The world of AI is having a profound impact on Martech. Advanced algorithms are helping organizations analyze risk, provide insights and help sales teams hit their targets. AI is being used to connect online and offline activity, using geolocation technology to track history of places visited by a customer. It’s also being used to create content personalized for each customer based on what they would like to see, effectively changing the marketing landscape as we know it.
The need for AI in Martech comes from an ever more complex media environment. Customers generate huge amounts of data, and marketers need to be able to react in real time. Customers won’t notice AI in action, but they will notice when an ad feels personalized and relevant to them.
However, while new ground is being broken in AI at fast pace, even experts will admit that there is a lot of work still to be done. We’re still quite a long way from fully integrated AI processes, even if we do live in a world of talking assistants. Let’s take a look at these challenges in more detail:
AI is reliant on huge amounts of data
It sounds almost axiomatic to say that AI relies on data, but many people underestimate the huge quantities of data required. Neil Lawrence, a professor of machine learning at the University of Sheffield said that even human-level performance is driven by “an unearthly amount of data.”
Lawrence points out that vision systems see way more labeled images than we require to recognize objects, and that speech systems require many more words than we do to understand words. AI systems will become easier to work with and more efficient, but this will take time. For now, any AI initiative must ask: what data do we have, can we trust and use it, and how can we get more?
Verifying and Validation
It’s essential to be verify and validate machine learning models. Is using a machine learning algorithm improving the performance of your campaign? What effect does an online campaign have on bricks and mortar store visitation? Can you measure brand uplift? Are the results statistically significant? Can you be certain that the model is more effective than a randomized control group?
Explanation and prediction is also a challenge, and you must know which you are optimizing for. For example, deep learning models are very good at predicting, but often less good at explaining why a decision has been taken. Simple Bayesian models and decision trees, on the other hand, are able to provide stronger explanations for why a decision has been taken, but are less effective at accurately predicting outcomes.
Choose the Right Use Cases and Monitor
AI is not a panacea for your business, but a patchwork of point solutions. Because there’s no such thing as a “general AI” yet, it is important to know which problems an AI solution is able to solve. AI in Martech is well suited to problems such as improving bidding and optimization algorithms, determining the best combinations of campaign settings such as pricing, budget and creative, to solve recognition challenges (such as click fraud), and segmenting audiences by categorizing unstructured data from logs.
Most of the time, you will be doing supervised machine learning: training algorithms against a known “validation” set before running on a wider data set. Even then though, you need to be careful about introducing bias. One study showed that predictive models had a tendency to “go where the signal is” – ads in a flashlight app overperformed relative to expectations not because the ad was engaging, but because when users use a flashlight app they tend to be fumbling around in the dark. Misfiring ad tags, accidental clicks and even incorrect upstream metrics can all interfere with the performance of a model.
Hire AI specialists
There’s a reason why the AI job market is dominated by people with data science qualifications and PhD’s. AI requires specialists who not only understand the technology, but who are able to explain how a conclusion was arrived at, and also how to avoid introducing bias into their work. While programming is a less specialized job than AI, engineering hires should still understand how to use languages such as Python and R that will enable them to get the most of of the data science work.
Create Data Science / Marketing Alignment
It’s no good if the AI part of your business is a research lab where AI technologies never reach production, and AI technologies will never be fully used if marketers do not understand the impact that they can have on their workflow.
Marketers live and die by their KPIs and metrics, and want to see return on investment before committing to using a technology. Marketers don’t have to understand the inner workings of a technology, but they do have to understand how to use it and how it will improve their workflow.