Amazon Nova 2.5 Sonic: voice agents in Italian with tools and human control

Amazon Nova 2.5 Sonic is a model for voice conversations that can dialogue in Italian and use software tools during the call. His most concrete business employment is an assistant who collects a request, consults authorized information and passes the conversation to a person when needed. It is not enough to connect the model to the phone to get a reliable switchboard.

AWS announced its general availability on 5 October 2026. The announcement lists Italian among the supported languages and availability on Amazon Bedrock in the Virginia, Oregon, Stockholm and Tokyo regions: it does not announce a Milan region for this model. These are features stated by the provider, not results of a Giallo test. AWS Announcement.

For an SME, the useful question is which conversation to entrust to it and with what limits. Booking an appointment already regulated by a calendar is different from handling a complaint, promising a refund or giving a health claim. The design of the service must start from this difference.

What changes with a speech-to-speech model?

A traditional voice system can concatenate transcription, response generation, and voice synthesis. A speech-to-speech model directly handles audio exchange, along with any text messages and instrument calls. For the company manager, the potential advantage is a less fragmented conversation; however, the acceptance criterion remains what the customer manages to achieve.

AWS describes asynchronous tools, a context window from 256k Token and seven languages supported. These specifications do not tell how many phone calls a company can serve nor do they guarantee the understanding of every accent, name or product code. The ability available in the account and the behavior on your conversations should be verified separately. Release Specifications.

A practice number, for example, must be confirmed before the search. An ambiguous date must be clarified before booking. If the customer interrupts the entry, the system must retain the correct status of the request. They are application requirements: a natural entry can make a mistake more convincing, not less harmful. The Guide to voice AI for companies helps to separate telephone experience and process automation.

What tools can you use during a call?

The model does not automatically have calendar, CRM, or warehouse access. The application exposes specific tools and decides which parameters to accept, which data to return, and which operations to allow. A function to read availability should return only the relevant times, without delivering the entire calendar with names and notes of other customers.

Distinguish tools for reading, preparation and writing. Reading a case’s status may be permitted after identity verification. Preparing an appointment proposal is not the same as confirming it. Writing to the business management software requires field checks, authorisation and duplicate prevention. The voice should communicate confirmation only after the operation actually succeeds.

If an internal system is reachable through a private network, the link also becomes an infrastructure project. The new support of AgentCore Gateway for Private APIs It concerns this level, distinct from language skills. No voice model solves authentication, management software timeout, or expired credentials on its own. Separating these problems makes it possible to understand where a disservice arises and who needs to intervene.

When is supported Italian not enough?

The official support of Italian is an initial requirement. A business conversation, however, includes surnames, addresses, abbreviations, brands and industry terms. Even a grammatically correct answer can associate the client with the wrong practice. For this, tests must include noise, overlaps, hesitations, and caller corrections.

AWS service card excludes real-time speech-to-speech translation among supported uses. You don't have to present Nova Sonic as a universal telephone interpreter: the ability to converse in different languages is another function. The performance and improvements reported by the manufacturer also require a context-of-use assessment. Official Service Card.

A good test includes questions that the assistant must refuse or transfer. If the customer asks for a diagnosis, a contractual exception or an unexpected change, the correct answer may be the switch to the operator. To design this part, it is also useful to read how a Business Voice Assistant: utility arises from the perimeter of the service, not from the number of possible answers.

How do you divide tasks between voice, rules, and people?

The table is a design proposal, not a list of functions automatically activated by the product. Each line requires company rules and a contact person who can approve them. The transfer must also be concrete: number, availability band, synthetic pattern and information already collected, without forcing the caller to start again.

For some requests a dedicated classifier may be needed before the voice path. Perplexity Decider for ticket sorting shows the logic of categories, thresholds and revision. The choice of the classifier remains distinct from the choice of the entry and must be evaluated on the actual process.

Hypothetical example: a review center collects bookings

Imagine an Italian vehicle inspection centre that wants to handle enquiries while its staff are busy. The assistant asks for the type of work, checks the calendar and suggests availability. Before confirming, it repeats the day, time and service. It does not provide quotes for work absent from the approved price list.

The management software returns a reservation ID. Only then the voice communicates that the appointment is recorded. If the writing fails, the assistant retains a callback request instead of declaring success. If the customer changes the time during the conversation, the previous proposal is invalidated before generating a new one.

This is one hypothetical scenario, not a customer or a proof of savings. The criteria to be verified would be: appointments entered in the correct service, absence of duplicates, understanding of corrections and successful transfers. The duration of the call can be observed, but it should not reward short conversations that leave unresolved problems. A AI switchboard for phone calls should be designed around these outcomes.

