fbpx

5 real world examples of how AI helps grow businesses

Examples of how AI have helped increase revenue and decrease cost across different industries.

5 min read

We’re going to bring you on a tour of real-world examples that we have worked on over the years, to give you an idea of the kinds of challenges that machine learning is very suitable for, and achieves good results.

These examples provide a glimpse of the versatility of using machine learning, and how you can leverage the power of AI to:

  • Increase revenue (e.g. understand target market)
  • Decrease cost (e.g. reduce churn)
  • Increase capabilities (e.g. attract great employees)
  • Increase efficiency (e.g. automated pricing and quotations)
  • Drive ROI (e.g. marketing campaigns)

Also, we’ll tell you a bit about what data was used to achieve the result. So that it doesn’t seem like some fairy dust was sprinkled!

Here’s the 5 examples at a glance:

Here’s the 6 capabilities:

1. Predicting employee attrition for HR

Aim

A HR agency wishes to help its clients figure out which employees are likely to leave.

The HR agency also wants to know the likelihood of a candidate they are recommending leaving within a short period of time (as the agency has to replace the candidate).

Rationale

An employee that leaves your company is very expensive. Especially good employees, and those that you have invested time and resources to train.

Again, retaining an existing employee is far more cost-effective than acquiring a new one.

Having an accurate model to predict which employees will leave a company, when they may do so, and also why, allows you to target those underlying causes, and work to prevent or minimize employee attrition.

In order to build a prediction model, we needed to provide some variables (or what we call “features”), so that the model can learn to predict churn based on these features.

Data set (46 variables)

  • Profile
    • Age
    • Marital status (single, divorced, married)
    • Distance from home
    • Department (sales, R&D, HR, engineering, etc.)
    • Education level (primary, secondary, polytechnic, …, PhD)
    • Gender
    • … other profile-related data …
  • Job
    • Job involvement (score 1 – 10)
    • Job level (score 1 – 10)
    • Job role
    • Job satisfaction (score 1 – 10)
    • Average overtime hours per day
    • Performance rating (score 1 – 10)
    • Has standard hours (binary)
    • Has business travel (binary)
    • Business travel class (economy, business, first)
    • No. of direct reports
    • Has flexible hours option (binary)
    • Has remote work option (binary)
    • No. of days of paid annual leave
  • Compensation
    • Monthly income
    • Last performance bonus amount
    • Last AWS amount
    • Has stock options (binary)
  • History
    • Number companies worked for previously
    • Total working years till date
    • No. of years worked at company
    • No. of years in current role
    • No. of years since last promotion
    • Number of managers under since commencement
  • Satisfaction
    • Has public-facing achievements
    • Office environment satisfaction (score 1 – 10)
    • … other satisfaction-related data …

Results

After testing, we developed and settled on a neural network, such that it was able to predict a score, which represents the likelihood that a given employee would leave.

Scoring accuracy: 78%

After which, we interrogated the model by simulating different employees, to find the factors which are more significant towards them leaving.

2. Customer targeting for an insurance agency

Crop businessman giving contract to woman to sign

Aim

An insurance agency wants to know who are the most likely customers to buy (and re-buy) a life policy them, and what value of policy they are likely to take up.

They can then target these customers with marketing campaigns, promotions, and “upgrade” benefits.

Rationale

A customer targeting model helps you identify the people who are most likely to buy from you, so you can focus your marketing resources at those most likely to become your customers.

This can be extended to any product, as well as product upgrades.

Data set (69 variables)

  • Profile
    • Age
    • Gender
    • Marital status (single, divorced, married)
    • No. of marriages
    • No. of children
    • Age of each child
    • Education level (primary, secondary, polytechnic, …, PhD)
  • Income
    • Average annual income
    • Department (sales, R&D, HR, engineering, etc.)
    • … other income-related information …
  • Assets
    • Average amount saved per year
    • Amount of savings
    • No. of property owned
    • Average value of property owned
    • … other assets-related information …
  • Liabilities
    • Amount of bank loans instalments unpaid
    • … other liability-related information …
  • Campaign
    • Number of times the client was contacted during the campaign
    • Was client contacted in a previous campaign
    • Number of days since the client was contacted in the previous campaign
    • Outcome of previous campaign
    • … other campaign-related information …

Results

After testing, we developed and settled on a support vector machine (SVM), such that it was able to predict a score, which represents the likelihood that a given person would purchase a life policy at this point in time.

Scoring accuracy: 92%

We also developed and trained a neural network which predicted the dollar value of the life policy the person would likely purchase at this point in time.

Scoring accuracy: 76%

3. Credit risk assessment for a bank

Buildings With Glass Windows

Aim

A bank wishes to be able to predict the likelihood a bank customer is going to default on their loan repayments, and therefore decide whether or not to grant a loan.

