Google I/O 2026 took place on 19 and 20 May. Google announced Gemini Omni, Gemini 3.5 Flash and a broader set of agentic product capabilities. For a Swiss SMB, the useful conclusion is not that phone bots can suddenly take over every conversation. The practical change is that real-time models can react faster, process several input types and call tools while a conversation is in progress.
There are still several technical and organisational layers between a polished keynote demo and a dependable customer call. Evaluating only the model ignores the phone connection, audio quality, interruptions, calendar access, error recovery and the transfer to staff.
What Google actually announced
The official Google I/O 2026 collection covers new models and agentic experiences. For voice projects, the current Gemini Live API documentation is also relevant. Google describes a preview service for bidirectional real-time audio, image and text interaction, with features such as user interruption, tool use, transcripts and multilingual conversations.
The preview label matters. A preview service can change and needs independent monitoring, fallback behaviour and version control before production use. An SMB should treat the technology as a new pilot option, not as a service guarantee.
Six places where a phone experience can lose quality
- Telephony: the call is converted into an audio stream and back again.
- Turn detection: the system must distinguish a natural pause from the end of a request.
- Model reasoning: it interprets intent, context and spoken details.
- Tools: calendars, CRM systems and knowledge bases may respond slowly or fail.
- Speech output: voice, pacing and pronunciation affect trust.
- Handoff: when confidence is low, staff need the reason and the captured context.
An initial response in two seconds may feel natural. The same architecture feels slow when a calendar request adds another three seconds. Latency therefore needs to be measured at each layer rather than as one average across the entire call.
Swiss German is not a tick box in a language list
Google documents broad multilingual support. That does not tell a garage in Bern how a deployed system will understand local names, municipalities or dialect over a compressed phone channel. French in Romandy and Italian in Ticino also require their own tests.
A useful dialect set contains at least 30 short phrases from the target workflow: surnames, towns, times, vehicle registrations, product names and typical corrections. The team scores both the transcript and the final task. A spelling error in a transcript may be harmless; a booking for the wrong time is not.
Callers must be able to interrupt, correct and leave
People do not speak like form fields. They correct themselves, interrupt and change the subject. The Live API documents barge-in, but the application still decides what happens next. Is a started booking discarded? Are confirmed fields retained? Does the system clarify when two different times were mentioned?
Every action needs a transaction boundary. Reading is less risky than writing. Offering an available slot is less risky than booking it. In the first pilot, the voice agent can capture information and prepare an action; a final confirmation follows only after an explicit read-back or human approval.
A realistic service-team pilot
A trades business could collect callback requests outside office hours. The assistant asks for a name, phone number, municipality, type of problem and preferred time window. It promises neither an appointment nor a price. Emergencies, existing complaints and unclear safety situations are escalated to a defined number or queue.
Before launch, the team creates 50 scenarios. They include quiet callers, dialect, background noise, interruptions, incomplete phone numbers and an immediate request for a person. The pilot handles only a small share of real calls until the critical handoff scenarios consistently work.
Privacy and recording are architecture choices
The Swiss FDPIC explains that the FADP applies to AI-supported processing and that affected people require transparent information. A phone system should therefore state plainly that an AI is answering and why the requested details are being processed.
A transcript, an audio recording and a CRM note are three different data objects. Each needs a purpose, retention period and access group. If audio is not required for quality review, it should not be stored indefinitely. Sensitive content should not be copied unfiltered into analytics or prompt logs.
A two-week voice scorecard
| Measure | What it reveals | Stop signal |
|---|---|---|
| Time to first understandable response | Perceived responsiveness | Long or highly variable pauses |
| Task completion | Whether a callback or appointment request is complete | False confirmation |
| Recovery after interruption | Whether corrections are retained correctly | Lost or duplicated details |
| Handoff with context | Whether staff can continue without restarting | Transfer without summary |
| Dialect errors with operational impact | Whether speech errors cause wrong actions | Errors in place, time or identity |
These measures are appropriate for an AI phone assistant for Swiss SMBs. They reveal where the flow is safe to expand, unlike a simple count of completed conversations.
Procurement questions before selecting a provider
- Which model and API version runs in production, and how are changes tested?
- Where are audio, transcripts and tool logs processed and stored?
- What happens when the calendar, CRM, model or phone provider fails?
- Can the caller request a person at any point?
- How is an incorrect action reversed, and who receives an alert?
- Which dialects have been tested with the company's actual target group?
Primary material for the technical review
- Google: I/O 2026 news and announcements
- Google AI for Developers: Gemini Live API
- FDPIC: AI and data protection
Voice AI became more capable around Google I/O 2026, but it did not become automatically production-ready. The sensible next step for an SMB is a narrow pilot with real dialects, measured latency, restricted permissions and a human handoff that still works on a busy Monday morning.
FAQ
Does a faster model automatically produce better phone calls?
No. Perceived quality also depends on telephony, speech handling, tool calls, interruptions, text-to-speech and the handoff to a person.
Can voice AI understand Swiss German reliably?
That must be tested with real calls from the target region. A general language list does not replace a dialect, accent and noise test for the deployed system.
What is a suitable first pilot task?
Choose a narrow task with an explicit handoff, such as collecting a callback request, explaining opening hours or structuring an appointment request.
Should callers be told that they are speaking to AI?
Transparency is essential. The FDPIC says that the purpose, operation and data sources of AI-supported processing must be communicated transparently.