Everything You Ever Wanted to Know About Agentforce Voice (But Were Afraid to Ask)

You've probably heard of Agentforce Voice, but the details might be a little fuzzy. “Is it a separate product?” “How does it connect to my existing agents?” “I already have AI chat — do I really need it for my telephony, too?”

The truth is, customers don't distinguish between channels. Whether they're typing, emailing, or calling, they expect the same level of service every time they engage with your business. With AI, those expectations have never been higher. Every interaction is an opportunity for a personalized white-glove experience. But as AI agents rapidly transform the way businesses engage with customers through text channels, telephony has emerged as a new blindspot for many organizations. As text-based interactions go AI-native, voice calls remain stuck in the past.

Much of that disparity is self-inflicted. Many enterprises approach voice AI as a separate procurement, with a new vendor, new infrastructure and second AI stack. That means two separate systems to maintain, two sets of guardrails to define and keep in sync, and a seam between them that customers feel every time they switch channels. Voice AI shouldn’t live in a silo. It should be a core capability of the agents you’ve already built, using the same reasoning engine, knowledge base and escalation paths.

That doesn’t mean voice is simply chat with a microphone. Three things make it harder:

This guide explores the architecture we’ve designed around these constraints and the configuration decisions that separate a natural-sounding voice AI from one that feels robotic. It also covers the use cases where AI voice agents are already delivering ROI today.

Agentforce Voice isn't a separate product but rather a capability that adds speech-to-text, real-time reasoning, and text-to-speech to the agents you've already built. The configuration, knowledge grounding, and escalation logic you defined for chat carry over. This not only changes the implementation timeline from months to weeks, but ensures that from day one, your voice agent knows everything about your business and customers.

Agentforce Voice integrates natively with your Salesforce data, actions, and workflows, with no customer connectors or parallel AI stack to build and maintain. When calls require escalation to a live staffer, the handoff carries full conversation context, so the representative picks up where the agent left off rather than starting over.

The six surfaces voice shows up on
Agentforce Voice deploys across six surfaces, each optimized for a different moment in the customer or employee journey.

Most voice AI deployments stitch together a transcription service, an LLM, a text-to-speech (TTS) layer, and a Contact Center as a Service (CCaaS) with custom integrations. Agentforce Voice is a single architecture with discrete components, each with a specific job.

How calls get in

A call arrives via one of two paths: Public Switched Telephone Network (PSTN) or Session Initiation Protocol (SIP) from an existing CCaaS platform (Amazon Connect, Genesys, NICE, Vonage, and others), or WebRTC from a browser or custom app via the Voice Mobile SDK. The audio stream enters Einstein Speech Foundations — Salesforce's speech processing layer, powered by Deepgram for real-time speech-to-text. The transcription streams word-by-word as the customer speaks so the reasoning engine doesn't need to wait for a completed utterance to start processing.

The ingress path also determines what data can travel with the call. SIP trunking allows metadata — caller ID, account context, prior session data — to be passed alongside the audio stream, giving the agent a head start before the customer finishes their first sentence. Agentforce Voice also supports PSTN to maximize compatibility with legacy phone systems. While this path offers less robust data passthrough, Agentforce can compensate through asynchronous data retrieval during the call.

Routing and conversation management

Transcribed text flows to the Salesforce Contact and Routing Technology (SCRT) platform, which manages the conversation session and handles Omni Routing. This determines whether an interaction stays with the AI or gets escalated to a live representative. SCRT is the layer that enforces routing rules, manages session state, and carries conversation context forward regardless of what happens next in the interaction.

On escalation, SCRT is also what carries full conversation context to the human representative — every customer utterance, every action the agent attempted, every piece of data it retrieved — populating the AI Work Summary automatically. The representative doesn't receive a transfer and a caller ID. They receive a case with context.

The reasoning engine

The Agentforce reasoning engine interprets intent, determines which actions to take, and executes them. It maintains context across turns so the conversation doesn't reset with every exchange. It grounds responses in enterprise data through Data 360, which connects to external systems without extracting or copying the underlying data. That zero-copy architecture doesn’t just eliminate the need to move and replicate data, it also helps secure legal and IT sign-off in regulated industries, where customer data can't leave its secure environment.

Native integration with Salesforce also means the reasoning engine has full access to the Customer 360 data model — not just records in a particular system, but a unified customer profile across all of them. That's what allows an agent to answer a warranty question, check order status, and route to the right team in a single conversation without any of those lookups requiring a separate API call to a separate system.

Trust and model infrastructure

Every response passes through the Generative AI Gateway before anything reaches an LLM or gets converted into speech. The gateway enforces the Agentforce Trust Layer, which ensures every response is screened for toxicity before being spoken back to the user. All agent responses are traceable to the knowledge source, mitigating the risk of hallucinations and helping validate where agents pull information from.

