Ask a support team which tickets to hand to an AI agent first and the textbook answer is "high volume, low risk". That is true and not very useful, because it does not tell you which tickets those are. We have a lot of conversations with support teams about where to start, and the answers cluster. Last week two of them stood out because the buyers said the same thing, unprompted and in almost the same words. One runs support at a software company. The other runs customer service at an energy utility. Both picked the tickets about money.
Two buyers, one number
The support lead at the software company had counted before we spoke. "Weekly we receive about 4,000, 5,000 actionable tickets, and then a third of those are billing, and that's what we are looking to target."
The head of customer service at the energy utility had a different product and the same share. "About 30 percent of our requests are: I did pay, just late." When the dunning case came up, his verdict was short. "That is exactly bullseye. That is exactly what I want to do first."
Two industries, two ticket types, one starting point:
- Software company: a third of tickets are billing.
- Energy utility: about 30 percent of requests are late-payment questions.
Money tickets are standard, and everyone knows the rules
What makes these tickets the obvious first candidate is not that they are easy for an AI. It is that they are the same every time, and the answer is written down somewhere.
The software support lead described her billing queue as "refunds, plan switches, questions about charges that they're unfamiliar with". Her summary: "Billing is where it's the most ripe. They're pretty standard billing questions."
For the utility, the ticket is a customer who received a reminder after paying. The answer depends on a payment record, not on judgement. Both teams already have a policy for these cases. What they do not have is the time to apply it a few thousand times a month.
The rest of the queue needs a human eye
The same buyers were just as clear about what they were not automating first.
The software support lead on the other two-thirds of her tickets: "The remaining two-thirds are extremely technical. They require a lot of troubleshooting within tools that we have internally. There's a lot of nuance in the troubleshooting that for now requires a human eye."
The utility's digital lead chose dunning on purpose, as a stress test. "We thought we would take a crowbar case, because then we can test the whole system on one case, and challenge it a bit." Later in the same conversation: "I know the case is extremely heavy, and I deliberately picked it, to test the limits a bit, so we know where we stand."
So the first use case is the one that is big enough to matter, standard enough to work, and hard enough to show where the AI stops.
Even inside the money case, some tickets stay human
Choosing billing or dunning first does not mean handing over the whole category. Both buyers drew a line inside it.
At the utility, the digital lead wants objections, complaints and instalment requests routed to people. "Depending on severity and situation there are different instalment agreements, but that is usually decided by the respective agent, the real agent who has the case at that moment." Questions about the claim itself, he said, "we could maybe even leave out for now".
He was also clear about the goal. "We don't want 100 percent AI resolution either. They said themselves they would already be helped enough if only the emails that actually need handling landed in their queues."
That is the practical version of "high volume, low risk": start with the tickets about money, automate the standard ones, and keep the disputes with the people who are allowed to decide them.
We have a broader list of customer service tasks to automate with AI agents, which takes a more cautious view of billing than these two buyers did, and a comparison of AI customer support platforms for billing and refund requests.
Quotes are translated where the conversation was in German, and anonymised. Only the customer's side is quoted.
