Can I speak to your human?

After Octopus Energy’s Greg Jackson recently revealed that artificial intelligence has taken on the workload of 250 staff at the energy retailer within a matter of months, Utility Week speaks to industry experts about the potential prizes and pitfalls of using AI for customer service.

If an interrogator, conversing in text with both a fellow human being and a computer, is unable to identify which is which, then the machine can be said to possess a form of artificial intelligence (AI).

No computer has yet been able to convincingly pass this test set out by code-breaking computer scientist and philosopher Alan Turing in the 1950s. However, the release in November of Open AI’s ChatGPT, available to anyone with internet access, has enabled everyday people to experience first-hand how close computers are to succeeding in Turing’s “imitation game”.

Unease over the potential societal implications of AI has been steadily growing among pioneers of the field for a number of years, but the launch of ChatGPT, among other things, has helped to catapult the issue into mainstream discussion as a matter of pressing concern.

Among the latest figures to speak out is Octopus Energy’s Greg Jackson. Writing in the Times earlier this month, the company’s founder and chief executive warned that AI will transform workplaces at a “terrifying” pace, making huge numbers of people redundant in the blink of an eye: “It’s like going from muskets to missiles. Everywhere. All at once.”

Jackson said Octopus had begun experimenting with AI in December and by February they’d plugged it into their systems, initially allowing it to reply to a “handful” of customers emails.

The responses produced by the AI had delivered a higher satisfaction rating than trained staff – 80% versus 65% – and at the time of writing, the programme was responding to more than a third of customers emails – the workload of around 250 people.

“That’s not the worst,” he added. “Today, Octopus AI writes messages, but soon (in weeks) it’ll make decisions and carry out actions too. It won’t just tell you your balance but will also suggest a change in payments and do it for you. It’ll order your meter installation or spot a vulnerability and send you an electric blanket.”

When asked about potential use cases for AI in the energy retail sector, ChatGPT responds: “AI-powered virtual assistants and chatbots can handle basic customer inquiries, provide instant responses, and guide customers through common processes such as bill payments, service requests, or energy usage analysis. This can significantly reduce wait times, improve efficiency, and provide 24/7 support.”

AI could also analyse individual customer data to offer customised assistance and tailored products and services, it explains: “This level of personalisation can enhance the customer experience, increase customer satisfaction, and foster customer loyalty.” This could include using smart meter data to provide “detailed insight” into their consumption patterns and personalised advice to help reduce their energy usage and bills.

Furthermore, AI algorithms could “detect patterns and anomalies in energy consumption data to identify potential fraudulent activities or security breaches,” while voice recognition and natural language processing technologies could enable customers to interact with energy retailers through voice commands or text-based messaging: “This can provide a seamless and intuitive customer experience, allowing customers to make inquiries, receive information, and perform transactions using natural language.”

Oliver Schoen, a director at Eon Next, says the prizes on offer from the use of AI are indeed substantial. But over the short term, at least, he believes both the opportunities and risk are somewhat overstated: “We’re probably in the midst of a typical hype cycle right now.”

Schoen says there is a spectrum of different customer service roles that AI could fulfil with varying levels of ease. At the left side of the spectrum are “simple, self-contained” questions, for example, about government schemes, that do not require access to customers’ accounts or their personal data.

At the right side of the spectrum are complex interactions, perhaps with vulnerable customers, to deal with very specific problems or sensitive issues. These may include referring customers to third parties or taking a series of interconnected actions.

He says attempts to leverage AI will obviously start at the left side of the spectrum and gradually move over to the right, initially with the AI giving human agents suggested responses to enquiries that they can then “fine tune and send off”.

Schoen says this is “not too far off” what many customer service organisations have already been doing for a while with the use of less sophisticated chatbots and automated phone lines.

But moving along the spectrum, Schoen says energy companies will quickly start hitting the “corporate reality” of dealing with issues of system integration, security and privacy. He says exposing customers directly to AI without a human filter will come with significant risks.

Schoen notes the example of chat programmes trained to respond to people based on interactions on social media, which started regurgitating some of the bigoted opinions they had come across. While acknowledging that this is an extreme example, depending on how programmes are trained, he says there is still the risk of “strange behaviour” emerging at “this stage of maturity.”

