How to segment employees and personalise experiences
In 2011 we published an article about employee segmentation which is still one of the most popular pages on this site. As we noted then:
The days of ‘one size fits all’ approaches is over. Segmenting the workforce is important, especially in tough economic times as it enables you to be as efficient and effective as possible, getting resources and initiatives to those most likely to respond / benefit.
One of the approaches in that article describes segmenting on the needs of the employee. Over the last decade this has been the dominant approach that we’ve used.
It should be noted that segmenting on behaviour is arguably the approach used by most ‘predictive analytics’, machine learning approaches that aim to classify a behaviour based on characteristics. Techniques such as trees can help identify target segments for action
Why segmentation should be an important part of your HR approach
In many firms there is a one-size fits all approach to how employees are treated. Employees are offered the same set of benefits, career development and working conditions regardless of their needs or desires. For example a benefit package might include childcare arrangements regardless of whether the employee has children or not.
Alternatively changes to address issues, such as changes to career development or management coaching for a supervisor is applied to everyone regardless of whether it is required or desired. We can think about survey action planning falling into this category, often with cascading priorities applied to groups who could be overperforming.
Obviously this approach creates a lot of wasted resources, and rarely meets any individuals ideal requirements (even though meeting these requirements might cost less to deliver).
Some HR teams offer a degree of differentiation, though this is often by relatively crude measures such as grade. Even though this approach, effectively increasing the size of the benefit package at more senior levels can make sense from a reward perspective it is still crude. The employees might prefer a different balance of offerings.
As individuals we’d all prefer more tailored treatment that matched our needs. Standard economic theory suggests that there is often significant lost ‘utility’ for the individual if they have to accept a bundled offering.
One approach to meeting this would be to provide a completely individual offering to each employee based on a certain cost. Unfortunately the delivery cost of this would be considerable. Another approach that goes part-way is to offer a menu of options and give employees a ‘budget’ to allocate. This is better but only tends to work for items it’s easy to allocate a cost towards.
Segmentation offers a different approach. With segmentation we identify groups of employees with needs similar-enough that we can design and deliver approaches that feel personal, but offer the scale to enable them to be delivered in an efficient manner.
Employees as consumers of work.
OrganizationView has always believed that we should be adopting ideas and techniques from the consumer-parts of our businesses to manage people.
As we noted in our first insight article in 2010:
For employers to compete effectively they must make themselves more relevant to their employees than their competitors.
Marketing has for long tackled such an issue by understanding and measuring consumer needs. The study of consumer behavior has provided answers to when, why, how and where people do or do not buy products or services. Central to this is an assumption that individuals seek to maximize their total utility. As two key constraints are their time and their money, how they wish to balance work and non-work parts of their life must be interdependent.
The way individuals think about managing this relationship between work and non-work parts of life (the so-called work / life balance) is inherently individual. We can’t assume that all individuals and striving to improve their careers, working to further a social belief or even just working to provide resources to support their lifestyles. We really need to get data to understand their needs.
Segmenting on needs through survey data
Of all the data organizations have on employees the best source for understanding needs is survey data, and most critically the qualitative answers to open questions provided in text.
Marketers have long used sophisticated approaches to do this, most notably using ‘discrete choice’ experiments. In these a respondent has to trade-off one offering to acquire something that is more likely to meet their needs. They have the advantage that the respondent can’t say they want everything, an unrealistic situation for most.
Conducting these experiments well is expensive and time-consuming. What we’ve increasingly seen is that other data sources, especially large volumes of text opinions, can provide a very similar set of results to the best-run discrete choice exercise, but at a tiny fraction of the cost. You might be paying 20x more for the discrete choice experiment yet achieve 2% more accuracy.
Typical employee surveys comprise two questions which can be valuable at providing the data needed for these segmentation exercises:
What are the best things about working for COMPANY?
What could we change to improve your experience of working for COMPANY?
It is possible to ask different pairs of questions specifically for segmentation. Whether you do that really depends on the purpose of the segmentation you’re doing. The above questions probably provide the broadest use-case.
Classifying the data
The first part of any segmentation exercise is to map the thousands of answers against a classification model, effectively grouping them into similar themes.
The best way of doing this is to build an inductive model - one where the themes in the classification are built from those mentioned in your data. This ensures that your classification model is as close to the reality of your data as possible.
A second approach is to start with a general classification model built on a large volume of data and add any categories which arise in the data but not in the model. This approach, effectively using an inductive approach on any data not sufficiently coded by the deductive model, can provide accuracy approaching that of the fully inductive approach.
The final approach - using a fully deductive generalised approach - works best when the model itself was built bottom-up from an inductive approach using a large cross-company dataset. These models can usually be identified by having a large volume (100+) categories. Be wary of a model which neatly matches the other survey questions as it’s likely to be hypothesis-led.
