Ada Support

Conversational AI best practices: voice automation for customer service

Selin Tamer
Conversational Interaction Designer
AI & Automation | 21 min read

Phone support is still the most popular and most expensive customer support channel. Companies around the world are spending over $1.3T annually on customer calls. Needless to say, this is a channel that could benefit from automation powered by conversational AI.

Voice automation helps companies serve customers quickly and efficiently, reduce average wait times, save contact center costs, and provide a quality phone support experience.

"Voice automation is beneficial to both your customers and agents. We know that 63% of customer-facing teams say it’s challenging to find the time to build meaningful customer relationships. Leveraging AI to resolve repetitive tickets not only speeds up service for customers, but frees agents to devote more time to the high-impact conversations that are so important.” – Emily Gregor, Senior Content Marketing Lead, Aircall

But if you think you can port over all the conversation design principles you’ve learned so far, think again. There are special considerations to keep in mind for voice automation, some of which are different from the best practices on chat and messaging channels.

Overall, it's important to consider the limitations and challenges of phone interactions, such as decreased customer patience, tricky information capture, more unhappy paths, or caller demographics. Thoughtful design and scripting can help address these considerations and create a successful automated phone support experience for callers.

Without further ado, let’s dig into some of the conversational AI best practices we teach our own clients regarding voice automation for customer service:

  1. Define your user persona
  2. Craft an inviting Greeting
  3. Write right
  4. Pause strategically
  5. Hand off at the right time

1. Define your user persona

The user persona is the profile of the person who will be using the phone to get in touch with your support team — i.e. “the caller”. Keep in mind that there’s a good chance that the caller is different from the user persona you may have defined for chat or messaging channels.

Understanding the caller and defining the parameters of this persona will guide your voice automation efforts and help you build more successful flows with higher resolution rates. Ask questions like:

  • Will the caller be comfortable speaking with an automated voice assistant?
  • Will they be calling from a landline/able to receive SMS messages?
  • Will your callers be speaking in dialects or using terminology that you should accommodate for in your training? For example, older callers tend to ask for escalation by saying “Operator”, which is probably not a word you have trained in your handoff answer yet

2. Craft an inviting Greeting

The Greeting is the opening message that callers hear when they call the voice chatbot number or when they are transferred from an IVR to the voice automation platform. Getting the Greeting message right will invite callers to the conversation and guide them on next steps.

You can craft a great Greeting by following these steps:

  1. Inform the caller that they are speaking with a conversational AI. You can also tell them how to speak to the bot — for example, complete sentences or short phrases.
  2. Notify the caller that the call is being recorded. There are many jurisdictions that have privacy laws that require you to notify customers of recorded calls, such as California’s Invasion of Privacy Act. 
  3. Account for “hello” interruptions. Consider the experience if a caller says "hello" part way through the greeting — which is more likely over voice than messaging). It’s often a good idea to have a Secondary Greeting Answer set up with variations of “hello” to prevent repeating the original Greeting message in these instances.
  4. Provide the caller with clear instructions on what to do next. Asking an open-ended question like “what can I help you with?” may be jarring for some callers. You can guide them by suggesting questions they can ask or tasks that the bot can help them with.
  5. Prompt callers to give the voice chatbot a chance. Sometimes, particularly with older demographics, callers may be resistant to speaking with a bot and will either hang up or ask to speak with a human immediately. This can affect your automated resolution rate and your ROI as a result. You can use your Greeting to get callers to try the bot by highlighting your highest volume drivers, saying 
    something like "I can help you with X, Y, or Z”.

We also strongly encourage creating a “Give the bot a shot” Answer. You’ll notice that many more people ask to speak with a live agent at the Greeting level in Voice than they do in Messaging. This is very typical caller behavior as most people who pick up the phone to contact support are expecting to speak with a person to resolve an issue that they believe requires a human touch. By defining the intents your Voice bot is equipped to handle, you will inspire more confidence in the bot’s abilities and encourage your caller to cooperate with the bot.

3. Write right

Phone calls require more attention from customers than chatting does. They’re either solely focused on the call and want to get through it as quickly as possible, or they’re multitasking and you only have a fraction of their cognitive processing abilities. Either way, brevity and clarity are the keywords when writing conversational AI flows for voice automation.

Keep the information succinct

Conversational AI flows should get to the point quickly and allow users to dig for more context if required. Shorten the information you deliver and add a prompting question to the end.

  • Before: You’re all set! Your plan has been changed and you should see the changes in your account within 2 hours. We also have a promotion on data-roaming. Would you like to hear more?
  • After: You’re set! You should see the changes in your account within 2 hours. Would you like to hear about our data roaming promotion? 

Reduce cognitive load wherever possible

Cognitive load is the amount of mental effort required to process information. As you’ve gathered by now, conversational flows for phone support have to deliver information with as little cognitive load on the caller as possible.

