Guides

9 minutes to read

Nov 11 2026

Which Calls to Give an AI Voice Assistant — and Which to Keep Human

Evgenia Panyova

Every support line takes two kinds of calls. Some are the same question for the hundredth time: where is my order, when is my class, what time is check-in. Others need a person who can listen, make an exception and take responsibility for the outcome. An AI voice assistant should take the first kind, so your agents have time for the second. The hard part is drawing the line between them before launch, not after the first complaint.

If the line is drawn wrong, customers notice first. In a Gartner survey of 5,728 customers, difficulty reaching a person was the top concern about AI in customer service, named by 60% (Gartner, 2024). The answer is not less automation but clear rules: which calls go where, what triggers a handoff and what the agent receives. This guide sets them out for ecommerce, fintech, travel and e-learning teams in the EU and UK.

How to Split Calls With an AI Voice Assistant: Three Lanes

Sort every call reason into one of three lanes before anyone writes a script. In lane one, the AI voice assistant resolves the call from start to finish. In lane two, it starts the call, identifies what the caller wants, collects details and completes routine steps, then hands over to a person. In lane three, a human takes the call from the first second.

An AI voice assistant is not a smarter Interactive Voice Response (IVR) menu. It understands free speech and follows a conversation that changes direction, so it will try to handle anything unless you define where it stops. For how a voice assistant works and how it differs from IVR, see our guide to voice assistants for sales and support.

Four questions decide the lane:

• Is the outcome reversible? A rescheduled delivery can be changed again; a refund outside policy cannot.
• Is money or a legal right in dispute? Disputes need someone with authority to make an exception.
• How heavy is the emotional load? Bereavement or a stranded traveller needs a person, even when the facts are simple.
• Does the answer come from a fixed source? Order status does. A judgement call on policy does not, and that is where AI hallucination risk is highest.

Lane two needs the most design work: two parties and a transfer. When BSG co-creates call flows with a client, fallback handling is configured with the main intent flows during setup, not after launch, so the agent starts informed rather than from zero.

The Call Matrix: Ecommerce, Fintech, Travel and E-Learning

The matrix applies the four questions to common calls per niche. Treat it as a first draft; your own call data decides the final split for your AI voice assistant.

Here is lane two in ecommerce. A customer calls about a phone delivered with a cracked screen. The AI voice assistant confirms the order, records the damage, checks the return window and explains that a colleague will decide on a replacement or refund. The agent picks up already knowing all three, so the call opens with a decision, not identification questions.

Fintech carries an extra layer in the UK. The FCA found that 26.4 million UK adults, or 49%, showed at least one characteristic of vulnerability, such as poor health or low financial resilience (FCA Financial Lives, 2024). At that scale, vulnerability is a routine routing condition, not an edge case.

When to Transfer a Call to a Human: 7 Handoff Triggers

Triggers decide when a call moves mid-conversation, whatever lane it began in. Your AI voice assistant should start an AI-to-human handoff on any one of these conditions:

1. The customer asks for a person. Transfer on the first clear request. Asking "are you sure?" is how an assistant earns a complaint.
2. A complaint is raised. It belongs with someone who can log it formally and own the outcome.
3. Money is in dispute. Refunds outside policy, chargebacks, contested fees: any decision that moves money against the standard rule.
4. Emotion is rising. A raised voice, repeated interruptions, distress. BSG's AI voice assistant recognises tone as well as intent, so a shift in tone can trigger the transfer without a keyword.
5. The request doesn't fit. If the AI fails to understand the need after two attempts, transfer.
6. The caller may be vulnerable. Signs of illness, bereavement, financial hardship or confusion. The AI should not diagnose; any signal is enough.
7. There is a fraud, safety or legal risk. Suspected scams, account takeover or legal threats go to trained staff, ideally in a dedicated queue.

Trigger one builds trust. Gartner advises that service AI must tell customers it will connect them to an agent when it cannot solve the problem (Gartner, 2024). The two-attempt rule is a starting value; adjust it once you have real transcripts.

What the Agent Receives: The AI-to-Human Handoff Package

A handoff is only as good as the context that travels with it. In a warm transfer, the agent receives context before or as the call connects, so the customer never repeats themselves. In a cold transfer, the agent gets a ringing phone and nothing else.

On the BSG side, Conversational AI Voice records transcripts, results and actions for every call in real-time analytics and connects to your CRM through One API, with CRM triggers configured during setup. So the fields can be written to the customer record your agents already use; which fields appear on screen is agreed with our team during flow design.

Without the package, the agent repeats what the AI already asked. With it, the agent opens with: "I see the screen arrived cracked; let's arrange a replacement." Before transferring, the AI voice assistant should tell the caller that a colleague will see everything covered so far, but only if your integration delivers that.

How an AI Voice Assistant Says It's an AI: EU AI Act, Article 50

Since 2 August 2026, Article 50 of the EU AI Act has required AI systems that interact directly with people to be designed so that people know they are dealing with AI, unless this is obvious (European Commission, 2026). Non-compliance can bring fines of up to €15 million or 3% of worldwide annual turnover, whichever is higher (Cooley, 2026).

On a call, a natural voice is the point of the product, so "obvious" rarely applies. Your AI voice assistant should disclose in its first sentence and offer the way out: "Hi, you're speaking with [Brand]'s AI assistant. I can help with orders, or connect you to a colleague at any time."

That line tells the caller a person is one sentence away, which answers the top concern from the Gartner survey. If asked whether it is human, the AI must answer honestly. For lines outside the AI Act's scope, such as UK-only call flows, we recommend the same opening: one script for all markets. Agree the final wording with your legal team.

Start Narrow, Then Map Your Calls With BSG

Launch your AI voice assistant with three to five frequent, fixed-answer call reasons, such as order status or class schedules. Switch on all seven triggers from day one, then review transferred calls weekly. Each handoff shows either a call the AI could have resolved, a candidate for lane one, or a trigger that worked as intended. Track containment rate alongside handoff quality: containment built on callers giving up is a failure, and CSAT on transferred calls shows whether the package works.

This is where time comes back to your team. A client review on our product page describes Conversational AI Voice handling 40% of repeat requests without live agents within a week of launch. Each repeat request it closes gives agents more time for complex conversations. For the bigger picture, see why AI voice is becoming the front line of customer support.

Send us your top call reasons. We'll sort them into the three lanes with you, agree the triggers and define what agents see on transfer, with Conversational AI Voice and the AI Support Assistant. Talk to the BSG team and let's go through your calls together.