Agentforce Voice grows up: testing, observability and disclosure before you put an AI on the phones

Most coverage of Dreamforce 2026 focused on the big interface story. Tucked into Salesforce's list of 21 announcements is a quieter set of voice updates that matter more to anyone planning to put an AI on the phone. They are less about a smarter voice and more about the unglamorous machinery around it: how you test a call, how you see what went wrong, and what you owe the caller.

What Salesforce announced for voice

According to Salesforce's own announcement list, the voice updates fall into three groups.

More places to use voice. Agentforce Voice for field teams is generally available in the Salesforce mobile app, with more apps expected in early 2027. Before a visit it can read out summaries of accounts, cases and opportunities, and afterwards reps can dictate notes, status updates and follow-ups. A voice option for employee-facing agents, which started on mobile, is heading to the Agentforce chat client on Lightning Desktop, currently as a pilot.

More languages. Six were added and are generally available: Spanish, Italian, German, Portuguese, Japanese and Canadian French.

Tooling for quality. This is the part we find most interesting. Voice and Conversation Testing, currently in beta, grades multi-turn voice and text conversations, includes a task-resolution measure and uses simulated customer personas, including frustrated and distracted ones. Voice Observability is generally available and monitors audio quality, flagging low-confidence words, noisy environments and latency, and pointing to the moment a conversation went wrong.

A partner write-up of the event also repeats a Salesforce claim that Agentforce Voice has resolved more than five million conversations on Salesforce's own support site. No time period or method is given, so treat it as a vendor figure rather than a benchmark for your business.

Why the testing tools matter

A voice agent fails differently from a chat agent. A misheard product code, a long pause, or a caller on a noisy site can derail a conversation in ways a text transcript never shows. Until now, many teams have tested voice by having colleagues ring in and tell them how it felt. Scripted personas and per-call diagnostics move that towards something repeatable, which is what you need before changing a prompt or a knowledge article without worrying what else you broke.

The caveat is that these tools are only as good as the scenarios you feed them. A persona built on Salesforce's assumptions will not know your customers' accents, your jargon or the three questions that make up half your call volume. Build the test set from your own call recordings and complaints.

The disclosure question is getting sharper

The other pressure comes from regulation. A recent vendor summary of AI voice disclosure rules (so verify against primary sources) reports that the EU AI Act's transparency obligation, which requires people to be told they are dealing with AI no later than the first interaction, has applied since 2 August 2026. It also describes a patchwork of US state rules, and says the FCC's proposed AI-call disclosure requirement is still only a proposal. For Australia, the same summary says there is no AI-specific duty yet in the telemarketing rules.

The legal position varies by jurisdiction and call type, so we would not hang a design on a blog summary. But the direction of travel is clear enough to act on now. Telling callers up front that they are speaking with an AI is cheap and defensible, and it fits with other research we've covered: Gartner found that 87% of customers say it is essential to be able to reach a human.

What we'd do before going live

• Pick narrow first journeys. Status checks, bookings and simple account changes are easier to test and safer to get wrong than complaints or cancellations.
• Fix the data first. A voice agent grounded in out-of-date knowledge articles, or one that cannot see the customer's case, will confidently say the wrong thing out loud. This is the same advice we give for any Agentforce rollout.
• Build your own test scenarios. Use real calls, accents and background noise, and include the awkward cases such as interruptions, silence and anger.
• Define success as resolution. Containment tells you the AI kept the call. Resolution and repeat contact tell you whether it helped.
• Design the escalation path early. The handoff to a human should carry the caller's intent and what has been tried, so nobody repeats themselves.
• Decide your disclosure and recording stance. Settle the wording, call-recording notices and what the agent may do on a customer's behalf before launch, not after a complaint.
• Watch the cost. Voice usually prices differently from text. Check the current Salesforce pricing and licensing for your edition rather than relying on third-party guides.

Voice is no longer a research project for Salesforce customers. The platform now ships the basic equipment to test and monitor it, which removes one excuse for waiting and one for skipping the hard work. The organisations that benefit will be those that treat testing, observability and disclosure as part of the build rather than a checklist at the end. If you would like an independent view on scoping a pilot, the guardrails or the compliance side, that is the kind of work we do.

Sources: Salesforce, "21 Things We Announced at Dreamforce 2026"; Cloudgaia, "Dreamforce 2026 Announcements"; JustCall, "AI Voice Agent Disclosure Laws: Federal, State, and International Rules 2026"; Gartner, "87% of customers say companies using GenAI for customer service must provide access to a human agent".