Suggesting the next action with artificial intelligence
Improving the human and machine interaction for commercial excellence
In a previous blog, we alluded to the point that artificial intelligence and machine learning could support the work of sales representatives by optimizing the next action to be undertaken. We also discussed the idea that solely providing a recommendation on the channel and message was not enough. The artificial engine should provide a perspective input on all variables influenced by the reps. This piece reflects on the principles that should apply and therefore should be included in the selection of the proper engine. Most of the CRM platform for life science nowadays supports the plugin of artificial intelligence. Nevertheless, not all predictive models would yield the same results.
The objective of leveraging artificial intelligence in such a context would be to optimize the actions performed by reps. The result of the implementation should be improved prescription numbers from the visited HCPs. it is therefore of utmost importance that the engine creates a model leading to sales volume. Hard sales numbers may sound like a very simplistic way to tackle a complex problem but it is also one of the only objective dimensions. Everybody must keep in mind that the reps’ behavior would remain only a part of the equation. Even with the best behavior from the sales force, some products with limited clinical benefits, for example, would not materialize in sales.
In BASE life science, we believe that the model can be extremely complex but that it should only try to predict sales within the locality of the HCPs. This is the only way to have an objective measure of the results. It should, of course, include the feedback from the sales force, but the initial starting point should be the creation of a model of the sales within the area of the HCP. Given the current laws, it is not allowed to know the prescription volume of a specific healthcare professional, but the prescription within its sphere of influence should be the output of the model.
Creating a model to optimize the human and machine interaction
The dependent variable of the model must be sales, but for the input the independent variables must encompass a broad range of aspects. There should be a characterization of the physician and the overall context (such as disease prevalence within the area or economic situation). Those should be used to normalize each sales rep’s territories. Finally, the last set of variables should be linked to the activity of the reps. This could encompass the number of calls, naturally, but also more interesting data points such as the type of messages presented to the HCP as well as the timing, etc.… The model could be extremely detailed with data points going to the lowest level of granularity such as the sequence of the slides presented. Artificial intelligence is not required within this step but would greatly help. Even if regressions are doable within excel or even more advanced tools, computers would be much better at defining the most suitable set of variables. The objective is not to have the most perfect model but more to get something reliable enough to be able to run an optimization.
The actual mechanism of suggestion should originate from this step, once the model is well defined. Essentially, when the reps click or tap within the system to ask for a suggestion, the system should run an optimization (similar to how the solver of excel would do). Essentially, the engine should run the model with the specific, controlled variable (e.g. the variables defined by the HCP) and deliver the set of actions that yield the highest sales number possible given the model. Again, this could be done manually but would represent a huge effort for a human being. Even with the help of software, optimization usually takes time and the traditional solutions do not have user-friendly interfaces, hence the value of using a CRM plug-in. Remember that ultimately, the AI will only optimize the sales force aspect of the variables and won’t be able to result in benefits beyond this dimension.
Assessing a machine learning engine in the field
The next step should be to include the feedback of the reps on the action suggested. This is important as it is a data point that will allow to make the machine learn but also to improve the rep’s behavior. Given the sales data is more objective, in the next period it will be possible to define when the rep was right to refuse an action but also when he was wrong. Using two perspectives would define the following matrix:
The perfect match category should be the metric of the process as it demonstrates when the machine and the human are marching in the same direction. Especially, it counts the number of times the rep followed the advice of the machine and it worked, meaning it led to an increase in sales.
The coaching opportunity is whenever the rep didn’t follow the machine’s model and it led to declining sales as predicted by the model. This is a golden opportunity for a specific coaching session, personalized by the sales manager on what can be learned from the mistake and how the system can improve.
The model failure category is the traditional learning path for a machine learning solution. This category should significantly decrease over time as the model gets smarter and smarter. Reps need to receive some feedback from the global team about the learning curve of the machine. They should be thanked as well for the constant feedback to ensure they continuously adopt the system.
The learning loop is the situation where both parties were wrong. The model suggested to act, the rep executed it but instead of increasing sales, it had a negative impact. It is important to indicate the number of occurrences to the rep. Again, this will reinforce the message that the machine can also be wrong, and that the role of the rep is of crucial importance. This would allow the machine and the human being to work together, ultimately improving the entire process.
To support the continuous improvement of the model, it is important for the reps to understand how the model reached the conclusion, whatever the category. This aspect, named interpretability[1], is necessary both for the user to learn and to identify bad data. Expanding on this aspect, it is also of great importance to have a mechanism supporting the constant improvement of the data.
What to include in the selection of an AI-based suggestion tool?
The evaluation of a suggestion tool requires a more in-depth analysis of the feature’s aspects. Most companies would be able to meet the requirement with configuration. This won’t allow differentiating between the vendors. To find variations between the engines, the following points could be included:
- Inclusion of objective variables to assess the model quality
- Whether the model also includes an optimization perspective
- How would the engine support the coaching of the reps?
- How could the reporting of the errors of the engine be communicated to the field force?
- Is the model logically explained to the user?
- Can the plug-in support the identification of bad data?
Adding those aspects to the more traditional vendor evaluations framework would enable a company to differentiate the solutions more accurately, ultimately not letting cost being the sole deciding factor. Selecting the right tool would support companies in gaining a more appropriate suggestion, but ultimately it will not result in more sales than the products available in the market can achieve.
About BASE life science
BASE life science is a fast growing, fast paced consultancy focused on the life science industry. Established in 2007 and based in Copenhagen, Denmark, BASE targets a local as well as a global customer base.
Since inception, BASE life science focuses on helping Life Science companies create real business value from digital platforms and data within its area of expertise; Commercial Excellence, Clinical, Regulatory Affairs and Quality & Compliance. Since 2007, the company has been active globally from Denmark and employs more than 50 employees.
Want to learn more?
Read our whitepaper: Commercial Technology supports Pharma in increased Competition
About the author — Jacob Dziegiel, Partner, Head of Customer Engagement
Jacob is a highly skilled consultant with more than a decade of experience within the life science area including a vast number of projects within pharma sales & marketing and enterprise mobility. He excels within project management and process optimization and he has a strong track record of converting new technology into high business value solutions.