There is also a knowledge management issue given that companies often have important information spread among various service providers around the world; their products and services “iterate so rapidly”; and there are regular government interventions in the market.

He questions how companies will ensure they give customers consistent responses as this knowledge base changes: “They don’t want to call and get today a very different response to the same problem than they got yesterday, or they will get tomorrow. Now, how do you ensure this?”

Schoen says one big potential benefit of AI is the ability to quickly scale up and down to meet demand.

At the moment, if you want to launch a new service-heavy venture, the main constraint is hiring enough staff. There is also the issue of synchronising this hiring with the generation of the revenue needed to pay them. Schoen says the use of AI for customer service has the potential to completely change this dynamic, allowing companies to quickly scale up in line with the emergence of demand.

Schoen says AI could also help existing enterprises to meet spikes in demand, particularly when it comes to enquiries on the left side of the aforementioned spectrum. He says Eon has seen a big uptick in enquiries over the last year relating to the energy crisis, with customers seeking information about how their bills will be impacted, how they can keep them down and what support is available through government schemes.

Rowanne Fleck, lead user researcher at the Energy Systems Catapult (ESC), says utilising AI to reply to routine enquiries could enable energy companies to massively reduce response times, which would be “great for everybody”.

ChatGPT says some customers may have reservations about interacting with AI programmes and prefer human assistance: “Energy retailers need to strike a balance between automated AI services and human customer support options to cater to different preferences.”

But Fleck says some customers may prefer to communicate with chatbots, which could “go through things more slowly, step by step… There’s no time constraint like there is with a person on a call.” People with social anxiety or autism may actually be averse to speaking to another person, preferring a “more stable and predictable” chatbot.

She says there is a “huge opportunity” for AI to help disseminate information to customers about decarbonising their homes: “One of the biggest problems in the energy industry at the moment is people getting the right information about home retrofits, knowing what to do to decarbonise their homes.

“One of the things that people tell us constantly is they just don’t know where to start looking. And when they do, they don’t know what sources to trust. And you can imagine something like this could really start to help people get the right information”.

In general, Fleck says: “The ideal future scenario would be where the AI augments the service that is provided rather than replaces it.”

As an example of this working in practice, Samuel Young, data science and AI practice manager at the ESC, cites a study of the use of AI for tech support at a large Fortune 500 company, which found “the big benefit was it enabled customer support agents with a few months experience to be performing as if they were much more experienced.

“And that really has benefits for kind of all consumers, but vulnerable consumers in particular, where it’s a bit luck of the draw as to do you get someone who has lots of experience or someone who doesn’t”.

Although not necessary to solve the problem, Fleck says AI could also help to identify vulnerability and provide a more consistent service through its ability to quickly trawl through records of previous actions and interactions with customers.

Fleck and Young agree that one of the biggest risks of the application of AI to customer service is that energy retailers absorb the benefits as costs savings, reducing the number of staff that are available to deal with vulnerable customers and complex cases.

Young worries that the development of AI may focus on optimising the experience for the majority of typical customers, whose main priority may be speed, at the expense of the minority of vulnerable customers, whose priorities may be different.

Fleck says the existing disparity in the experiences of, and costs to serve, these segments could grow, increasing the incentive for companies to find ways to cherry pick the first set.

Young says companies need to develop supportive technologies for their specialist support teams, including by giving AI “the time and the data and the experience to learns in those kinds of contexts.” When humans do make adjustments to AI generated responses, these alterations need to be fed back into their training to create a positive “feedback loop”.

He says it is important that AI developers spend time “sitting next to a call service agent, listening to live calls” to they can personally experience the “many kinds of subtleties” of customer interactions and identify the potential “blind spots” of their software.

As AI programmes begin to interact with customers with less oversight from human agents, Young says people must have the ability to opt out, speak to a human call agent and have AI decisions reviewed.

AI and human agents could create a much better experience, Fleck concludes, but it must be a “symbiotic” relationship working for the benefit of all customers, and not just a way for companies to cut staff and costs.