Writing succinctly will get you a long way towards that goal. Here are a couple of other tips to keep in mind too:

Send answers with heavy cognitive load to other channels. Sometimes an answer will inevitably contain a lot of information that is difficult to remember, such as an address or an order number. Wherever possible, it’s better to send this information to a text channel instead, such as SMS. This way the caller will not have to worry about looking for a pen or remembering a random string of words; they’ll have the information they need on hand whenever they want.

  • Before: In order to exchange your item, you will need to go online to and enter all of your information. Make sure you have a copy of your receipt and the purchase date on hand, as well as the reason for your exchange. 
  • After: You will need to go through our return portal online to exchange your item. I’ve just sent you an SMS with all the information.

Don’t inundate callers with too many options at any one time. If you are prompting callers and want to give them multiple options to choose from, stick to 3 or 4 and offer them the ability to hear more if they want to.

Make sure callers know how to navigate their way through the automated system. Callers will want to pause, slow down, and repeat the content that is read out to them. Make sure your build incorporates these commands and that your callers know what the magic words are to trigger them. While you will do your best in your prompt-design to ensure that callers know what information to provide and how to provide it, you will also inevitably have situations where your callers are stuck in a state where they’re unable to cooperate with the prompt that is presented to them. Make sure your callers know the magic words that can get them out of these situations. 

Follow Acknowledge, Confirm, Prompt rules

While this writing structure is a great format for conversation design on most support channels, it’s particularly important for voice automation.

  • Acknowledgement lets callers know ‘we heard you, and it’s our turn to speak’
  • Confirmation allows callers to know what answer they’re in. If the bot got the answer wrong, it allows them to barge in so they don’t have to sit through a lengthy answer to realize they didn’t get what they were looking for
  • Prompt keeps the conversation flow moving forward within guardrails that you can control in the build

Ensure complete clarity in your prompts

Whenever you are prompting callers, make sure that the prompt clearly indicates how they should respond or what actions they can take. The more clarity you provide in exactly what information you need, and exactly how the caller should input it, the stronger your build will be. Example:

  • Before: What is your date of birth?
  • Better: Using a complete date, such as May 23rd, 1990, what is your date of birth?
  • Best (verbal reply): Using a complete date, such as May 23, 1990, please tell me your date of birth.
  • Best (dialpad input): Using your dialpad, enter your date of birth in Year, Month, Date format. For example, May 23, 1990, would be 1 9 9 0 0 5 2 3.

Create sample dialogue to know what info you need to ask for

What is often helpful when determining how to build net-new voice automation flows is to see the use-case played out in dialogue. It becomes clear what information will be required to complete the happy-path of the flow, and allows you to call-out these requirements during the project scoping stage of implementation.

Once you’ve created your sample dialogue, remember to find a partner to read it back and forth with. It may sound silly, but you’ll be amazed at how hearing the intonation and speed of the interaction can highlight areas of content optimization.

Below is an example for determining what information is needed for an order exchange:

Use conversation-markers to communicate the progress in the flow

Add conversational markers to let callers know where they’re at in the conversation. Words like “first,” “almost there,” and “finally” help set expectations about how much longer this part of the conversation will last, or how many steps they may still have to go.

Create clear error messages

Keep error messages clear and informative to make sure the caller knows what went wrong and how they can retry.

  • Before: Oops, something went wrong! Would you like to try again?
  • After: I wasn’t able to find an order associated with your phone number. Is 555-555-5555 the phone number you used to place your order? 

Review and optimize

Review your copy and refine verbiage wherever you can. Remember to optimize for brevity and clarity. If you’ve already built out automated flows for messaging channels in Ada, you can copy these over to Voice and adapt them instead of starting from scratch. I’ll dive into some best practices for adapting Messaging content to Voice at the end of this article.

4. Pause strategically

Much like live verbal conversations, automated voice conversations must account for pauses. This is good practice when you need to collect information from a caller that they may not have on hand. There are some best-practices to keep in mind for length of pause when expecting input.

Let the caller know a pause is coming. If you need to collect an order number or a policy number, you can add a pause and let the caller know “I’ll give you a few seconds to gather this information before we start."

Plan out when to “timeout” the pause. There is an average length of time it takes between when the pause starts and when the caller ought to have started saying something. If that length of time passes and the caller hasn’t said anything, you should “timeout” the pause, i.e. say something to prompt the caller and make sure they are still on the line. 

Figure out the timeout lengths. Depending on what the task at hand is, the time needs to be shorter or longer, and you need to make sure you’re accounting for the right length so that the bot and the caller don’t interrupt each other. Here are some examples of longer or shorter timeouts:

  • Longer - grouped numbers: When callers read a number that’s naturally grouped, such as a credit card, they naturally pause between groups. You want the timeout to be longer so you don’t cut them off.
  • Longer - information recall: When callers need to find a piece of information, or recount a story, they’ll likely speak a lot or hesitate naturally. Here too, you want the timeout to be longer so you don’t cut them off.
  • Shorter - quick answers: When callers are expected to give short answers like “yes” or “no”, it’s better to set a shorter timeout to keep the dialog moving.

