Building Your AI Model for Communications

If you believe in self-reported survey data, you can agree most corporate communications and PR agency people are already using AI for efficiency gains in content creation and story pitching. Other uses like targeting and social optimization are less advanced but growing. When working with your resident data scientist to improve in each area, as well as others like corporate strategy and brand alignment, it’s essential to know which machine learning models are being used on your data.

Continuous improvement in each communications discipline will rely on your understanding of AI Model Selection. Meaning, you should be able to ask and answer which approach you are using to improve prediction accuracy and recommendation (text/image generation) accuracy.

Case Study: After experimenting with many different Machine Learning models (previous blog post), Microsoft asked itself a big question. Is it more important to have large volumes of quality data for an average AI model to "learn" on, or is it more important to have a superior highly tuned AI model that was done better from the start, even if it took longer to make and cost more to produce?

Asked another way, is it better to run large volumes of data through a decent learning model over and over again? Or is it more important to keep tweaking the model using lower volumes of data, allowing the model to improve itself before it is considered ready for use by the communications team? Microsoft concluded that it was far better to have large volumes of great data and run it many times through the good-but-not-great model it had created. The result? A Large Language Model.

By running data through an AI model many times it turned out to be far better at prediction and generation. With that bit of knowledge, a productive discussion with a data scientist can lead to an AI model for the communications function that truly differentiates you and your department/agency from others. You don't have to perfect your AI communications model before launching it. Work with your data scientist to select the right AI to train on your data. With the right amount of data over time, your chances of success are much higher.

Engineering point: If it costs you $100 to build a product that solves 80% of your problem, and it costs you $1,000 to build a product that solves 100% of your problem, spend the $100. By the time you spend $1,000 to perfectly solve the business problem completely, the problem you started trying to solve will have changed.