Voice Agent for EV Charging Support

What a charging support call actually is

Support volume in charging is dominated by a small number of situations, which is exactly the shape that automation suits.

• I cannot start a session, which is usually authorisation, payment or a vehicle that has not latched
• It stopped early, which is usually the vehicle, occasionally a fault
• It is charging slowly, which is almost always the vehicle's onboard charger and almost never the charger
• I was charged incorrectly, which needs the session record
• The bay is blocked, which is not a charging problem at all
• Where is the nearest working charger, which is availability

Every one of those is answerable from data the platform holds. None requires judgement about the equipment.

How it sits with your charging platform

The agent is not a separate support product with its own view of the world. It integrates into the charging management platform you already run, and reads and acts through it.

That integration is the part that decides whether an agent is useful or merely polite. A voice agent with no access can apologise fluently. One that can see the session, the charger state and the payment record can answer the call.

It is in pilot today: multi-language, answering the common cases, and able to start and stop a session remotely for an authenticated caller.

What an agent can resolve alone

1. 1

   Explain a session

   What happened, when it stopped, how much energy, what it cost. Reading a record back accurately resolves a large share of billing calls.

2. 2

   Restart or stop remotely

   Where the caller is authenticated and the charger is reachable, the most common actual fix.

3. 3

   Explain slow charging

   The most frequent complaint and almost never a fault. A clear explanation of onboard charger limits closes it.

4. 4

   Direct to an alternative

   The nearest working bay, which is more useful than an apology.

5. 5

   Raise a ticket properly

   With the charger, the session, the symptom and the caller already attached, so a person does not start from nothing.

What must reach a person

The handoff rules matter more than the capability. An agent that fails to escalate is worse than no agent.

• Anything involving safety: smoke, heat, a spark, an exposed conductor, a vehicle that will not release. Immediately, without attempting to diagnose
• Anyone in distress, stranded, or with a vulnerability the call makes evident
• Any dispute about money beyond reading the record back
• Anyone who asks for a person. Once, without negotiation
• Anything the agent has already failed to resolve. Repeating is worse than escalating

The last two are what determine whether people trust the line. An agent that will not hand over is the reason people hate automated support, and it is a policy choice rather than a technical limit.

The access it needs

• Charger state and recent history, so it is talking about the real unit rather than in general terms
• The caller's sessions, once they are identified
• Tariffs, so it can explain a charge rather than guess
• Availability nearby, for the redirect
• Remote start and stop, scoped to that caller's own sessions and nothing else
• Ticket creation, with everything already attached

It should not have access to change configuration, alter tariffs, act on chargers unrelated to the caller, or issue refunds without a person. The scope is deliberately narrow.

Languages, and why coverage decides the value

Language coverage is not a refinement on a support line, it is most of the value. A line that answers in one or two languages covers a fraction of the people who will actually call, and the ones it does not cover are not served worse, they are not served at all. This holds in India, where a network spans several language regions, and it holds across Europe, where it spans several countries.

• Breadth of language coverage matters more than conversational polish in any one of them
• Callers switch languages mid-sentence routinely, and an agent that cannot follow loses the call
• Accent range within one language is wide and is where recognition actually fails
• Technical terms are frequently used in English inside a regional sentence, which the agent has to handle rather than be confused by

An agent that handles several languages competently is worth more than one that reasons elegantly in a single language. For the callers it does not cover, the alternative is not a worse answer. It is no answer. The agent is multi-language today and runs in pilot.

Safety boundaries

Charging involves high current and, in DC, high voltage. There are things an automated system must never say.

• Never instruct anyone to open equipment, remove a cover or touch anything internal
• Never advise forcing a connector or cutting a cable, whatever the caller wants to do
• Never suggest bypassing or resetting protection
• Never say it is safe. It cannot see the situation
• On any indication of a safety issue, escalate immediately and say plainly to move away from the equipment

These are hard rules rather than preferences, and they should be enforced structurally rather than by prompting.

Measuring it honestly

• Resolution rate without transfer, which is the headline and the easiest to flatter
• Escalation rate, and whether escalations were appropriate rather than merely counted
• Calls it made worse: repeat calls within a day, callers who gave up, complaints about the agent itself. This is the number that gets omitted
• Time to resolution against a human baseline, honestly compared
• Coverage by language and by hour, since the value is largely in the unstaffed hours

The third one is the one worth insisting on. An agent that resolves seventy percent and quietly loses ten is not resolving seventy percent.

How a caller is identified before anything is done

An agent that can restart a session has to be certain whose session it is, and a phone line gives it less to work with than an application does.

1. 1

   The number it was called from

   First factor, and on its own only a reasonable guess. Numbers are shared, spoofed and reassigned.

2. 2

   A one-time code sent to the registered email

   Second factor, confirming the caller holds the account rather than just the handset.

3. 3

   Scope after that

   Only that caller's own sessions and their own charger. Nothing configured, nothing touched on a unit that is not theirs.

With both factors satisfied, a remote start or stop is a reasonable thing for the agent to do unattended. With only the first, it can read a session record back and nothing more.