Artificial Intelligence in Marketing | The Ultimate AI Canvas

Machine generated transcript…

Hey everybody before we get started just to let you know you can download a high-resolution version of the canvas that were gon na, introduce Ive, put a link in the description or you can simply go to this URL lets jump into it. The press tends to focus on the latest academic findings of AI and a growth. Try we like to focus on the low-hanging fruit, the practical and more tangible uses of AI for marketing and growth, the stuff you can start applying tomorrow, if you havent already done so. Please

Check out the earlier videos where we actually address these easy low-hanging, fruit, applications of AI from marketing and growth over the course of the last year, through the trainings that weve delivered and through the use cases, weve worked on weve developed, an AI canvas which will allow You to determine whether you should apply AI for a business case at all and ensure a good scoping and preparation of an AI project. Although other canvasses exist. We believe this one best meets the needs of analytics translators, domain experts, marketeers heads of growth and data scientists alike, so heres, a quick

Overview, the canvas is separated into five columns, which should be covered from left to right. First, the business side of things and the data and the actual testing phase right in the middle next is the evaluation phase to see if this is actually working and, finally, the action phase when you have a model that’s worth implementing weve, also added a practical little Indication of

Which sections are the most difficult and/or, the most important, with this little yellow dot, youll probably need to look out for these sections theyre more critical, and they typically require more time. First of all, the business and funnily enough, this is actually a very often overlooked. Part of an AI project weve often said that data science shouldnt be left to data scientists alone. It should also be in the hands of

Domain experts or heads of departments who understand the business implications of running an AI project check out this video, where we talk about a new role called the analytics translator, which is a mix of domain knowledge, AI intuitive understanding, project management and enterprising spirit. The first section. What is our goal as youll see, has an important thought here. You should answer what question do we want the AI to

Answer are you an online platform that wants to anticipate which customer will be worth the most and which will leave the platform? Are you a b2b company that wants to score leads? Are you an energy company that wants to cluster customers, or are you a new startup? The ones to automatically translate customer feedback or improve the product and more do we even need AI to help answer this business question sometimes descriptive, analytics business intelligence, a simple Excel file, or even simply talking to

Customers is enough to get the job done. Sometimes the project stops right here in this box, because we realize we dont need AI to solve this business problem. Sometimes bringing in AI is like bringing in a bazooka to kill a fly. Of course, dont kill flies for a great list of questions that we believe a. I can help answering go ahead and check

Our navigation cheat sheet on this URL. The next section is project priority, and this is an absolute showstopper. One of the top three reasons why weve seen AI projects not come to fruition is simply because they were addressing an issue that wasnt a priority for the organization. Ai projects involve time, commitment and multiple stakeholders. If you work on something that’s, not a high priority or high on the roadmap, you might not get the internal buy-in and the stakeholder support to make the project successful. Make sure you work on

Something that matters something that you can showcase internally, which leads us to our next section about possible blockers. What possible hurdles or blockers should we expect and begin to alleviate access to data access to automation, tools, privacy issues, lack of resources, poor team commitment, pushback from other departments list anything that could block your project, which then leads us to stakeholders which internal people and stakeholders Should we involve early on to remove possible blockers and to get support? Some usual suspects are CTOs chief data officer x, product owners executives, but also sales

Marketing UX or crm teams now that weve validated that this is a valid project to work on now that we think we can get company buy-in its time to move to the oil of the project, the electricity of the project, the actual data we begin with the Most essential part of this column in some regards this is a throwback to the. What is our goal box, but its a bit more precise. What specific outcome are we looking for in the case of prediction models? What metric are we trying to predict in the

case of clustering what patterns are we looking for in the case of natural language processing what are we looking for in the text sentiment topics in the case of image detection what are we trying to uncover from a tensor of images now based on this outcome that we want to achieve what data or data sources do we need in order to achieve this outcome in this section Id suggest not limiting yourselves to what you.

think is available really focus on the data that you would want in the ideal scenario sit down with your team and start listing them on a whiteboard and try to be as creative and even as unrealistic as you want now this next section is going to be the data reality check and as you can see this one has an importance dot its critical and difficult based on the data that we want what do we actually have at the.

moment and how is the data actually structured this is the part where you need to involve other people who manage the data within the organization and typically the larger the organization the more people are involved critical question here do we already have all of the data or do we need to start acquiring it do we need to start gathering it and by the way in this.

section youll usually realize that you forgot a few key stakeholders in the stakeholders box some of them will start coming to mind before we move to the next column just a quick reminder you can download this canvas at this URL cool now lets move on to the actual testing phase in this section youll need to use your AI knowledge to navigate the testing phase of the project without the right understanding of AI as a tool this might be a trickier part of the canvas but that’s also why we give training in AI for non data.

