Want to offer good customer support? Don’t turn to AI.

While people often say that customer support is an ideal place to use AI, we strongly disagree that it’s good enough yet. And if you’ve ever gone ten rounds begging a chatbot to connect you with a human, you’ll understand why.

We recently doubled down on offering human support and even had a recent hiring round to expand our team. This means we’ve now got 23 people working on support for us, based all around the world. We’ve discussed our approach to customer support before, but with so many companies leaning into AI as “the way of the future”, we’re quite happy to be going against the grain—at least for now. For us, it’s all about the human touch.

Why we don’t think AI belongs in customer support yet

The quality of an AI response is poor

This one goes without saying. Is there anyone out there who’d rather explain a complicated problem to a chatbot rather than a person? AI can’t understand what you’re telling it—it can only regurgitate data, and not even with guaranteed accuracy.

There’s too much potential for inaccuracies

AI is notorious for hallucinating and giving misinformation with a lot of swagger and confidence. Being led astray by a chatbot that’s supposed to be providing the answers is the last thing that we want for someone wanting help with Cliniko!

What about AI as a behind-the-scenes tool?

Our founder, Joel, spoke more about our approach to AI last year and why we tread cautiously around it. But as a tech company, we’re also always keen to try out the latest tools and we’re constantly evaluating whether AI can be useful—to both our team and our customers. So when Intercom (which is the online platform that we use to provide customer support) offered us a trial of their latest AI offering, “Copilot”, we were interested to check it out.

What we trialled

Intercom advertises Copilot as “the personal AI assistant for every support agent.” It doesn’t interact with customers; it’s an exclusively behind-the-scenes tool that’s intended to help the people providing support.

The way it works is simple. Whoever is using it can ask it questions related to a customer inquiry and Copilot creates an answer that (at least in theory) our team could edit and use as the basis of a response to a customer.

Like any Large Language Model (LLM), we needed to train it—so we fed it our entire support guides and the last 3 months of customer conversations. In other words, it was trained on the most up-to-date information possible about Cliniko.

For two weeks, we trialled it. None of us are using it anymore.

Why we disliked Copilot

Low quality writing

For the most part, the answers generated by Copilot were accurate, but we’d still end up rewriting them from scratch. AI has a recognisable voice and tends to give responses that are wordy and convoluted. Copilot would regularly spit out answers that were 3-4 paragraphs long, when a human would convey the same information in 3-4 sentences. When it comes to tech support, it’s essential that we can give clear and concise instructions; busy practitioners using Cliniko don’t have time for anything else!

Unfortunately, however, Copilot is a one trick pony—wordy is all it can do. The software even comes with a button that allows you to “adjust tone” and one of the options they provide for this is “friendly”. Interestingly, when you click this, it simply adds more words. Not helpful.

For example, when answering the question “how do I re-send a confirmation email?” this is how Copilot responded.

Firstly, here’s an example of the “normal” tone (46 words):

And this is its “friendly” tone” (60 words):

Aside from adding 14 unnecessary words, the “friendly” answer also sounds weird and unnatural!

Giving wrong answers

There were other occasions when Copilot was just outright wrong. Despite being trained on the most up-to-date information we could provide, it still tended to conflate ideas. For example, it would confuse information that was intended for a completely different part of Cliniko and spit out an incorrect answer.

Even though it was mostly accurate, it was wrong frequently enough that we didn’t want to start relying on it (and inadvertently feed incorrect answers back to our customers), so it had to be constantly checked. This essentially eradicates its potential to save time. We also recognised that the accuracy of Copilot’s answers was likely to deteriorate over time as it started to cannibalise its own misinformation—so it was only going to get worse.

Jeopardising the quality of our support

Lastly, we realised that relying on Copilot was going to lead to worse support over time. This is always one of the risks with AI; it makes people lazier at the task they’re doing and ultimately starts to impact on their skillset. At Cliniko, rather than opting for an AI solution, we’d prefer to put our time and energy elsewhere. For example, we’d be more inclined to keep providing solid internal training on what constitutes good technical writing and continue investing in our humans.

Situations where AI could be helpful to a customer support team

There are only a few instances where we think AI tools in their current state could be handy to a support team:

1As a search tool

One thing we did like about Copilot was the fact that it would include links to old support conversations that we could click on. But this was useful because we could read back over previous responses we’d offered—written entirely by a human—and use these as the basis for a new answer. In other words, Copilot was helpful as a search tool.

2As a replacement for “tier 1” support

Although we don’t have tiered support at Cliniko, many companies operate this way. Tiered support is a way of managing inquiries where the customer is escalated upwards across different levels if their problem is not resolved. In these cases the entry point support tends to be pretty basic, and we can see AI tools being useful for businesses who use this model. But it’s not our style.

Don’t get us wrong, we’re not against AI at Cliniko. It has definite utility in some situations, and we’re very interested to see how the technology progresses with time. But to date, we’ve found that AI rarely lives up to what it promises. Although on the surface it often seems like it’ll make things quicker, in practice we’ve found that it usually ends up costing much more time.