Every Charge Point Operator (CPO) feels the same pull right now. As EV charging networks grow, support tickets scale with every new charger and every new driver, expectations for the driver experience keep climbing, and a dozen vendors promise an AI customer support agent that will resolve most of it overnight. The temptation is to treat the decision like a switch: turn it on, watch the queue shrink.

There's a playbook forming for launching AI in customer service, and the honest version of it isn't about flipping a switch at all. Launching an AI customer support agent well is a sequence: understand the fundamentals, choose a deployment approach, build a business case, define your criteria, evaluate, deploy, and then optimise. Read as a whole, the message is that a good launch is engineered, not gambled. Here's how that logic applies to EV charging support specifically.

Start with a business case, not a switch

Before you evaluate a single vendor, you should be able to answer three questions: what problem are you actually solving, which conversations are worth automating and what the real economics are.

For a CPO, the first answer is usually hiding in your own ticket data. A large share of contacts are predictable and repetitive: a session that won't start, a payment that failed, a charge card that isn't recognised, a driver asking why they were billed twice, or a "the app says available but nothing happens" report. These are the workflows where an AI customer support agent can either fully resolve the issue or do most of the work before a human steps in — and they're where the return shows up fastest.

The economics are where most business cases go wrong. Cost-per-conversation matters — but on its own it's only half the picture. You also have to measure outcomes: how many of your drivers' questions the agent can actually resolve. And that number is largely set by the data layers the agent can reach. An agent that only sees tactical charge-point data might resolve around 30% of contacts, because everything touching billing, a driver's charge card, or CRM history sits outside what it can answer. A more full-stack agent that integrates across those layers can resolve far more — closer to 65%. Every question it can't answer still lands on a human, and an angry driver stranded at a charge point is not a ticket you want to leave half-handled.

So when you build the business case, don't stop at cost-per-conversation. Map which data layers and domains the agent can actually reach — is it only the tactical charge-point level, or also the user and charge-card level and your CRM? — because that's what decides how many of your business's questions it can resolve. An agent that handles 65% of your EV charging contacts is worth far more than one that handles 30%, and for an operation fielding thousands of driver contacts a month that gap compounds into real savings and capacity your team can redirect toward the hard cases that really shape the charging experience.

Decide what "good enough to launch" means

You wouldn't energise a new charger without acceptance criteria. An AI customer support agent deserves the same. Before go-live, define both the capabilities you need and the bar the agent has to clear.

For EV charging, the capability list is fairly specific. The agent has to handle more than FAQs — it needs to read live session and billing data so it can tell a driver why their session failed, not just what a session failure is in general. It has to work across the languages your drivers speak and the channels they use, hand off cleanly to a human when a charge point is genuinely faulty, and stay inside your policies on refunds and reimbursements. Then set the criteria you'll judge it on: resolution rate (fully handled, no human), driver satisfaction on par with your human agents, and quality across accuracy, behaviour, and experience. Deflection — pushing the driver away from a human — is the wrong target. A resolved issue and a better driver experience is the right one.

Treat launch as day one, not the finish line

This is the part most teams underestimate. Launching well is the start of a loop, not the end of a project. The strongest operators run it as a flywheel: train, test, deploy, analyse, over and over.

In practice, you don't switch everything on at once. You go live on a narrow, high-volume, low-risk slice — say, "where is my session and why did it fail" — prove it, and expand. Train adds new skills: billing disputes, refund handling, roaming errors between networks, multi-step workflows like home-charging reimbursement. Test validates each new skill in simulation before a driver ever sees it, so a change to your refund logic doesn't quietly break something else. Deploy rolls it out to a specific market or channel first. Analyse looks past individual tickets at the macro patterns — which topics keep escalating, where the agent is unclear — and feeds that straight back into the next round of training. Each cycle compounds the one before it, and the customer experience improves with it. The agent you launch with should be noticeably less capable than the one you're running three months later.

The takeaway

An AI customer support agent isn't a coin you flip and hope lands right. For a Charge Point Operator, it's an asset you build a case for, set criteria around, launch deliberately, and then grow skill by skill. The CPOs who treat it that way won't just deflect more tickets — they'll keep resolving more of what actually strands their drivers and steadily raise the charging experience, month after month, without their headcount rising to match.