scientists so were gonna kick off with how are we gonna measure a success of our AI project what metric will we use to define the success criteria of the AI project here we will use conventional measures of success depending on the type of AI project lets give some examples for supervised learning projects like classification models we would use recall precision or accuracy.

for value regressions we would use our two scores for clustering cases wed use the silhouette for natural language processing weve used the LDA and so forth now these are some of the things we teach in our courses and these are learnings that are acquired through a little bit more than a 10 or a 15 minute YouTube video next up is pretty specific about what tools and what algorithms were going to use to try to answer the question of the project what algorithms.

are we going to use to reach our outcome variable based on the data that we have it can sometimes be hard to navigate the wide and complex range of algorithms that can be used in AI projects which is why we created the AI cheat sheet which I mentioned before next one is really interesting what is the current human performance human performance on this task thanks to this will later be able to measure the money time and performance impacts of actually using a model should we even use a model some examples in the.

absence of machine learning models we are creating personas and clusters simply by hand another example in the absence of a model we are treating all customers the same regardless of whether or not theyre likely to churn in the absence of a model we are analyzing tons of customer feedback by hand rather than using topic modeling these are just some examples here you really want to set the current human benchmark and now to the final section of the actual testing phase do we need to start from scratch.

or can we use machine learning as a service and off-the-shelf solution that’s already been prepared many image recognition algorithms have already been pre trained would you want to start from scratch same thing for nudity detection models or profanity detection models but in the majority of cases your use case will be specific to your business and we will need to start from the beginning great now onto the next section after weve tested the model its time to assess its performance are we happy with.

the results for prediction models are there too many false positives are we happy with the ROC AUC for a clustering exercise a segmentation exercise are we happy with the number of clusters are we happy with the score of the silhouette so basically put are we happy with the output of the model or do we need to optimize and/or retest it this next section byproducts is actually really interesting when running machine learning models some types of models are what we call interpretable models you.

could interpret them this means that you can actually dive under the hood and go into the brain of the model to see how the model took its decision how it came up with the outcome this is the case with regressions or with forests for example but its not the case with neural networks which we call black boxes although some people are working on this at the moment trying to make.

black box neural networks interpretable so one great byproduct of interpretable algorithms is that they give us extra information sort of a nice-to-have so for example in the case of a predictive model we might be able to see which features which factors which variables are linked to the outcome in an unsupervised model in a segmentation model for example we could see which.

features and factors are the key drivers of differentiation between the different clusters okay now lets move on to the next part called interventions to test and this is where your brain starts to be important again based on the outcome of the algorithm which experiments do we want to start testing ai models are not like a/b tests they dont prove causality you still need to test whether when live the automation will have a positive effect so in this section youll need to come up with ideas for new features new offers new campaigns new nudges new cop.

new marketing or product or sales efforts that you want to run based on the outcome of the model and then its up to you to create the experiment so that you can reliably tell whether it had a positive a neutral or a negative effect on the metric that you’re trying to improve and in the next section we would put the results or the findings from that experiment did we get it uplifting conversions where were able to drive more traffic where were able.

to increase Net Promoter Score were able to increase retention increase engagement was the experiment effective now lets move on to the fifth column the last part of the canvas now were in the action part the part where we take action in the first section its about automation so if were happy with the model do we actually want to automate this can be for example automating the lead scores with the direct message to the sales team automatically putting.

customers into clusters once they hit the website and do enough actions automating a recommendation engine automating your topic modeling algorithm to understand what topic a customer on a chat is talking about this is the part where all your hard work and research can finally be duplicated and systemized and automated so youve run your experiment youve done your AI test and.

the results are interesting or even amazing but how to communicate this to other people within the organization you need to start asking yourself that question this is where you gain love and traction and internal Evangelization of this AI way of working do you need pivot tables showing the churn rate per account type and user type the list of primary features related to churn if youve created new customer clusters give them meaningful names that everyone will understand and describe those personas based on the most important features to ease visualization for key stakeholders use graphs and charts that.

they can easily understand and then the last section which also has a dot is about driving adoption people shouldnt be dependent on you to be able to keep running and optimizing this model how do you drive more AI adoption within the organization and although this is a continual non-stop process we have included it in the canvas how do you make sure you’re not the only one who.

can run this model at least another person data scientist or somebody within the specific departments should be able to also train and test and run this model so the basic question is how are we going to bring other people on board to be able to understand this and run it themselves so that’s it I hope you enjoyed this journey through the AI canvas as always please support us by commenting and subscribing and well see you really really soon.


Comments

Leave a Reply

Your email address will not be published. Required fields are marked *