How AI is changing the economics of customer service

How and where is enhanced voice interaction being used?

Jon Aniano: At Zendesk, we’re largely concerned with customer-facing interactions. I would emphasise that voice is still a high-friction interaction. The experiences at their best are amazing, but at their median are still not so great, so you can call a company and the IVR will ask you to press one for sales or two for service. These experiences will be with us for a very long time, but voice AI agents are taking away a lot of the friction, a lot of the pain and a lot of the complication of the old-style voice experiences.

There is a lot to be done to clean that up, and Zendesk loves being a part of that. You can take an IVR experience from “press one…” to “how can we help you?”, followed by a fast and accurate answer or actually taking action, and that’s a big improvement in customer service.

Parts of the industry have been calling for the end of the voice channel for a very long time, but that’s just not happening. Voice continues to grow as a customer service channel, and is still the right place for high-emotion, high-impact or high-value interactions with customers. It can drive loyalty when it’s done right. It can solve things with at least the perception of more urgency than other channels. So it’s not going away.

Some industries, such as retail and e-commerce, have a very high potential for automation. If your AI voice agent can successfully handle enquiries about order and return status, you can create better and faster experiences for consumers, especially as there are many times when voice interactions are more convenient for them.

Even Zendesk customers in highly regulated industries such as financial services see the potential for this type of automation. There are high-friction, low-risk interactions that happen on the voice channel, and if they were automated, their customers would still love to call a toll-free number and get their problems solved. This also frees up human agents to handle high-value, high-impact, high-emotion interactions with less waiting time for callers.

We’re already doing this with text-based channels such as chat, messaging and email. So, the reasoning and the logic and all the technology to solve those problems is there. The only remaining question is how well can we deliver an automated AI voice experience? I think with the latest models, the language processing and the reasoning is at the point where you can deliver a better experience than those high-friction experiences.

The emphasis is moving from average handle time and first call resolution to successful resolution, customer satisfaction and other customer-side indicators. Did the customer get their problem solved? Is the customer happy? Changing those metrics changes behaviours and roles

There’s a big difference between everything pre-2024 and post-2024. Prior to GPT-3.5, we had no chance of delivering a great automated customer experience, and a lot of that older technology still exists. So, if you are talking to a customer service chatbot today, you can’t be sure that it’s using LLM technology. There are still legacy vendors and there are still bad experiences.

The latest generations of models from Anthropic and OpenAI, and even open-weight models, are so good that we can put together an agentic system – and this is what Zendesk AI agents does – that is better at knowing when it’s wrong or knowing when it doesn’t have the answer, and setting thresholds and guardrails around when it escalates.

Zendesk’s voice AI agents are in an early access programme and are expected to reach general availability by the end of the quarter.

Another unsatisfactory experience occurs when people turn to text chat to avoid long waits on hold, only to find the agent is handling so many simultaneous chats that they spend just as long waiting for a reply. Can AI help?

The same economics are at play with chat, messaging and email as they are with voice. Human customer service agent time is so finely rationed in the customer service world today that agents are expected to handle simultaneous and concurrent interactions to shave seconds from average handle time.

But that’s the wrong metric. When we can automate 80% of the resolutions, it doesn’t necessarily mean we’re going to shave 80% of the human time out of the system. Instead, we’re going to use that human time in a far less rationed way, so by the time you are talking to a human, you’re getting a really big slice of that human’s time, and you’re going to get a better outcome and a better result.

Waiting on hold or for a human agent to respond is a symptom of rationing, and automation can take the pressure off. In the past, we wanted to avoid escalation because customer service agents’ time was rationed. With robust automation, we can escalate much sooner because we have more human agent availability. The experiences that we’re putting in the market today for Zendesk customers are night and day compared with the previous experiences.

The metrics and roles for customer service teams are changing. The emphasis is moving from average handle time and first call resolution to successful resolution, customer satisfaction and other customer-side indicators. Did the customer get their problem solved? Is the customer happy?

Changing those metrics changes behaviours and roles. If we know we can automatically resolve – not just deflect – 80% or 90% of the inbound volume, we can do completely different things with the humans in the customer service world. For example, they can monitor borderline AI agent escalations, or assist AI agents when human judgement is needed, but without escalating the call.

So, we move from a rigid hierarchy of tier one agent, tier two agent and so on to a flexible pool of agents with varying specialities. There is also a new role for customer service experience architects to constantly review the automation path and its success rate.

10 years ago, deflection rates were in the range of 10% to 15%. Now we’re talking about full resolution rates of 70%, 80% or 90%, which completely changes the economics of customer service.

During the keynote, you mentioned the Quality Score feature, which is in beta. What will it do for customers?

In the past, it was impossible to review every single interaction that happens in your customer service operations. You couldn’t listen to all the recordings; you couldn’t read all the transcripts. So, everything was done via sampling. You would sample 5% or 10% of interactions to try to be statistically accurate, but ultimately, humans were reviewing a small sample and making decisions based on that.

Now, we can analyse every single customer service interaction through LLMs. With our customers, we are doing that in real time as the interactions are happening to provide real-time guidance. And we can do that after the fact and provide aggregated analytics and insights.

Our Quality Score is going to ship to all of our customers so that they have a quality score for every single interaction, and we think that changes the game. But we also go beyond that with our full QA [quality analysis] product, where essentially you can write a custom rubric for any type of interaction and we’ll run a quality analysis using that rubric for every single interaction that happens in the system, and we’ll aggregate those analytics and provide you with all those insights.

So, there is a move from reading the insights, thinking about them and making changes, to treating the insights as a massive body of content that can be used to generate recommendations about what should be done: changes you can make to your workflows to make sure things are going to the right people, trends in the words your customers are using and how your agents are reacting to them, and suggested changes in your procedures or your scripts.

Now, because we’re analysing everything, because we’re inspecting everything, we can surface actionable insights directly to agents or directly to the contact centre operations managers.

We’ve been talking about customer service, but how is this technology being applied to internal functions?

Zendesk has been in what we call employee service – IT help desk, HR help desk, legal shared services and so on – for a long time. Organisations of all shapes and sizes buy Zendesk and self-implement for customer service, and then their various departments use it to handle email enquiries, to build a web-based knowledge centre, and generally deliver employee service with the same high-quality experiences that they provide for customer service.

In the past couple of years, Zendesk has been actively marketing to these use cases, and has built a specific product on top of the core platform to handle them. Employees are people in the real world who have expectations around how customer service should be: how quickly they get answers, the manner of those answers, whether it is conversational, whether it is AI-powered. They bring expectations from the real world and expect their employer to match those experiences, and that’s where Zendesk really shines.

One difference is that communication channels may vary between customers and employees. An organisation might use WhatsApp and the web for customer service, but employee service might be better delivered over Slack or Teams, so Zendesk’s AI agents and Copilot have to operate in a Slack or Teams world.

Another is that AI agents for employee service should respect the relevant permissions, and that’s a stricter and more complicated issue than it is for an external help centre or a product knowledge base. So we’ve done a lot of work there. We acquired a company called Unleash that has a really strong permissioning model on internal content, and now we’ve applied that to our AI agents for employee service.