Prediction and attribution are age-old problems in marketing and sales (henceforth referred to broadly as just "marketing"):
For sales prioritization, most B2B revenue teams still resort to ad-hoc rules like those shown below. Attribution is similarly based on arbitrary methods like last touch, equal-weighted, linear and so on.
In other words, hunches and instinct rather than data and scientific rigor. To nobody's surprise, these ad-hoc methods lead to wasted resources and lost deals!
This state of affairs in the marketing world is especially ironic given the vast amounts of customer journey data now logged in automation platforms. What has been missing is a flexible, universal engine that intelligently extracts patterns from this type of sequential data; a tool that observes hundreds of thousands of prospect journeys, and learns to predict their behavior, and accurately quantifies how much past touch-points contribute to these predictions.
One cannot help but wonder: with all the recent AI advances, surely some of these can be leveraged to help finally solve the prioritization and attribution problems? At first glance it is indeed shocking that this has not happened yet, and one can be forgiven for concluding that marketing has somehow been left behind by the AI revolution. But a deeper dive reveals two key technical hurdles for advanced AI solutions in customer journey intelligence:
(1) Marketing data consists of event-sequences, which have largely been ignored by recent AI advances. The reason for this is twofold. Firstly, this type of data is inherently messy to handle. Unlike say periodic, numerical time series data from the stock market, customer journeys consist of mostly non-numerical event-types that occur at arbitrary time-points. (Event sequence data is of course not unique to marketing; they occur in numerous important business settings, such as in process mining, healthcare, security and finance). Secondly, event sequences are significantly more challenging for AI model training. Simplistic machine learning approaches do not perform well and require a lot of manual work for each new dataset.
(2) Trustworthy AI is critical for adoption by marketing teams. If a model predicts that an account has an 80% chance of a Closed Won within 3 months, a marketer usually would not just trust this score at face value; they would want to know what factors drove this score. This is the attribution problem, and Trustworthy AI techniques provide an elegant and principled solution. The transparency provided by these techniques can not only build trust, but also identify weak spots in a model, or oddities in the data, so the AI model can be improved over time.
Even among the relatively small number of folks with technical expertise in both these areas (i.e. modeling event sequences with modern deep learning, and layering on Trustworthy AI to ensure adoption), there is only a fraction that have interest and domain knowledge in the broad marketing space. This is the third factor that has contributed to the paucity of advanced AI in marketing.
XaiPient represents a unique confluence of these three factors. Building on our deep expertise in event sequence models and Trustworthy AI, over the past two years we have built a universal, flexible and explainable prediction and attribution engine for event-sequences. Our first products focus on customer journey prediction and attribution in marketing.
Just like natural language sentences have patterns and predictability, event sequences in customer journeys are predictable, except that these have a notion of time. Starting with this intuition, we build on recent advances in Natural Language AI models such as Recurrent Neural Networks, Transformers and Self-Attention (the latter two underlie the popular BERT and GPT3 language models), and the mathematical framework of Temporal Point Processes. Our prediction engine is powered by our proprietary innovations on top of these ideas. An early version of our models can be found in a paper we published at a top ML conference (ICML 2020).
Our prediction engine powers XaiPient’s HubSpot app, and we now also have a plug-and-play version (see below) that can generate predictions and attributions from any other marketing data (as shown below), such as from Account Based Marketing (ABM) platforms or data aggregators. If you are a B2B revenue team looking to boost your close rate and marketing/sales efficiency, feel free to reach out to us -- we'd love to show you how our customer journey intelligence engine can help!