We can finally deliver the promise of improving customer service.

Dawid Naude
3 min readDec 22, 2023

AI in call centres, customer service, back office

AI can finally deliver on the promise of automation massively transforming the call centre/ service experience… and therefore the customer experience. Here are 12 areas and some ideas. All of these are possible today.

1. Understanding the caller question to direct it to the right agent.
2. Getting the right agent on the call
3. Making sure the agent is trained enough for the call

Understanding the problem
4. Understand the existing customer profile across 6+ systems
5. Understanding the customer problem through conversation (knowing the right questions to ask)

Finding the right solution
6. Finding the right solution
7. Getting extra support if needed

Taking action
8. Finding the right forms, systems and processes to complete the transaction
9. Completing the processes

Wrapping up
10. Sending information to the customer
11. Summarizing the call steps
12. Reflecting on what could’ve been improved (this is impossible given the next call is seconds away)

Here is how AI completely changes that interaction.

  1. Understanding the call question.
  • From: Terrible voice bots “I’m sorry, I couldn’t catch that, could you repeat it”.
  • To: Doing a detailed fact find via an AI driven bot whilst you wait for the agent.

2. Getting the right agent on the call

  • From: Basic skills routing via a skills table matched with the type of issue
  • To: Matching based on previous conversation history transcripts where agents have successfully solved issues with high satisfaction

3. Making sure the agent is trained enough for the call

  • From: 9 weeks classroom and shadowing training (at which point at least 30% leave)
  • To: Simulated customer calls with intelligent customer bots from day 1.

4. Understand the existing customer profile across 6+ systems

  • From: Partial data consolidated into one key system, but still needing the user to read through multiple tabs and records to understand.
  • To: A simple 3 sentence summary of the entire customer position, noting key recent events and likely what they’re calling about. (Combine this with the fact find whilst the customer is on hold and suggest resolution would automatically be shown)

5. Understanding the customer problem through conversation

  • From: Navigating to a knowledge article, which may contain some of the questions to ask, but likely it’s nested in a few articles.
  • To: Having the system suggest other questions based on listening and understanding the conversation.

6. Finding the right solution

  • From: Searching for a knowledge article, going through multiple articles and finding bits and pieces to sort through the resolution
  • To: Automatic creation of most likely resolution, sourcing information from previous solution summaries, combined with knowledge articles.

7. Getting Extra Support If Needed

  • From: Putting the customer on hold or transferring to another agent, requiring the customer to repeat everything.
  • To: AI reviews the context of the existing problem, matches it with an agent who has successfully solved similar issues, and pulls them into the call automatically as soon as available.

8. Finding Forms and Completing Processes

  • From: Navigating through systems, forms catalogues, PDF forms, mailboxes.
  • To: Dynamically suggesting the appropriate form (e.g. Refunds) based on the call’s context.

9. Completing the Processes

  • From: Manually completing the form.
  • To: Form automatically populates based on the call transcript to date.

10. Sending Information to the Customer

  • From: Either no summarization of the call, manually typed up or sending meaningless automated messages.
  • To: Rich summary texted and emailed to the customer, providing clear and concise information.

11. Summarizing the Call Steps

  • From: Manual typing for 2 minutes afterwards, risking the inclusion of inaccurate steps or omission of important ones.
  • To: 99% of the call summary automatically generated, ensuring accuracy and completeness.

12. Reflecting on What Could’ve Been Improved

  • From: This step is often not done due to time constraints.
  • To: Automatic quick tips or advice shown after the call. If it’s a highly successful conversation, this advice is sent to be included as new training material.

All of the above could theoretically be implemented from today in a combination of voice recognition and transcription, large language models and API’s.

I’d pick one and start there, run a lightweight, ugly experiment with a control group, then another. You don’t need to do major surgery on any systems to achieve this, it can be lightweight, open-source and api driven.