Every customer support helpdesk is rushing to release a suite of native AI features to keep up with the market.
Zendesk is no exception.
Here’s a critical assessment of how Zendesk’s AI agents feature works, its benefits, and some best practices to consider.
What are Zendesk’s AI agents?
Zendesk AI agents are bots designed to resolve a range of customer queries.
Unlike traditional chatbots (like Zendesk’s former bot builder functionality), these bots can be powered with generative AI, making them able to identify and answer various customer questions, from simple to complex. The goal is to solve as many issues as possible from end to end.
Most bots can handle a customer requesting to return a product. But what if that customer doesn’t have a receipt? That would usually get escalated to a person.
The new AI agent can collect information from the customer to search for that order in the database, suggest alternative products they can swap the item with, and confirm the new order. Here’s an example of Zendesk demoing this in practice.
The benefits of using Zendesk AI agents
AI agents are a new and improved version of the normal customer support chatbot, which has been around for years.
Most customer support teams implementing a self-service solution like that are looking to:
Improve response times. An AI agent can generate an immediate response around the clock with no extra cost. They can also help streamline ticket routing and reduce the volume of tickets the team has to handle, improving response times across the board.
Enhance customer satisfaction. Quick and accurate responses lead to happier customers. If AI solutions can truly resolve the customer’s question instantly, they’ll be more satisfied.
Increase efficiency. Automating repetitive tasks allows human agents to focus on more complex, high-value interactions, boosting productivity and morale.
Scale support. As businesses grow, the volume of customer inquiries typically increases. AI agents can handle many interactions simultaneously without increasing staffing costs.
Best practices for using Zendesk AI agents
We’ve gone through the process of setting up an AI agent from scratch in Zendesk.
Many of its features are intuitive. Zendesk built AI agents to enable their customers to launch a bot quickly, which is a huge plus if you have a simple product paired with a robust knowledge base.
It can be riskier with a more technical product, though. Quickly setting it up and leaving it might mean that you aren’t getting the most value.
These are the six key steps to setting it up well:
Experiment with AI personas
Design the AI agent’s behavior
Input translations manually
Ensure your help center is up-to-date
Review and assign intents
Create answer flows
Experiment with AI personas
When you first set up the AI agent, Zendesk prompts you to choose a persona. You can opt out of using a persona, but that might result in your bot sounding a little more robotic than it needs to.
These are the default and pre-selected settings.
What’s simultaneously good and bad about this feature is that you have very few options.
That means you don’t need to experiment for hours to understand exactly how each selection impacts the way the bot will communicate. It’s easy to select the one option that’s closest to your brand voice.
It also means you’re restricted in the types of personas you can build. There’s no way to input specific instructions or prompts about communicating with customers. And there might be nuances in other languages (such as formal and informal versions of the second-person pronoun “you”) that you can’t account for in the persona.
Design the AI agent’s behavior
Zendesk sketches out six simple flows for the AI agent.
You can keep the default responses or edit each one to match your brand’s tone of voice more closely.
Again, the simplicity here is a big plus. Zendesk’s interface allows you to configure and modify the AI agent’s behavior without needing advanced technical skills, which makes deploying an AI agent accessible for teams with limited resources.
The only major restriction is that the bot will only answer one intent at a time, even if it recognizes multiple intents, which can result in a choppy experience.
Zendesk’s standard recommendation is to state that this is a bot (mention it in the name, include it in the communication, and so on). That way, you can set your customers’ expectations accordingly.
This is a big win for Zendesk and its approach to implementing AI solutions. Some solutions try to avoid dealing with a potentially negative response by simply hiding that the response is AI-generated.
Input translations manually
For companies offering multilingual support, Zendesk offers a flexible solution.
You can implement the bot in multiple languages and select these in the settings.
Enabling languages automatically translates every intent, answer, and interaction you input. But you can always overwrite these with manual translations in most languages.
This level of control means you can ensure a high-quality experience. At the very least, you can double-check translations for the important pieces of text (like how the bot starts a conversation) or for any common questions.
Ensure your help center is up-to-date
As always, any AI solution that integrates with your help center will only be as good as the help center itself. You need to make sure that your help center is up-to-date.
