The state of AI customer service in 2026
AI customer service is now more of an infrastructure element than a mere experiment. More and more companies are willing to trust AI to handle tasks. That, in turn, frees customer care agents to tackle more complex issues that require a human touch.
A recent study shows that around 25% of contact centers have fully integrated AI into their daily operations, but 88% still only use it sparingly. This is further supported by the global AI customer service market hitting $15.12 billion in 2026, and is expected to more than triple by 2030.
Those are impressive numbers, yet there is still a large gap between contact centers that use AI in some form and those that have woven it fully into their operations. Most are still stuck in the pilot phase or think the AI customer service they have is good enough (it often isn’t and is just frustrating for customers).
That gap is the story of 2026, and it also explains the mixed customer sentiment towards AI. However, it’s not “mixed” because when it works, it works great, and when it fails, you can’t type “I want to talk to a real person” fast enough.
AI in support gets a range of reactions, but people tend to push back when it can’t actually help them.
The customer sentiment paradox
Let’s look at some figures that might not make sense right away and may even seem contradictory at first glance.
Nearly one in five people who’ve used AI for customer service saw no benefits from the experience. That’s much higher than when using AI for other general tasks.
Yet, 62% of customers prefer chatbots over waiting for a human response.
And still, 75% would rather talk to a human than to a machine.
These stats tell the same story, just from a different angle: if AI doesn’t help, it frustrates people. Speed, accuracy, and responsiveness are the driving factors behind these numbers.
When the answer is fast and accurate, trust goes up. When the bot falls into an endless loop and refuses to send a customer to a human agent, trust goes down. Hard. So much so that it might turn people off entirely from accepting customer service from AI.
What good AI customer service looks like
Before we go any further, it’s important to understand the types of AI companies usually employ to good effect.
There are four general categories based on their purpose, strengths, and weaknesses. This will help distinguish and clearly outline the use case. Many people think AI bots are a catch-all in customer service, but you’ll see that’s not the truth.
Customer-facing conversational bots
This is what many people think of when they hear “AI customer service.” A chat box that’s at first handled by AI instead of a human.
Customers interact with these bots directly (important distinction), and when executed well, they are great for handling routine tasks like billing questions, password resets, order status, account changes, etc., that don't necessarily require a human to explain them.
A good example of what a mature implementation of such a bot is our own customer-facing bot, Orbi. As Panos puts it:
A customer, when they want to pay an invoice, doesn't have to go to a human and get. That can result in a 15 to 20-minute live chat. Today, they can come on chat, talk to Orbi, and it will take them 30 seconds to pay an invoice. Both sides are happy.
However, when it comes to complex technical troubleshooting, emotionally charged interactions, or nuanced billing disputes, Panos makes it clear that a human is required.
Things like actual website troubleshooting and actual email troubleshooting, they do require human attention, and they will always require human intervention.
Forcing an AI chatbot to handle situations it shouldn’t only leads to a drop in trust and brand reputation. While it can help out immensely, it should never be left to tackle all questions alone.
AI-assisted human agents
If you’ve ever worked as a customer care agent, then you have no doubt spent a lot of time looking up knowledge bases, pricing information, policy documents, and so on.
Not only can that get tedious and repetitive, but it also takes precious time; in a live chat , you don’t have much of it. Tools like internal AI chatbots help alleviate this pressure by offering human agents what’s essentially an internal, interactive knowledge base.
At hosting.com, we have Pinky Bot, which reduces manual searching. As Panos explains it:
Instead of trying to memorize all the different pricing for all the different products, they can just focus on troubleshooting websites and fixing problems. That is what the customers actually want.
The customer doesn’t care how we fix their problem, just that it gets done.
We have found Pinky Bot to be extremely helpful given our extensive product portfolio. It also takes the pressure off new agents who previously had to memorize everything.
Quality monitoring and real-time sentiment detection
This is perhaps the most underrated case for AI in customer care. The chatbots get all the attention, but a well-trained AI that reads every chat and evaluates it based on politeness, helpfulness, and customer satisfaction is just as significant.
Quality control is still one of those things that requires a lot of human effort to maintain, and it’s usually done after the fact. An AI system trained for the purpose, however, can not only monitor and provide feedback, but also react in real time.
