How to plan customer activities? (part 2)

Case study: Implementing Next Best Action Recommendations for Sales Territory Planning

This is Part Two of our blog series on AI-powered “Next Best Action” tools for field forces.

In our last story (see part One of this blog series), we discussed the needs of one of our clients. A large bio-pharmaceutical organization hired us to work on its customer activities planning (cycle targeting and call planning exercise).
The objective was to empower the sales force to create their territory plans for the cycle. Indeed
with autonomy comes responsibility and motivation to do things right!  

Context 2

That doesn’t mean that we shouldn’t nudge the salespeople in the most optimal direction. Indeed, let’s have a bit of empathy: making a plan for an entire sales cycle isn’t easy at all and requires support.

  • The typical rep might have 200 or more diverse individuals worth visiting on her territory. Some of them she already knows well, others she has never met.
  • Those individuals are either happy to work with the pharma industry or restrict access. Some are brand advocates; others favor the competition, but maybe could be won over.
  • How can she best allocate her limited time?


Let’s take the GPS metaphor. You might have noticed that when you enter a destination into Google Maps and ask for an itinerary, the first thing you see is not an arrow that says “turn left in 500m”. No, the first thing you see is a map with your location and 2 or 3 itineraries options overlayed.

What if I could have such recommended activities options in my resource allocation?


Inspired by that example, we provided the reps with an accurate map of their territories. Not literally a geographical map, of course, but a digest of the key information that they need to know to make a good plan.
We gathered and collated information from different sources: sales and performance, obviously, but also past activity and additional account & hospital metrics (more than 70 metrics). We used open data sources that they often had never seen previously. We provided them with a high-level summary of their territory, as well as brick level and account level information. When skillfully arranged in a meaningful brief, the data constitutes a map and shows the rep where they currently stand.

The next, harder, step is to show the direction to take. The itinerary options, if you will. In an ideal world, each rep would be treated as the “CEO of his sales territory” and would receive individualized advices from an experienced & knowledgeable consultant. We could have done that with a couple of territories, but certainly not with over a hundred. We had too much data, and the situation was too complex for traditional management consulting approaches. To solve that issue, we created an AI that plays the role of an automated management consultant.


microsoft GBM

  • 1st STEP: the AI needed to understand the drivers of high sales performance.

We built a model that predicts market share growth in geo-bricks based on all the available metrics (70+ metrics, i.e., historical sales, market share, geography potential, HCP demography, historical promotional activities, and more). We used Microsoft’s implementation of an algorithm called Gradient Boosted Machines. That AI software “thinks” in the same way a management consultant would: the machine uses statistical methods to build a long series of if-then statements branches. Those statements define a large decision tree that explains every possible situation observed in the data.

  • 2nd STEP: the AI needed to create micro-segments of customers with similar situations.tree 1

Again, this is similar to what a management consultant would do. But instead of a 2 x 2 segmentation matrix, the AI looked at more granular data and defined automatically over a hundred micro-segments. Geographies with similar inputs end up grouped in the same micro-segments with similar predicted output. We then use decision trees to explain that performance prediction.

What combination of “if-then” explains that we expect low growth in that particular geography?

tree 2

  • Some of the reasons might be things about which we have no control over – like the healthcare professionals (HCP) demography
  • but others are actionable variables – like visiting such specialty, or performing such activity in such local situation

If an actionable choice explains low predicted performance, is there a different choice that we can make that would send us toward a different path with higher growth?
A human being would have a hard time looking at all the possible paths across the branching graph of possibilities. Our AI had no such problems.

  • 3rd step: armed with our map and the recommendations from our AI, we provided each sales rep with detailed and granular recommendations.

For each geography and each account, we showed the reps what their situation to date was, what they had done in the past, and how our AI would improve on their past activities to reach higher performance. We compiled that information into personal planning briefs for each salesperson. We also deployed the data on an online platform so that they would have all the relevant information, individual by individual, when creating their territory plans. The platform, also developed by KARETIS, was very important to allow reps to work collaboratively with their teams and managers.


feedback The feedback from both management and sales force was great. The reps appreciated the autonomy and were impressed by the information provided. Most of them had never had access to such relevant data about their territories.  They also appreciated that it wasn’t just raw data. We provided them with a clear opinionated analysis of what it meant for them: our AI recommendations. Salespeople are not analysts and hopefully don’t stay at their desk evaluating figures.
Providing only the map without clear actionable and refresh-able direction options (the recommendations) would have been less helpful and more confusing.

validation methodologyThe more skeptical reader would be thinking now: “This is all very feel good, but what about ROI? ” Indeed, one of the great benefits of using AI for business analytics is that results can be validated and quantified.

We validated our recommendations by looking at what would have happened in the past. We tricked the AI in thinking that we were a year in the past (computers aren’t that smart). We then asked for the recommendations for the next 6-months cycle and compared to what actually happened. For a more robust analysis, we did it with 2 separate time windows and averaged out the results.

  • Some sales reps, unknowingly, came up with the same ideas as the AI and “followed the recommendations.” On average, 70% of them achieved higher growth compared to their historical baseline.
  • For comparison, overall, only 45% of the reps increased their growth rate.

If every rep had “followed the recommendations,” the market share would have been 10% higher after 3 years. Those are incremental results of course. There is only so much additional performance that you can squeeze. That sales boost represents several millions of euros/dollars.  The ROI on the cost of the project is over 10x… in only six months! 

Is it possible to reconcile greater autonomy and collaboration in the field force with data-driven and AI-powered management of sales operations? We think yes. It’s not only possible but very profitable. No turn-key solution is available yet, but development is surprisingly quick and cheap considering the advanced technology involved. As our experience shows, change management can be almost painless.

Have you thought about or even implemented similar solutions? We’ll love to hear about your experience. Get in touch with us to continue the discussion.
Wants to implement it? Contact any partner of KARETIS