The bank uses this to grant instant loan approvals when either an existing customer, or a new customer, applies for a loan.

The bank also wants to know the likelihood of a loan applicant defaulting, rather than a simple yes/no on whether the model thinks a loan applicant will default.

This allows the bank to assess the predicted risk.

Data (18 variables)

  • Profile
    • Age
    • Marital status (single, divorced, married)
  • Loan application
    • Loan amount
    • Loan duration
  • Customer status
    • New to bank
    • Existing
  • Income
    • Average annual income
    • … other income related information …
  • Existing loans
    • No. of existing loans
    • … information related to late- or non-repayment of existing loans…

To keep things brief in this post, we did not include every single variable.

Results

After testing, we developed and settled on a neural network, such that it was able to predict a score, which represents the likelihood that a given loan applicant would default.

Scoring accuracy: 83%

4. Predicting customer loss for a telco

Man With Luggage on Road during Sunset

Aim

A telco wishes to know which customers are most likely to switch away from them to a competitor, and thereby offer loyalty benefits to such customers.

Rationale

It costs you more to bring in a new client, as opposed to keeping an existing client.

If your business can understand what triggers clients to leave, it gives you something very specific to act on.

For instance, you may consider loyalty programmes and customer retention campaigns that directly target the reason customers leave in the first place.

Data (72 variables)

  • Profile
    • Age
    • Gender
    • Nationality
    • Estimated salary
    • … other profile-related data …
  • Relationship
    • Duration of being the telco’s customer
    • No. of family members with telco
    • Previous churn (binary between 1 if the customer has left the telco before, 0 otherwise)
    • Customer service satisfaction score (on a scale between 1 if the customer has never interacted with customer service or is completely satisfied, and0 if the customer is absolutely dissatisfied).
  • Spending
    • Post-paid vs. pre-paid
    • Mobile phone purchased with contract (vs. sim-only)
    • Average monthly spending
    • Current plan with telco
    • … other spending-related data …
  • Usage
    • Talk time in {morning, afternoon, evening, night}
    • SMSs sent in {morning, afternoon, evening, night}
    • Data used in {morning, afternoon, evening, night}
    • International calls
    • Voice mail
    • … other usage-related data …

Results

After testing, we developed and settled on a neural network, such that it was able to predict a score, which represents the likelihood that a given customer would leave the telco.

Scoring accuracy: 87%

After which, we interrogated the model by simulating different scenarios which the telco identified, to find the factors which are more significant towards customer loss.

This allowed the telco to act on those key factors, and develop campaigns and programmes and communications to turn it around.

5. Predicting donations for a non-profit

photo of flight of red and white rescue helicopter during snow daytime

Aim

A non-profit organization wants to understand which segment of people are more likely to donate money to their cause, and adjust their requests and appeals towards those groups.

The non-profit has a challenge. For some data, not all the variables are “complete”. This is because the non-profit cannot demand such information from people donating. It only has data that people give willingly.

Therefore, the information is available “best effort”, and the model needs to be able to deal with such missing information.

Data (17 variables)

  • Profile
    • Age
    • Is employed (binary)
    • Marital status (single, divorced, married)
    • Job (including student)
    • Average annual income
    • … other profile-related data …
  • Donations
    • Days since last donation
    • Days since first donation
    • Average days between donations
    • Last donated amount
    • Total number of donations
    • Did donate in last donation campaign

Results

After testing, we developed and settled on a neural network, such that it was able to predict a score, which represents the likelihood that a given person would donate.

Scoring accuracy: 86%

Conclusion

We’ve explored 5 different, but very useful and common situations, that AI or machine learning can bring actionable insights to you and your business.

These examples provide a glimpse of the versatility of using machine learning, and how businesses leverage the power of AI to:

  • Increase revenue (e.g. understand target market)
  • Decrease cost (e.g. reduce churn)
  • Increase capabilities (e.g. attract great employees)
  • Increase efficiency (e.g. automated pricing and quotations)
  • Drive ROI (e.g. marketing campaigns)

We’ve focused these examples on how your business can achieve insights, but that is by no means what machine learning is limited to.

If you have data, or are willing to start to build your business around having access to good data, you will quickly be able to leverage machine learning to derive insights that can bring incredible clarity to your business, your customers, and your employees, among many other applications.

In almost any business, and in pretty much an industry, the ability to predict an outcome based on a set of variables, gives you the edge over your competitors, and guides you in the direction to grow your business.

Eugene Ching Founder of Qavar, an AI and cybersecurity company. We use machine learning to bring insights into your business, and defend you against digital threats.

Don't miss out. Find out how leveraging AI or automation can help you.

Subscribe to receive practical tips, advice and ideas on how AI, machine learning and technology can help you grow your business.