At hosting.com, we use such a system to prevent situations that could become frustrating for our customers. If the system sees that a customer is not happy with the responses they are getting, it can flag it for escalation to a senior staff member.
This system has helped us maintain the quality of our responses, provide the help our customers need, but it’s also allowed us to take better care of our own support team.
A study by Gartner and AllAboutAI showed that frontline chat rep turnover dropped by 43% when supported by AI. In combination with Pinky Bot, this has led to less repetitive and mundane work and fewer hostile customer interactions.
Ticket triage and intelligent routing
When AI can read all incoming chat messages and tickets and understand their intent, it can route them without human intervention. That’s especially valuable when the support team is spread across multiple departments, or when request volumes are high.
At hosting.com, we are actively exploring this for our ticket and email channels. Panos explains it like this:
Customers who talk to us on live chat usually have simpler questions. Customers who talk on a ticket or email have more complicated questions. Those require human attention.
When the routing logic accounts for channel differences rather than just topic categories, it can significantly reduce the need to manually move tickets around.
Should you use AI for customer service? A decision framework
One of the biggest things we want you to get from this blog post is that not every company needs AI customer service (or AI in general). We are sure you’ve encountered many examples when AI was forced into a product that didn’t need it.
To avoid that, we’ve put together four simple questions that you must consider before pulling the trigger on AI for your company.
Do you have the volume to justify it? There’s no magic number here as every case is different. At hosting.com, we handle roughly 500,000 conversations per month, which is an extreme case. For us, the ROI is justified. If you handle fewer than a few hundred, though? That may not be worth the setup and maintenance costs.
Are most of the questions you receive the same few questions? AI is best at handling routine, repeated questions. If the majority of your support volume is the same questions over and over, then AI will definitely help. If most of your interactions require judgment, context, or empathy, however, AI will struggle with them.
What happens when AI gets it wrong? It will, AI is far from perfect, and the consequences vary wildly by industry. Imagine AI giving the wrong answer for medication dosage or a legal entitlement. It shouldn’t be used for high-risk questions like these. Air Canada is a recent example of what happens when AI gives the wrong answer in a high-risk situation.
Can you guarantee easy escalation to a human? Don’t make it hard to talk to a real human. We are customers elsewhere, too, and it’s endlessly frustrating when the AI refuses to loop in a human agent. That’s easily one of the best ways to lose customers if that’s your goal. 89% of consumers believe AI should be an easy gateway to human interaction, not a wall.
At hosting.com, we’ve been building AI into our customer service operations for about three years now. Panos has been the spearhead for the efforts, from the first proof-of-concept to the systems that now process half a million conversations per month.
We’ve learned a lot in this time, and here are the insights we gleaned from our journey so far.
What sparked the necessity?
In 2022, our care team was growing fast, and the interactions across live chat, tickets, emails, and phone calls were scaling up. We already mentioned how Panos was trying to keep up with as many of those as possible every single day, but it was simply not sustainable.
It wasn’t the AI hype at the time that inspired us to implement it in our own systems. It was the operational necessity to automate our processes instead.
We started with the newly released (at the time) GPT-3.5 API. It could read live chat conversations and surface what was happening, without a human having to do it. It’s what marked the start.
From QA to real-time sentiment detection
Our initial use case was quality control. It evaluated each case across the criteria we mentioned earlier in this blog (politeness, helpfulness, and customer satisfaction), then fed the results into an internal review dashboard.
Such a task would have been humanly impossible without AI. Instead, the task now takes about an hour per week. However, we didn’t stop there, and the system evolved further.
We could use such a system to detect customer sentiment in real time, as Panos explains it:
If you're a customer and you express visible frustration, maybe your website is broken, maybe the domain name isn't connected, we know in real time. We can intervene: a fast escalation, a fast connection to a manager. Customers get more quality and faster support when they need it.
This led to the creation of an internal channel for cases that needed urgent attention due to negative sentiment. It allowed the team to focus on the customers who need personal attention then and there. Because of this shift in procedure and mindset, our CSAT scores now average 85% to 90% daily.
Orbi, the customer-friendly chatbot
The first generation of our chatbot was called Xerxes, but even back then, we knew we weren’t done improving the experience. Thus, Orbi came about, and it’s now what our customers interact with every time they open a new chat window.
Orbi excels at several specific areas, none of which push it beyond its intended capabilities.
