Scribd
Upload
21 views

Module 3

The document discusses AI applications in hiring and people management. Some key points: 1) Machine learning algorithms can help with hiring by analyzing applicant data to predict which candidates will be the best performers, though the predictions may not be perfectly accurate. 2) There are challenges around ensuring algorithms don't incorporate bias from flawed training data and around explaining algorithm recommendations. 3) Predicting human behaviors like attrition is difficult since there are many unmeasured factors influencing decisions, and the available data has limitations in terms of completeness and accuracy of measures. 4) While predictions from people analytics may not be perfect, the tools can still provide some advantage over unaided human judgment by identifying meaningful patterns in available data.

Uploaded by

ishaan
Copyright
© © All Rights Reserved
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
21 views

Module 3

The document discusses AI applications in hiring and people management. Some key points: 1) Machine learning algorithms can help with hiring by analyzing applicant data to predict which candidates will be the best performers, though the predictions may not be perfectly accurate. 2) There are challenges around ensuring algorithms don't incorporate bias from flawed training data and around explaining algorithm recommendations. 3) Predicting human behaviors like attrition is difficult since there are many unmeasured factors influencing decisions, and the available data has limitations in terms of completeness and accuracy of measures. 4) While predictions from people analytics may not be perfect, the tools can still provide some advantage over unaided human judgment by identifying meaningful patterns in available data.

Uploaded by

ishaan
Copyright
© © All Rights Reserved
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
You are on page 1/ 33

AI Applications in People Management

Hiring as an Example

Peter Cappelli, George W. Taylor Professor of Management and Director of Wharton’s Center for
Human Resources
Hiring

• Hiring is currently the most important people management issue, and the
one on which the most money is spent
• Replacing someone costs about $4,000 on average
• Higher up in an organization, turnover can cost the equivalent of 2 years
salary
• Staffing industry is $200 billion
Hiring Today

Reality looks much different now


• Majority who changed jobs weren’t looking to move
• Short lists are created by applicant tracking software, not people (just
automation on key words)
• People Scout, a recruitment process outsourcer (RPO), hires 300,000
people per year
Hiring Today

Reality looks much different now


• Interest in passive candidates
• Find people who are not applying and try to bring them into our system
• Goal seems to be to get more applicants into the “funnel”
• About 2% get job offers
Better Hiring

• Find the right applicants


• Picking the right ones to hire
• What is a good hire?
• Who are our best workers?
• Define what is “good”
• Then try to identify everything we know about our best employees
• Get the same information for the not good employees — need variation
Better Hiring

• Split the data gathered in half


• “Training” data on which the machine learning software is going to learn —
it will build a model that predicts the performance score we are using to
identify good vs. bad employees
• Test it on the other half of the data
• Go to applicants and enter their data for all the attributes we used in our
model
• At the end of the process, we get a score as to how closely they appear to
match our best employees
Advantages and Issues

• Likely to be much better at predicting


• We need a lot of data — need thousands of applicants to build a machine
learning model
• Bias
• Good news – algorithms treat every candidate the same
• Bad news – if there is bias in the training data that built the algorithm,
there will be bias in the algorithm itself
Advantages and Issues

• If it is a good algorithm, you don’t need applicant tracking systems


• Some people who turn out to have a good fit for a job don’t have the
attributes you thought were important
• In order to know that the algorithm continues to work, you need to do some
hiring randomly
Other Issues

• Some results may not seem intuitive


• Communting distance
• Facial expressions
• Explainability
• Can you tell somebody why this works in a way that seems sensible?
• Thousands of vendors now — each selling a solution based on machine
learning type algorithms
• Can they provide evidence that the algorithm works based on real job
performance?
• If you can’t validate the algorithm with your own data, this could lead to
trouble
AI Applications in People Management
Broader Issues of Privacy and Ethics

Peter Cappelli, George W. Taylor Professor of Management and Director of Wharton’s Center for
Human Resources
Broader Issues

• Hiring Example: Do you go with the candidate you know or the one with the
higher score on the algorithm?
• When asked, almost everyone picks their own candidate
• People want to have a say
• Good relationships with supervisors are built around an exchange
• If you turn decisions over to the algorithms, you start to erode the
relationship between supervisors and subordinates, potentially weakening
the whole system
Broader Issues

• Algorithms generate a score based on the entire sample


• If you break the sample up by some demographic attribute, you might find
that the algorithm gives systematically different scores to different groups
• Law won’t allow separate models by attribute
• The law lags practices by quite a bit
Broader Issues

Privacy issues — what do we feel okay predicting?


