How does AI actually handle a phone call?

It turns the caller's speech into text, uses a language model connected to your systems to decide what to do and retrieve any facts it needs, then speaks the answer back. The loop repeats through the conversation and runs in about a second.

What should we automate first?

One high-volume, repetitive call type whose answers come from a clear source of truth, such as status checks, FAQs or appointment booking. Prove one, then expand.

Will AI automation cut my team?

The usual goal is not to cut staff but to move them off repetitive calls onto the ones that need human judgment, while capturing calls you were missing outside working hours.

How do I know if it is worth the cost?

Look at your call logs, work out what share of calls fit automation, and estimate how many the agent would fully handle. If the addressable volume is small, it may not be worth it yet, and a good partner will say so.

How AI Automates a Call Center \ Vandri

The mechanics, in plain English

A voice agent does in software what a human agent does on a call, broken into three steps that repeat through the conversation:

Speech to text. The caller speaks. A speech recognition model turns the audio into text in real time. For Slovenian and Croatian, this step is where quality is won or lost, because a model not tuned for the language mishears names, streets and dialect.
Understand and decide. A language model reads the text, works out what the caller wants, and decides what to do. If it needs a fact, it looks it up in your connected systems. If the request is out of scope, it hands off to a person.
Text to speech. The agent's answer is turned back into natural speech and played to the caller.

This loop runs fast enough to feel like a normal conversation. Vandri builds it on the vCore platform, using Soniox speech-to-text for Slovenian and Croatian, LiveKit for the live call and OpenAI Realtime or Gemini for the reasoning, all on EU hosting.

What to automate first

The instinct is to automate everything. That is how projects fail. The calls worth automating first share three traits: they are high volume, repetitive, and answerable from a clear source of truth.

Good first candidates:

Order, ticket or claim status checks.
Frequently asked questions with stable answers.
Appointment booking and rescheduling.
Simple outbound reminders and follow-ups.

Leave for later, or for humans:

Emotional or sensitive calls.
One-off situations with no repeatable pattern.
Anything where the answer is not written down or systematised anywhere.

Pick one call type, get it working well, then expand. A narrow agent that handles one thing reliably beats a broad one that handles everything poorly.

The honest ROI math

You do not need invented statistics to reason about the return. The math is simple:

What a call costs you now. Agent time per call, plus the cost of missed and abandoned calls, plus any after-hours or overflow coverage you pay for.
What share of calls fit automation. Look at your call logs. What percentage are the repetitive types above? That is your addressable volume.
What the agent deflects. Of that addressable volume, how many calls the agent fully handles without a human.

The value shows up in three places: calls handled outside working hours that you were losing, faster answers on routine calls, and your people freed to handle the calls that actually need judgment. The point is not to cut the team. It is to stop wasting them on work that does not need a person.

Run your own numbers before you buy. If the addressable volume is tiny, automation may not be worth it yet, and a good partner will tell you so.

Where it goes wrong

The common failure modes are predictable, which means they are avoidable:

Wrong language layer. Using speech recognition not tuned for Slovenian or Croatian. One misheard sentence and the caller asks for a human, and trust is gone.
No source of truth. Pointing the agent at questions whose answers live only in someone's head. The agent cannot retrieve what was never written down.
No handoff. An agent with no clean escalation path traps callers. It should hand off to a person the moment a call goes out of scope.
Boiling the ocean. Trying to automate every call type at once instead of proving one.
Ignoring compliance. Storing recordings in the EU but processing audio outside it. For a call center that handles personal data, both storage and processing need to stay in the EU.

Keeping it compliant

Calls contain personal data, so compliance is not optional. A voice agent should run inside EU data residency, with the audio processed in the EU and not only stored there, plus role-based access and audit logs, following GDPR and ISO 27001 principles. We cover this in detail in our guide to GDPR-compliant AI in the EU.

A sensible rollout

The path we use at Vandri:

Look at your call logs and pick one high-volume, repetitive call type.
Connect the one or two systems the agent needs to answer it.
Ship a working agent for that call type.
Listen to real calls, fix what it gets wrong, and tune.
Expand to the next call type once the first is proven.

You stay in control throughout, and the agent always hands difficult calls to a person.

Ready to look at your numbers?

Tell us what your call volume looks like and which calls eat the most time. We will help you work out the addressable volume and whether automation pays off, honestly. If you want the buyer's view of the finished product, read our guide to an AI call center in Slovenian and Croatian, or book a call.