Use specific phrases when the timeout occurs. For example, rather than just “Can you repeat that?” or “I didn’t catch that” add something about the topic at hand like, “Can you please repeat your account number?” or “I didn’t catch that. What’s your account number?” You can also use these phrases to guide callers, for example, “Sorry, I didn’t get that. Your account number can be found at the top of your statement. Please say or type it in, or say, ‘I don’t know it.’”

You don’t always need to timeout the pause. Sometimes it can be appropriate to do nothing when there’s silence/when no speech is detected. This may be true when:

  1. There’s another way for the caller to move forward like a button in SMS or mobile.
  2. Their lack of response doesn’t break the conversation.

5. Hand off at the right time

With well-designed conversational flows, you’ll be automatically resolving most of your customer inquiries. But there are certain cases that it would make sense to hand off the conversation to a live agent instead. These are the main 2 to consider:

  • Repeating too often: In the event that the automated conversation is repeating too often, the bot should realize that it is looping and therefore not being helpful. It might be that the caller has a very unique case that’s better suited for an agent. The bot should hand off the call then.
  • Business rules: Some companies will want specific inquiries to be handled by specific departments; for example, you may want to pass along all cancellation requests to a loyalty department. You can set these rules however they make the most sense to you.

Adapting Messaging content to Voice

If you’re already using Ada, you can easily copy all the automated flows you’ve already built for Messaging over to . With a few adaptations, you can deploy your new Voice bot in a few hours instead of having to build everything again from scratch.

Here are some considerations for adapting Messaging builds to Voice.

Evaluate existing flows

Evaluate how many of the existing flows in the Messaging lead to automated resolutions and determine how much of that can be automated in Voice. If many flows require handoff to live agents, it may result in lower containment rates for the voice bot.

Information capture

  • Type of information: Consider what type of information needs to be captured for the voice bot to function effectively. Some information may be easily captured over the phone (such as a first name), while other types may require methods with more friction, such as SMS capture (such as an email address or confirmation number).
  • Amount of information: Keep in mind that callers may be easily frustrated and drop calls if there are multiple questions or steps involved in capturing information. Voice flows may need to be streamlined to minimize the amount of information that needs to be captured.
  • Phone number-based information capture: Consider using the caller's phone number to pull as much customer information as possible, rather than gathering multiple pieces of information through different methods such as voice, dial pad, or SMS. This can help create a more seamless experience for the caller and minimize the need for additional confirmations and validations.

Answer flows

  • Type of answer flows: Voice bots may need to provide quick and actionable flows, as lengthy instructions or complex URLs may not be valuable over the phone. Consider adapting the answer flows to be concise and easy to follow in a voice conversation.
  • Metavariables: If the messaging bot heavily relies on metavariables, the voice bot may need to be structured differently.

Unhappy paths

  • Validation and fallbacks: Voice flows may have a higher potential to travel down unhappy paths, so it's important to build in validations, confirmations, and fallback answers to guide the caller back to the happy path. Consider the need for additional checks and confirmations to ensure accurate information capture and minimize errors in speech-to-text transcriptions.
  • Caller impatience: Callers are generally more impatient over the phone, so it's important to account for scenarios such as caller barging-in during bot responses or requesting a live agent immediately. Thoughtful design and scripting can help address these challenges and keep the caller engaged with the voice bot.

This 2 minute clip has a good example of good barge-in design. Note the following:

  • “When you hear your option, call it out” being used for lengthy list of options, encouraging barge-in
  • “When you’re done there, say Main Menu” being used to educate the caller on how to navigate the automated voice system

Ensure Messaging copy is optimized for Voice before saving

  • Double check the terms used in Voice and in Messaging when copying over blocks. For example existing messaging can say '...please type in a question'
  • Remove all emojis from Voice content
  • Consider brand-specific terminology, like unusual product names, used in Messaging that may be harder for a Voice bot to pick up when said over the phone. For example, maybe you have products like "QuickExam" or "Model CX-90." To increase understanding, add these terms to your Bot Vocabulary to teach your bot to better recognize them and ensure the right answer is served in response.
  • Be thoughtful about how you ask questions to reduce cognitive load on the caller and keep the conversation moving forward successfully. For example, if a bot says “How would you like to pay today? You can say credit or debit card,” a caller is likely to start responding as soon as the bot says “How would you like to pay today?” Instead, the message should read “We accept credit or debit for payment. How would you like to pay today?”

Ready to replace your legacy IVR?

Power 24/7, personalized experiences with Ada’s award-winning AI-powered automation. Increase agent efficiency, decrease wait times, and resolve more phone inquiries at a lower cost.

Learn more