Artificial intelligence and ecommerce
Voice agents for customer service and ecommerce
How real-time voice models open new support and shopping experiences, and what reliable business implementation requires.
In brief
Key ideas
- Amazon introduced Nova Sonic on 8 April 2025 for real-time, bidirectional voice conversations.
- Voice can reduce friction in support, order tracking, availability and assisted shopping.
- Experience depends on latency, interruptions, context, integrations and human escalation.
- Automation should be designed around tasks and outcomes, not a technology demonstration.
On 8 April 2025 AWS introduced Amazon Nova Sonic, a model designed to understand and generate speech in one architecture and sustain real-time bidirectional conversations. The announcement illustrated a broader change: interaction with artificial intelligence did not need to begin in a text box.
For customer service and ecommerce, voice can turn complex tasks into natural conversation. Convincing speech alone is not enough. Value appears when the agent understands intent, retrieves authorised information, executes safe processes and knows when to transfer the conversation.
Immediate use cases
In support, an agent can identify an order, explain its status, answer common questions, collect incident details or prepare a human handover. In ecommerce it can guide selection, check compatibility and availability, explain delivery and returns, or support a purchase.
Internal scenarios matter too: warehouse staff can ask for instructions hands-free, sales teams can update CRM records by voice and managers can request operational summaries.
Start with a focused, measurable need. Trying to handle every possible conversation from day one multiplies risk, cost and frustration.
A conversation is not a chain of audio files
Traditional architectures converted speech to text, sent text to a model and synthesised the answer afterwards. An integrated speech-to-speech model can retain more information about pace, tone and interruptions while reducing latency. AWS introduced Nova Sonic with unified understanding and generation, bidirectional streaming, function calling and grounding in enterprise knowledge.
Quality still depends on the complete system: correct recognition of names and references, quick responses and natural interruption, conversational memory, controlled access to catalogue or CRM data, explicit confirmation before payment or cancellation, and a handover that preserves context.
Design for errors
In a visual interface users can reread. A spoken error disappears while it is delivered. Summarise sensitive data, ask for confirmation and offer alternatives. If the system cannot identify a reference, it should disclose uncertainty rather than invent.
Privacy decisions must be explicit: what is recorded, for how long, for what purpose and who may access it. Credentials should never be requested casually. Authentication, consent and data minimisation are part of the experience.
From pilot to operations
A good pilot limits channel, language, hours and tasks. Begin with low-risk queries and anonymised real data or a test environment. Assess complete conversations before expanding.
Useful metrics include first-contact resolution, time to resolution, transfers, abandonment, repetition, identification errors, satisfaction and cost per interaction. Ecommerce should add assisted conversion, order value and returns linked to recommendations. Qualitative review remains essential because averages can hide awkward silences or unclear answers.
Connect voice to the omnichannel journey
The agent should not become an isolated channel. A journey may begin on the website, continue by voice and finish with a link or written summary. Relevant history should reach the human team without forcing the customer to repeat everything.
Sitelicon treats voice agents as a combination of experience design, integration and operation. We define the use case, connect the necessary sources, establish boundaries and measure the outcome. Voice is a powerful interface; trust depends on everything behind it.
Editorial note: originally published in April 2025 and reviewed on 11 September 2026 to preserve the editorial archive.
Editorial responsibility
Who is responsible for this content?
Sitelicon Team
Content created and maintained by Sitelicon’s multidisciplinary team.
- Publication
- Latest review
- Traceability
- 2 referenced sources
Sources and updates
Verified and dated information.
Last update recorded on . The features and terms of digital platforms may change over time.
Related services
We can help you put it into practice.
For each article, we select the Sitelicon capabilities most directly related to the topic.
Artificial intelligence and machine learning
We build assistants and agents connected to business data and objectives.
View service 02Integrations and automation
We connect conversations to CRM, commerce, orders, support and operations.
View service 03Custom software
We develop interfaces and workflows with control, traceability and oversight.
View service