Hypothetical example: a distributor assists installers

Instead, consider a spare parts distributor who receives phone calls from installers. The assistant can read the status of an already identified order and communicate authorized delivery information. It must not suggest technical compatibility when data is missing or invent availability to avoid a pause in the conversation.

The flow separates the verification of the caller from the search for order. The tools return an understandable state, along with the update date. If the system is not up-to-date, the entry declares it and switches to a person. A request to change the destination requires a different level of verification than a simple consultation.

This scenario is also hypothetical. The value to be searched is not an abstract percentage of automation, but a more orderly handling of repetitive requests. In the pilot, personnel must be able to report incorrect responses and consult the passage that generated them. The assistant does not become the commercial manager: he performs a delimited task, with a trace of the actions and a way out for the customer.

How to prepare a test before the telephone connection?

Open the card for a controlled voice test

Write down the allowed task, the approved source, the available tool, and the point where you need a person. Prepare conversations with correct requests, missing information, fixes, and unavailable systems. Separately evaluate understanding, response, software action, and transfer. Use test data until the path is stable.

The test should retain the version of prompts, tools and configuration. Otherwise a correction is not comparable with the previous test. It is useful to distinguish understanding error, wrong response, wrong instrument call and external system failure. Just saying that the phone call went wrong doesn't help correct it.

The Strands framework announced the general availability of Bidi Agents at the same time and supports Nova Sonic along with other providers. It offers a possible integration path, not a switchboard ready for any company. Strands Announcement.

Another avenue concerns the tools for analysing and modifying existing agents: ElevenAgents Architect work on configurations, drafts and tests. The useful comparison is on the operating cycle needed by the company, avoiding confusing a voice model with an assistant to maintain its application.

What mistakes make a voice agent fragile?

Allowing too large scriptures It's the first mistake. A generic tool that modifies any field of CRM makes it difficult to control effects. Better small operations, validated parameters and explicit confirmations for what changes a commitment to the customer.

Another mistake is treating every silence as consent. The customer may be distracted or the line may have fallen off. Confirmation of a reservation must be recognizable and the system must know what to do when it is missing. Even automatically repeating a write after a timeout can create duplicates if the management software has already received the first request.

Finally, a European region is not enough to declare the project compliant. It is necessary to evaluate transmitted data, logs, any registrations, personnel accesses, suppliers involved and storage. The choice of Stockholm is an infrastructure possibility indicated by AWS, not a certification of the business procedure. Announced regions.

How to recognize a pilot ready to grow?

A driver is ready to grow when the observed errors are understandable and the team knows how to handle them. If an incorrect booking depends on a missing rule, the rule can be corrected. If, on the other hand, no one can reconstruct what data the system used, increasing calls makes the problem more difficult to contain. Traceability must arrive before the service is extended.

It is also necessary to verify the route from the point of view of the staff. The operator receiving a transferred call should see an essential summary: reason for the contact, confirmed information and operations already performed. It should not receive a long reconstruction that mixes hypotheses and facts. When the system is not sure of a given, that uncertainty must remain visible in the passage.

Another check is for updates. A change in schedules, trading conditions or calendar categories must reach the assistant through a maintained source. Hiding this information in a very long prompt makes it easier to forget a change. The knowledge base and operating rules must have a manager and a review procedure.

Finally, the customer must be able to leave the automated journey. An explicit request to speak to a person, a repeatedly misunderstood conversation or a problem outside the agreed scope must have a defined destination. If no operator is available, it is better to collect a verified request for follow-up than simulate an impossible transfer. This part of the service may seem less spectacular than the voice, but it determines whether the experience remains reliable when something does not go as planned.

The final verification should also include the actual availability of staff. An agent can correctly recognize the need to transfer the call and fail anyway if the human path is not viable. The team must try operational bands, no response and callback request. The outcome communicated to the customer must correspond to what the service can really guarantee, without implying that an operator has received a request when there is no confirmation.

Which conversation is best to start with?

It is advisable to choose a frequent request with a verifiable outcome and an available human alternative. Before the pilot, the company should know who keeps the information, who corrects the errors and who deactivates the flow if a problem arises. These responsibilities matter as much as the choice of model.

The first integration may be limited to consulting data and preparing a request. The scriptures come when confirmations, permissions, and misbehavior have been verified. The cost should be estimated on the entire chain: model, telephony, cloud services, integration and maintenance. A universal price list is not indicated here, because plan and architecture must be verified on the official pages and in the account.

If you want to define a useful conversation, with access to management software and transition to staff, the work on Giallo AI agents starts from that concrete path. The result to be asked is a service that knows what it can do, recognizes its own limit and leaves the customer a clear result.