• Flight risk (turnover)
• Best flight risk models use social media
• Once you start monitoring social media, employees are likely to find out
• Race for “authentic” information
• Tattleware/spyware — monitors you while working to make sure you are
actually working
• About 1/3 of employees cover the camera on their computers
• Employees shift to using their cell phones to talk to coworkers to avoid
monitoring
Broader Issues

Privacy issues — what do we feel okay predicting?


• General Data Protection Regulations (GDPR) in the European Union
• “Right to be forgotten” – employers should not keep old data about
customers or employees
AI Applications in People Management
Data Limitations

Matthew Bidwell, Associate Professor of Management


Data Limitations

Technical limitations of AI
• How much accuracy we can expect from AI based models
• The role of inaccurate data
• The role of unmeasured factors
• How we should think about what these algorithms can do for us and also
what they can’t
Expectations for Accuracy

• Some accounts convey the impression that if we use the models right, we’ll
be able to get almost perfect predictions about how people will behave
• Misleading ways of representing the data
• Analogies to the success of AI in other simpler tasks
Expectations for Accuracy

Employee

A
B
C
D
E
F
G
H
I
J
K
Expectations for Accuracy

Estimate of Whether Someone is Going to


Employee Leave Next Month

A
B
C
D
E
F
G
H
I
J
K
Expectations for Accuracy

Estimate of Whether Someone is Going to


Employee Leave Next Month

A 0
B 0
C 0
D 0
E 0
F 0
G 0
H 0
I 0
J 0
K 0
Expectations for Accuracy

Example: Excel attrition model


• Not helpful to just apply the average attrition rate to everybody in your
organization
• What we really want is a model that explains the differences between the
people who leave and the people who don’t leave
• Can we get a model that will explain 95% of those differences?
• No — if we explained 30% we are probably doing well
• Why aren’t these systems more accurate?
AI Applications in People Management
Why Aren’t Our Systems More Accurate?

Matthew Bidwell, Associate Professor of Management


Why Aren’t These Systems More Accurate?

• Helpful to compare with problems where machine learning does much better
• Speech recognition
• Why is this so different to an attrition prediction?
• In speech recognition, taking something that has already happened and
then classifying it into categories — different than predicting the future
• In a classification task, you generally have all the information necessary
Why Predicting Attrition is Harder

• What do we need to make a really accurate attrition prediction?


• Demographics
• How long they've been in the job
• What the job is
• Performance evaluations
• Other job applications
• Posts on social media
• Is that enough to make a fully accurate prediction?
Why Prediction Attrition is Harder

All sorts of other information that might influence whether they stay or leave
• Fight with the manager • Set up their own company
• Plan to return to school • Join friend’s company
• Will they be offered a better job? • Easily bored and likes change in
• Leave for reasons of ill-health their lives
• Risk averse
• Leave to look after a family
member? • Secretly in love with one of their
• Relocate to follow a partner or be colleagues and unable to consider
working anywhere without them
nearer to family?
Complexity of Human Decisions

• Human behavior is complex


• So many aspects of personality that might affect decisions
• Many things outside the individual that shape behavior and decisions
• “Open system”
AI Applications in People Management
Measurement Problems

Matthew Bidwell, Associate Professor of Management


Measurement Problems

• Central problem in applying machine learning to human behavior is that we


never have all of the measures that we want
• Even the measures we do have can often be inaccurate
• Do people really want to tell us how they’re feeling?
• Surveys — people may misunderstand the question
• Analyzing text — people may be answering slightly different questions
• Measurements may only reflect a point in time, not how they are going to
feel later
• If we only have a rough approximation of what people think in our data, that
is going to make it harder for us to make accurate predictions
Data Limitations

• Always need to remember:


• Every model is based on incomplete data
• Every model is also based on inaccurate data
AI Applications in People Management
So Why Use Analytics?

Matthew Bidwell, Associate Professor of Management


“We are card counters at the blackjack table, and we’re gonna turn the odds on the casino.”
Moneyball, 2011
So Why Bother?

• Card counting is a way of skewing the odds when playing blackjack


• If you implement effectively, you’re still going to lose plenty of hands
• But the point is you will win more frequently than you lose
• It's a similar logic when it comes to implementing AI in HR decision-making
• The models will not be perfect
• The question is not “Is the AI model perfect?”
• The question is “Is it better than the alternatives?”
• Skew the odds of making the right decision in our favor
• If we can do this, we will end up being more successful

You might also like