Get the most out of data and AI with a marketing data strategy

I recently started working at Capgemini, having spent the last 6 months at home with my 1.5-year-old child. During this time, virtually all major technology companies launched their own version of generative AI or services related to it. While I was immersed in cleaning the highchair and diaper changing- the Twitter-verse (or should I say “X-verse”) was abuzz with hashtags like #ML, #AI, #GPT and #LLM. For a solution architect like me, the world looks very different from what it was just a year ago.

While the technology landscape hasn’t changed that much in a year, the dawn of generative AI has been a major eye-opener in some respects. The promise of AI in a marketing context is enticing: the ability to generate personalized experiences with speed and scale. Even though the learning curve may seem steep, the reality is that you don’t have to be a data scientist to understand the business implications of AI. Let’s dive a bit deeper and I’ll explain why.

How to train an AI?

A common misconception about AI models is that you always need to have lots of data to use one. Large language models (LLM) like GPT-4 do initially require vast amounts of data and training but that essentially constitutes as product development. An LLM requires human effort to “teach” it to weed out inaccurate and false answers to user prompts. After sufficient training, the model can answer most user queries with decent accuracy. At this stage, it can be deployed and made publicly accessible. These pre-trained models can be implemented to specific business needs in a couple of ways.

The first method is via fine-tuning the model. Instead of using huge computational power and human working hours to train a model from scratch, you take a commercially available LLM and train it further for a specific business application. The model will likely provide decent answers to user prompts right from the start, owing to the model’s extensive pretraining. However, human supervision is required to flag all biases, misconceptions and falsehoods that surface. Over time, the model becomes more attuned to its new environment, and its accuracy improves.

The alternative to supervised training is to use what is called in-context learning (ICL). Instead of humans supervising the model and giving it feedback, the pretrained model compares its output to its context, making predictions based on past information present in where the model operates in. A typical use case is deploying an LLM to a CRM environment, such as is the case with Salesforce’s Einstein GPT. The model looks for past records and uses that to provide more accurate answers to user queries. This method of training is attractive in that it has a greatly reduced need for human effort.

The inescapable truth of data quality

Whichever method you choose to train an LLM, you always run into the same conclusion: data quality is key. Bringing us back to the marketing context: what does good quality data mean? Marketing is unique in that it relies heavily on both external and internal data sources. Marketing deals with customer demographics; interests and preferences; website engagement; product or service usage; contact information; and purchase history. With the variety of data sources, there is also a greater risk of data management issues, such as:

• Duplicate leads
• Contradicting marketing permissions
• Outdated contact information
• False or spam contacts
• Mismatched marketing engagements

Identifying and fixing these issues should be your first plan of action if you plan on leveraging AI in your marketing. Failure to do so could lead to escalating already existing problems. For instance, if you provide the LLM bad data as context, you end up with biased, inaccurate, or simply false suggestions.

In the case of personalization, this can cause prospects to receive wrong versions of dynamic content. As for segmentation, bad data fed to the model can cause triggered automations to target the wrong people. The worst thing is that this kind of AI malfunction will go unnoticed until a customer complains about it. And then it will be too late. Bad data will undermine all efforts to incorporate AI into marketing.

Avoid pitfalls with a clear strategy

There are plenty of what I call “AI nihilists” out there. A common criticism is that LLMs will always produce biased, limited, or flawed results. There are others who believe LLMs’ potential to be fundamentally out of reach for most businesses. I’ve never found this sort of attitude particularly helpful. We saw the same thing with Electric Vehicles, and now we’re seeing an unprecedented surge in both demand and production of electric cars. Which side do you choose: the progressives or the laggards?

Assuming you chose the former, you are looking for a way to combat the problems I’ve outlined in this post. The best way to do this is by shifting perspective to a strategic level.

You need what I’d call a marketing data strategy. It sits between the company marketing strategy and data strategy. Essentially, we want to explore how data and AI relate to marketing strategy building blocks, e.g.:

• Segmentation (How do we identify our key segments? Which data sources do we need?
• Targeting (Which life events or milestones do we look for?)
• Positioning (How do we determine customer pain points? How do we know if our marketing is successful?)

The above are illustrative examples; your marketing strategy is your own, and there is no need to reinvent it. Ideally, you want to leverage what you already have as much as possible. There are, however, certain areas that you want to include in your marketing data strategy regardless of your chosen format. Here are my suggestions:

Marketing Data Strategy:

– Data Sources and integration
– Compliance
– Consent and preferences
– Personalization
– Reporting and analytics
– Storage, maintenance and retention
– Targeting and triggering

A marketing data strategy lays out rules and policies that intend to help organizations to develop their marketing technology stack, incorporate new data sources, build new marketing workflows, or adopt new processes. It provides a bedrock that you can anchor decisions on, and it ensures that you consider all relevant questions before making major data-related decisions. As is the case with any strategy, this too is a living document; don’t let it fall into disrepair – include regular checkpoints to assess and update the strategy.

Final thoughts

We’ve outlined the possibilities and risks associated with the use of LLM in marketing. By now, you should have a better understanding of what it takes to adopt an LLM in a business context. I’d like to add that LLM adoption requires a significant investment of company resources. To work properly, an LLM may require a Customer Data Platform, data lake, data management platform, analytics platform, or a combination of these. In addition, you need to hire or allocate specialists to supervise and develop your organization’s AI capabilities. Finally, I strongly recommend including a Chief Data Officer role for any business seeking to tap into AI. With top-level sponsorship, any such initiative has much better odds at succeeding.

Catch up on my session at Dreamforce where I explored this topic even further

You will need to sign up for a free Salesforce+ account to watch this video.

.