Harnessing AI to Become Frictionless. Part 1: Simpler, Smarter IVR Systems

I have hesitated to propose how to catch the shooting star that is the combination of Artificial Intelligence/Machine Learning (AI/ML), Large Language Models (LLMs), and ChatGPT with its Generative AI (GenAI), but there are some powerful applications available to help organizations to become frictionless¹. 

In this first in a series, I will lay out how to use these tools to simplify IVR systems. I’m starting here rather than chatbots and other channels since most organizations are still relying on IVR, albeit delivering poor customer experiences by introducing friction. I say this as someone who started working with IVR in 1986 and whose MCI Call Center Services division in the 1990s built thousands of IVR systems for MCI business customers to use with their consumers. I’ll work through these sections:

• Quick definitions of IVR.
• Acknowledging and determining how IVR is causing friction.
• Harnessing the power of AI/ML, LLMs, and ChatGPT for IVR to become frictionless.

Quick definitions of IVR

First off, let me split IVR into routing and resolution, using touch tone or voice recognition. Some IVR systems are only set up to route customers to the best qualified agent, either with prompts (“Press 1 billing”, etc.) or by inviting open-ended questions.

Other IVR systems are set up to answer customers’ questions, for example in response to “I don’t agree with my balance?” or “Where’s my stuff?” or “Where’s the closest shop that’s open now?”, either with prompts (“Press 1 for your balance”, etc.) or by inviting open-ended questions.

Acknowledging and determining how IVR is causing friction

Next, let’s tick off how IVR systems are frustrating customers. How many of these are you hearing from your customers? For routing IVR, you can determine this by tracking the percentage of transfers from the first agent to the agent who can help the customer², calculating sentiment score during the agent interactions, surveying them, collecting social media posts, and analyzing verbatim comments. I’ll express this as you might hear it from customers with some explanations in the parentheses.

• “You don’t recognize me and my contact or purchase history” and “You don’t know me” (both signaling the IVR – and company – being “dumb”).
• “I can’t find the option that I need” (Logic doesn’t offer required options).
• “I keep having to repeat myself or fish around” (Spoken inputs such as accents or terms not programmed into the IVR).
• “I’m trapped!” (No way to escape to an agent).

Harnessing the power of AI/ML, LLMs, and ChatGPT for IVR to become frictionless

How can these exciting tools remove this IVR friction? Let’s imagine these two customer journeys: (A) “Where’s my stuff?” and (B) “I don’t agree with my balance, again.”

A. “Where’s my stuff?”

• Sonya has been waiting for her partner’s gift-wrapped birthday book to arrive in three days, after getting reassurances that it would be on time.

• She calls the 800 number from her mobile phone³.

• The IVR recognizes Sonya’s ANI and opens with “Greetings again, Sonya. You are probably calling about your recent book order. We are experiencing delays with our gift-wrapping and will not be able to deliver it on time this coming Thursday. However, we will send a complimentary roll of gift wrap paper that you had requested. If that’s OK with you, please say “Yes” or press the 1 button on your phone. If not, please say “No” or press the 2 button.”

• Sonya really wanted it wrapped and had paid for it, so she says “Operator!“

• The IVR replies with “Sorry that we disappointed you. I will route you right away to our escalation team.” [instead of offering other options and keeping her in the IVR]

• The escalation agent connects immediately, opening with “Sonya, I am so sorry that we told you that we can’t deliver the book wrapped to you on Thursday, but I’ve sent a message to my colleague in the warehouse to move it to the top of the queue, and it will be there in time.” The IVR could do this, too, but talking with the agent will probably satisfy Sonya who will trust it more readily.

Which of these new tools were used? Using GenAI and LLM, the company matched Sonya’s ANI with her purchase history, extracting salient details including order specs and timing, as well as the conditions behind “missing the promise”⁴ with an alternative, the roll of wrapping paper. Then, using AI/ML when Sonya rejected this alternative, behind the scenes the tool will tee up something else the next time. The IVR then connected with the CRM to pass Sonya’s request, rejection, and exit state so fast that the AI used agent assist to send the request to the warehouse and provide the script for the agent.

Are these tools available today? Yes, albeit with some precise use of APIs across disparate systems.

B. “I don’t agree with my balance, again.”

• Rob reviews his telco’s monthly invoice and he sees that the same extra charges are on it, despite having emailed the company last month.

• Rob calls the 800 number from his office landline.

• The IVR can’t associate this ANI to a unique customer or contract, so it opens with “Welcome to Fast Telco. Please provide the last 4 digits of your account number or your mobile number, using the keypad or by saying them.”

• Rob enters his mobile number, which the IVR connects to his account, and says “Greetings, Rob. It looks like we didn’t apply your $15 credit to this month’s invoice, but your next invoice will show that credit and another $15 credit and $20 more with our apologies. The total credit will be $50. We are sorry for not getting it done in time. If that is OK,please say “Yes” or press the 1 button on your phone. If not, please say “No” or press the 2 button and hang on for our confirmation.”

• Rob says “Yes“.

• The IVR concludes with “Thanks, Rob. We will send a written confirmation to your email address for a $50 credit on your next invoice. Can we the number you are calling us from to your account? We won’t contact you on it unless you tell us.”

• Rob replies “Please add my office number“.

Which of these new tools were used? As with the other use case, here multiple tools are used … GenAI and LLM to add Rob’s ANI to his account, then extract salient details including billing and contact history plus the reason behind the delay. Then, using AI/ML when Rob accepted this alternative, behind the scenes the tool will tee up the same offer the next time.

Are these tools available today? Yes, again with some careful development.

Stay tuned for Parts 2 and 3 to harness the power of AI/ML, LLMs, and ChatGPT to become frictionless!

Notes

¹ As introduced in our latest book The Frictionless Organization: Deliver Great Customer Experiences with Less Effort (Bill Price & David Jaffe, Barrett-Koehler, June 2022).

² Transfers caused by misrouting ought to be <10%.

³ This use case hinges on already registering and being able to match the customer’s phone number or ANI (Automatic Number Identification).

⁴ Amazon considered accepted customer orders as “promises”. Whenever Amazon realized that it couldn’t deliver the order placed as the customer requested it, and as Amazon thought it would, a “missed promise” path kicked in. This is from my years as Amazon’s 1^st WW VP of Customer Service.