Enhancing Calls with Natural Language Processing

Intent bots in car insurance phone service

Role

UX Content Strategist

Industry

Fintech

Duration

July 2023 - Dec 2024

Voice bubbles showing a conversation rise above a phone.

A visual representation of the phone experience.

Identify the ask

Goal: increase the number of calls contained in the automated phone system ("IVR") by offering new self-service options.

I leveraged voice recognition technology ("voice bot") so the system could capture a range of queries and offer callers a more natural verbal interaction.

To drive self-service, I advocated for:
1. functionality that suited the caller's use cases
2. organization and vocabulary congruent with online resources and the caller's mental model
3. empathetic voice and tone

Problem to solve

I wanted to structure the experience like natural conversation by:
1. empowering and enticing callers to self-serve
2. accurately directing callers to the correct agent
3. delivering relevant information through an empathetic and supportive tone to satisfy confused or stressed callers

I had the opportunity to set the precedent for how our brand builds and documents future flows using voice bots.

Audience

Customers with automobile insurance agreements and auto dealers.

This IVR handles a wide range of intents across multiple products, but most calls concern the status of a claim or extending coverage. These are infrequent callers, unfamiliar with technical jargon and in a potentially stressful situation.

Sample conversation 1

Phone system:
In a few words, tell us what you need help with. You can say things like "start a claim" or "refund status." Or, to use your keypad instead, say "keypad."

Caller:
What's your website?

Phone system:
To give you the right website, what are you looking to do? Say: • cancel contract • extend contract, or • something else

Caller:

Extend my contract

Phone system:

If you have a mechanical or maintenance contract that is still active, talk to the dealer you bought your vehicle from to extend your coverage. We can't extend your coverage if you got it through Multiprotect, and we can't extend a GAP agreement.
Want us to text you a link to our website?

Caller:

Phone system:

If you're done, you can hang up. Otherwise, tell us what you need help with.

The first iteration of the Main Bot.

The first iteration of the Associate flows, with the bot and the voice-key flows.

Understanding the tech

I consistently met with the developers to design and document the experience in alignment with the technology's strengths and limitations.

Key takeaways:
1. The bot is easily foiled by background noise.
2. We can only pull certain types of data because we depend on a third-party database.

The voice bot performs two primary functions:
1. Identify the intent (the main reason for the call). The bot uses natural language processing to identify keywords and match them to predefined intents.
2. Specify the slot, which is a variant or subcategory of the intent, through follow-up questions. Again, the bot uses NLP to identify words.

Sample conversation 2

Phone system:
In a few words, tell us what you need help with. You can say things like "start a claim" or "refund status." Or, to use your keypad instead, say "keypad."

Caller:
Claims

Phone system:
Is this an existing claim?

Caller:
Yes

Phone system:
Which do you need help with? The claim status, sending documents, or something else?

Caller:
Documents

Phone system:
Is this for a GAP claim?

Caller:
No

Phone system:
Do you need a refresher on which documents to send us?

Caller:
No

Phone system:
Okay. Send your documents to scs@ally.com. That's scs@ally.com. You don't need to include any other info in your email.

Want us to text you the email address?

Caller:
No

Phone system:
If you're done, you can hang up. Otherwise, tell us what you need help with.

Understanding the business and the users

I became intimately familiar with the auto insurance products and how customers interact with them. This way, I could paint a complete picture of the user's journeys surrounding the phone call.

The phone experience cares for wide variety of callers and products, which means several questions are needed just to sort the call into the correct agent queue.

Iteration

Day 1
I built the existing functionality using the voice bot.

Day 2 and onward
Over many months, I added self-service functions. I refined the flow in response to business's and users' feedback and routing data. I revised work to maintain conversational flow while accommodating new functionalities.

A major enhancement was adding an authorization experience (i.e. logging in on the phone) so callers could access personalized account info. Read more about it in my project that discusses authorization.

A later iteration of the Main Bot, with self-service functions. The transition key depicts how the caller moves over to the voice-key (touch-tone) menu.

A later iteration of the Associate flow. The transition key depicts how the caller moves over to the voice-key (touch-tone) menu if the Bot can't understand them.

The Main Transitional VKM. The transition key shows how the caller gets to each point from the Main Bot.

The Identify Voice-key Menu is where a caller will start if they elect to leave the bot experience on the Main Bot.

Sample conversation 3

Strategy

Contextual awareness
I built variations of the content/flow to make human-like conversation—a contextual memory that accounts for what info the caller may have heard or provided. The bot avoids sounding redundant or oblivious, which motivates the caller's trust and engagement.

Supportive
I wrote content that keeps the the caller informed, explains what's going on, and is broken down into simple steps. To minimize a caller's uncertainty and maximize their autonomy, when the bot runs into an error, or must go down an unhappy path, it's transparent with the caller and provides alternative ways to move forward.

Integrated
I designed the queue routing (human agents) to align with the self-service intents, in case the user runs into trouble using self-service. The logic/information architecture is shared so the system doesn't ask redundant questions or send the caller to the wrong person.

Back-up plan
I built a touch-tone counterpart (traditional) for each bot flow, so the caller has the option to use familiar technology. The bot technology isn't infallible—if it can't understand the caller's verbal input, it will transition to the traditional/familiar (and more reliable) touch-tone phone menu. I also felt it was essential to offer callers the option to use the touch-tone menu from the beginning if they were not interested in using the voice bot.

Sample conversation 4

Phone system:
In a few words, tell us what you need help with. You can say things like "start a claim" or "refund status." Or, to use your keypad instead, say "keypad."

Caller:
I don't know what to say.

Phone system:
Okay, you can pick from these options: • Roadside assistance • Cancel contract • Claim • Sales support • Billing • Contract info • Enter an extension

Caller:
I just want to talk to someone.

Phone system:
You can say or enter a 7-digit extension now. Or, say "skip."

Caller:
Skip

Phone system:
Okay, let's get you to a team member.
Hang tight. An associate will be with you soon.