These are the options for how the bot behaves if it finds a relevant article.
In theory, you can choose “Don’t answer based on articles.” That’s the fallback option if your articles are out of date or if you don’t have a help center to begin with.
It still isn’t a great alternative.
A comprehensive help center is essential for your customer experience–many customers might engage with a help center without ever interacting with your support team or any chatbot solution you implement.
Writing help content that works for generative AI will ensure that your customers can find the right answers if they ask ChatGPT or Gemini for the solution rather than visiting your website.
Review and assign intents
Zendesk automatically generates a list of intents to help you categorize and group customer queries.
Intents are the best way to provide individualized responses (not based on help articles) or identify cases where you don’t want the bot to generate responses.
Because intents don’t have to be set up before you pilot and launch a chatbot, you can wait and see how people respond to the chatbot and start reviewing intents only after you collect some data to see how customers interact with it.
Intent classification is one of the biggest challenges of implementing a chatbot:
Customers can phrase their queries using different words. Zendesk would likely generate multiple intents, which you can group when you generate the answer.
As you can see from the screenshot above, some intents are generic. “Payment or checkout issue” could mean anything. In these situations, the bot's answer is built on asking clarifying questions first.
Bug identification is also extremely difficult for an AI solution. If you have a common question about a feature and it turns out that feature is broken and not working as described, the AI can generate the wrong answer many times before it catches there’s a difference in intent.
Not all customer queries will fit neatly into predefined intents. This is also exacerbated by a more complex product, where there might be a ton of small, niche questions that get asked once a month.
There are some ways to improve intent classification over time:
Start with broad, general intents that cover a wide range of common queries. This approach allows the AI agent to handle most interactions initially and gather data to refine and add more specific intents over time.
Regularly review the data collected from customer interactions to identify new patterns and refine your existing intents.
Establish a feedback loop where customers and support agents can provide input on the AI’s performance. It helps to have your agents review every interaction, especially at the beginning.
Create answer flows
The final fallback option to ensure that intent classification doesn’t result in a negative support experience is to design your answer flows with that in mind.
This is an example of the default answer flow if the customer chooses to talk to a human.
This feature embeds a mini-rule-based chatbot into the AI agent. It’s great for circumstances with specific user flows or if there are multiple potential outcomes that need precise handling.
You can use it to collect information from the customer based on their request and then design a flow that either escalates to your team or tries to resolve the case directly. Because Zendesk can also connect to your API, this can be a powerful feature that drastically increases the number of cases handled from end to end.
Zendesk’s AI agent pricing model
Zendesk will start charging for AI agent usage per resolution in August 2024. This is becoming standard in the AI space, and it has a few major advantages:
A per-resolution payment model makes it easy to compare costs directly with your current cost per ticket. This makes it easier to justify the investment financially and measure its return.
It provides financial flexibility and scalability. If the AI implementation is not performing optimally, the lower number of resolved tickets translates to lower costs, offering some financial buffer during the initial stages of implementation.
It creates an incentive for the AI provider to ensure that their systems are efficient and effective.
The main disadvantage of this model is that it misinterprets deflection as resolution. In some cases, a lack of customer follow-up will be counted as a successful resolution when, in reality, the customer simply gave up.
The worst-case scenario is paying for a ton of resolutions that ultimately result in customers bouncing because they were frustrated with the experience. That’s largely why closely monitoring the responses is important.
Boost your customer support team’s efficiency
Zendesk’s AI agents are available on all Support and Suite plans right now, so you can already experiment and see what the setup process looks like. It’s a pretty powerful feature that provides a significant opportunity to boost your customer support team’s efficiency.
Another way to enhance your support operations is to integrate advanced and dedicated tools like Swifteq. We’ve built a range of Zendesk apps that can help you optimize support workflows, streamline ticket management, and improve your help center–all of which pair well with an AI strategy.
Written by Nouran Smogluk Nouran is a passionate people manager who believes that work should be a place where people grow, develop, and thrive. She writes for Supported Content and also blogs about a variety of topics, including remote work, leadership, and creating great customer experience. |