Segmenting based on topics discussed
There are a large number of ways of segmenting the responses when they’ve been coded. Segmentation relies on an algorithmic technique called ‘clustering’ which differs from most other machine learning techniques used in HR because it is unsupervised, that is the patterns it finds are learnt from the data without having an outcome variable (like attrition) that they’re trying to predict.
The approach that we use for segmentation is to use a probabilistic model of how unusually common two themes being mentioned by the same individual is and then using these probabilities to create a weighted network of the interaction between themes. Effectively it’s grouping themes (needs) together into groups which are similar not by meaning, but how likely individuals share them.
In network analysis there are numerous clustering approaches - in network terminology ‘communities detection’ - which you can then use to spot relatively clear groups. Unlike some clustering algorithms these community detection approaches typically identify the ideal number of segments (some clustering techniques require an analyst to decide in advance the number of clusters required).
We can plot the result as a graph where the communities are shown.
Clustered themes from a survey: “What is good about working for COMPANY”
With the right data, and enough responses it’s possible to build good employee segments using this approach that relate to the needs of the employees.
Probabilistic mapping of segments to employee characteristics
When you’ve built your segments, based on the needs, it’s useful to identify which type of people are most likely to fit into which segments. Again we can use a variety of machine-learning approaches to do this.
These approaches usually fit a similar pattern. First you classify each respondent by the needs-based segment that their comment best fits into. Then you build a predictive model based on other data (not the text question) that you have on the individual. It could include other survey questions, demographic information, organization-type data, performance data etc. It’s important to use a machine learning approach that is interpretable (i.e. it produces a set of relationships that are easy for a human to understand).
Segmenting on other data
I mentioned above it’s useful sometimes to link needs-based clusters from the text classification to other survey questions.
One challenge that you’ll find is that in a typical survey employees often groups of questions in a very similar manner. To complicate things further employee survey data - especially Likert data - isn’t terribly rich. Each question often only has 5 possible answers.
A way of overcoming this is to use a segmentation technique to cluster the questions together. We can then build indexes of these clusters and look at employees responses on a distribution of scores, for example identifying those employees in the top or bottom quartiles for each group.
A weighted graph of partial correlations for a typical employee survey
Again we use community detection on a graph of linked questions to identify these clusters. To counter some of the common issues with this data type we use partial correlations between the questions.
Typically these clusters will show groups similar to the main categories in your survey though with some expectations. However, by creating groups of survey questions in this manner it’s easier to find meaningful groups of employees that can be used for the next step - bringing those segment to life.
Building personas
Personas are a set of techniques that have long been applied to bring different customer groups to life. By using personas we can humanise our findings, making them easier to relate-to and therefore empathise with.
Much of the empirical data that you need for creating personas can be drawn from the analyses detailed above. The clustering on text provides groups with similar needs. The mapping of other characteristics to these groups helps shape which types of people are likely to fit into each group.
It’s also possible to use various information extraction techniques to identify quotes in the body of responses that most simply illustrates the needs for each of the personas.
Creating personas therefore should be data-led, and much of this collation can be automated to a large degree. However, our experience is that this data, regardless of how ‘complete’ it seems should be used as an input only to creating segments. The process of a team developing segments from the data can be instrumental to ensuring the personas are accepted and used within an organization.
Developing an action plan and Prioritising segments
Not each of your segments will be equally valuable. Equally not each of their needs will be as easy, or costly, to address.
Given the output of the steps above it will, however, be far easier to identify a series of actions. Grouping them together for different segments ensures that you can be much more integrated in how they are delivered, often combining them in effective bundles. he act of mapping the needs to characteristics and then developing personas also makes it far easier to target the interventions to the relevant groups in your populations, thereby reducing cost and increasing effectiveness.
In many instances it’s important to create offerings that appeal to certain segments but not to others. In this way it’s possible to create employee ‘packages’ that are desirable to one or several segments but those with different needs can self-select out. As long as there is an alternative that is more appealing (ie it’s not a case of this-or-nothing) then you’ll likely increase the overall satisfaction of the population whilst reducing waste.
Moving to persona-driven HR
Since starting in 2010 we’ve been convinced that applying segments and personas to employees is a key way of delivering strong employee experience in an efficient and effective manner. They can be one of the easiest changes to increase the ROI of your HR activities.
The advent of advanced text analysis to find groups of employees with similar needs has dramatically improved the ease of moving to a segment-based approach. No longer do large firms need to conduct expensive and timely exercises to build robust segments. Now it is possible to construct good segments in a few days using data that is probably already being held by the organization.
One final word of caution. Any type of segment is not cast in stone. As people’s lives change they move between segments. A young graduate living with friends is likely to have different needs than when they are parents a few years later. Even segment needs evolve (slowly) over time as alternatives and circumstances change. Segmentation models, like any model need maintenance and updating.