for a cybersecurity blog about password managers.
1. Businesses and their employees ultimately need a lot of different online accounts, which means there are a lot of passwords that need to be sorted.
2. In the early days of the Internet, we used to have a couple of passwords for important online properties (email, website, etc.). But as technologies evolved, we are now creating hundreds of passwords, some of which we rarely use more than once or twice.
3. A password is the most crucial element to keeping. your company's data secure--as well as the weakest link when it comes to data security, though only 31% of companies use a password management solution to protect their data.We provide the AI a few examples of introductory sentences. These are introductions used by other cybersecurity companies in real blogs that we have archived. Sending that prompt to GPT-3 a few times, we get the following outputs:
AI: Cybersecurity professionals suggest that the best way to protect yourself from a cyber attach is to use a password manager, which is a software or application that remembers all your passwords and automatically logs you into your accounts.
AI: What is a password manager? A password manager is a software tool that stores your passwords online. The software generates strong passwords for you, stores them in an encrypted database, and even fills them in for you when you’re logging in to a website or app.
AI: A password manager is a single login that is used to access all of your other online accounts, including social media, email, and other financial accounts.These results are more specific, and are closer to the tone that we are looking for. Having relevant examples to feed into the AI is a component of “prompt engineering” where the way you interact with the AI is specifically designed for each use case.
Some AI products use targeted examples, some don’t
At Automata, we are focused on natural language generation applications for marketers to help them create better content, faster. In addition, we want to help writers create content that might be outside of their comfort zone.
By training our AI on content that is directly related to our customer’s industry, we can be sure that it is working with the most relevant and up-to-date examples. For example, if a user works for a Fintech company, the examples we feed our AI come from market leaders in the Fintech space and other content related to Fintech.
We have a database of tens of millions of blog articles written by thousands of companies over the past several years. Every time a customer sends a request through our system, we search our database to find the most relevant articles and send elements of them to the AI platforms as examples. This exhaustive database is what differentiates our approach from the barebones generic approaches used in many “quick and dirty” GPT-3 applications.
Baseline AI Models vs Fine Tuned AI Models
The baseline AI models are what you get “out-of-the-box” from companies like OpenAI, Cohere, and even open source models like GPT-J from EleutherAI. However, many platforms allow users to “fine tune” their models.
Similar to how we looked at the cases with and without examples above, fine tuning allows you to pre-train the AI models on tens or hundreds of thousands of specific examples. Using this approach, we are able to build a custom AI model around specific industries and text formats.
For AI language generation to be a valuable part of a sophisticated marketing stack, it requires a sophisticated strategy. This strategy begins and ends with the data that is made available to the AI platforms. By augmenting the baseline AI models for specific use cases, we are one step closer to bridging the gap between what is expected from high-end marketing professionals and what is offered by out-of-the-box AI language generation platforms on the market today.