Natural conversation with large memory context.
Account actions with zero hallucination.
Fast resolution of transactional requests.
It also has a better name, since “Xerxes” doesn't really sound inviting. That’s all customer-facing stuff, though. Under the hood, it’s much different than the first iteration. We’ve refined the instruction set, and the underlying model is better. Panos was candid about what it took to build it properly.
The architecture is extremely complicated. It does cost us a lot of money, obviously as a business, and a lot of development time. There were also a lot of caveats we had to make as trade-offs for this to happen. But it's an amazing tool.
The next step for Orbi is to extend its assistance to email conversations, which will require carefully building the routing logic and escalation paths before deployment.
Pinky Bot, the internal agent assistant
We already mentioned Pinky Bot earlier, but it's worth elaborating on, as it’s an amazing tool that helps our customer care agents daily.
Separate from our customer-facing bot, Pinky Bot connects to all our knowledge bases and legacy platform systems to serve as a single source of truth for agents. In Panos’s words:
A company like us needs a place called the single source of truth. That source is the same one the training team uses, the same one the AI internal chatbot uses for data retrieval, and where our team goes when they need to manually check something.
The practical effect is that agents no longer need to memorize all the information they need for their daily tasks. If they need something, they can ask the bot for accurate answers in seconds, while staying focused on troubleshooting and problem-solving.
AI in training
One application that rarely gets mentioned is AI-assisted trainings. For our team, Pinky Bot already makes onboarding much more straightforward.
Additionally, Kameliya, our Head of Global Training, has built an entire learning infrastructure using AI. It includes interactive quizzes, AI-generated video content, and a chatbot that asks agents deliberately tricky questions.
This is what closes the loop: well-trained agents perform better in live interactions, which improves the data the QA system uses to evaluate quality, which, in turn, improves training over time.
Where humans remain essential
Panos is unambiguous and firm about the boundaries. When it comes to technical troubleshooting, complex email cases, and emotionally sensitive interactions, humans will always be necessary.
AI doesn’t replace human judgement. Instead, it extends human capacity, and that’s what our AI customer care architecture reflects. Orbi handles what it’s good at and escalates quickly when it encounters a problem beyond its capabilities.
Common pitfalls that can be easily avoided
When it comes to AI customer service failures, most people think it’s because of the technology. The reality is that the technology (the AI bot) simply follows the instructions it was given.
Because of that, these failures are usually strategic, not technical.. You can’t expect an AI that’s implemented poorly to do well. So, here are the most widely spread implementation mistakes we’ve noticed.
Using AI as a wall: This is the most damaging implementation. Hiding your contact information, forcing users to interact with AI, and making it hard to reach a human operator all lead people to quietly give up on your company.
Stale knowledge bases: AI answers from whatever data it has access to. If you changed some of your policies or pricing, for example, but didn’t update the data your AI uses, it will provide customers with incorrect information. That’s worse than no information at all.
Deploying without guard rails: Legal liability for AI errors is real , and the consequences can be dire. Again, this isn’t a failure of the technology, but of the strategy. Companies must define which conversations AI will handle, and which it won’t. The AI must have boundaries well before it’s deployed.
No feedback loop: Improvement cannot come without proof. Companies that deploy AI but don’t pay attention quickly discover their problems through worse channels: customer complaints. By that point, the damage is done. The quality monitoring system must be in place before the chatbot processes even a single customer.
These are just four of the most egregious examples, but we are confident they are all very relatable (especially that first one). And as you can see, having a clear plan for before, during, and after the AI implementation makes all the difference.
What it takes to get this right
The benefits and efficiency of AI customer care are real. When the implementation is honest about what AI can and can’t do, the customer experience improves.
And the companies that are winning at AI customer care all have the same three things in common:
They started with operational necessity and weren’t just following the AI fad.
They built in feedback loops from the beginning to monitor customer sentiment.
They kept humans central to the processes and experience, rather than treating them as incidental.
Panos has spent three years building this. The work has been hard, the development time and costs real, but he’s confident we are out of the “early days” of AI customer care implementation.
I don't think it's in the early days. I think we're quite progressed. But the future is going to be crazy since there are so many more things about to come.
We focused on the problem, built a solution to solve it, but kept customer experience and human intervention at the core, and the tools evolved naturally.