Every interaction is logged with zero data retention on the LLM provider side, meaning customer data used during a call isn't stored or used for model training. The audit log contains a full record of everything the agent retrieved, generated, and acted on.

The LLM Gateway supports multiple model providers — Anthropic, OpenAI, Google Gemini, and others — without requiring trust enforcement to be re-implemented for each one. Organizations can run different models for different agent types, or change providers as the landscape shifts, while the same trust controls govern all of them.

Speech output

Approved responses are converted to speech via ElevenLabs, with a library of 40+ voices and controls for speed, tone, and pronunciation. These attributes help shape how authoritative, warm, or neutral an interaction feels. Spoken responses are delivered back over the same path the call arrived on.

Two questions determine whether a voice AI interaction feels natural: how fast does the agent respond and what happens while it’s thinking. Both are design decisions.

From speech to response

When a customer speaks, the audio streams into Einstein Speech Foundations and begins transcribing immediately — word by word, not sentence by sentence. The reasoning engine starts processing the incoming text as it arrives, so by the time the customer finishes a sentence the agent has already begun interpreting intent and retrieving relevant data. That streaming approach cuts the effective latency of the interaction, because the system isn't waiting for a signal that the utterance is complete before doing any work.

That parallel retrieval is the key to staying within the 2-second action latency benchmark. Waiting to look up customer data until after the utterance is complete costs time the conversation can't spare. The strongest deployments go further. They retrieve data asynchronously before the customer has said anything specific — pulling account information and order history while the greeting is still playing. The data is available the moment it's needed, not the moment it's requested. An agent that already knows who it's talking to before the customer identifies themselves doesn't just feel faster. It feels intelligent.

What makes it feel like a conversation

Making voice interactions feel human is as much an art as it is a science. Three factors play an outsized role:

• Silence handling separates a natural interaction from a stilted one. A pause with no acknowledgment reads like a dropped connection. Simply having the agent say "one moment" while its processing keeps the conversation moving.
• Response length matters more in voice than chat. In chat, users can read at their own pace, skim, and re-read. In voice, a long response is a burden. The caller can't skim it, can't re-read it, and can't interrupt without the agent losing its place. Voice responses should stay under 50 words. If more detail is available, the agent should offer it rather than deliver it unprompted.
• Pronunciation errors are invisible in testing and conspicuous in production. Don’t rely on TTS to correctly pronounce program names, product identifiers, numerical codes or proper nouns. Be sure to validate pronunciations during pre-launch QA, not after the fact.

What live representatives see on escalation

Some service requests require human judgment, whether it’s a topic the agent isn't configured to handle, a customer expressing frustration, or a regulatory requirement that mandates human involvement. When a call is escalated, Agentforce generates an AI Work Summary containing the full conversation context. Representatives can review all the details before they pick up the phone, including what the customer asked for, what the agent attempted, and where the conversation stands. No re-explanation required on the customer's end.

How you design your escalation paths matters almost as much as the design of the agent itself. If your voice agent is deflecting 80% of calls autonomously but routing the remaining 20% to the wrong team without any context, your CSAT numbers will drop. The best voice deployments explicitly define which topics require a human, which emotional states trigger handoff and which teams receive which call types.

The voice AI deployments that deliver results fastest aren't the most ambitious ones. They're the most disciplined, scoped to use cases with high volume, predictable intent, and resolution paths that don't require human judgment.

What does that look like in practice?

For Adecco Group, the world’s leading talent and technology expertise company, Agentforce Voice powers candidate prescreening for thousands of job applicants. Recruiters who once reached only about 10% of candidates can now prescreen 100% of them, saving as much as 15 minutes per interview.

Florida Prepaid, a state-administered college savings program, uses Agentforce Voice for FAQ and policy answers, handling 75% of business-hours calls and 100% of after-hours cases. AI-handled calls average roughly 1.6 minutes against 10.9 for human-handled ones.

Canada Goose, the luxury apparel brand, uses Agentforce Voice to handle routine customer service inquiries. During peak demand, the agent cut wait times by 13.9% while CSAT rose 21.2%.

These are just a few examples of real customers driving real-world impact with Agentforce Voice. Below are five of the highest-value production use cases we’re seeing in live Agentforce Voice deployments.

The highest-volume contact driver for most service organizations is also the most containable: customers calling to ask something that already has a known answer. Policy questions, eligibility checks, coverage details, account status — these calls don't require human judgment. They require fast, accurate access to the right information.

What makes FAQ and policy calls a strong entry point for voice AI is the predictability of the interaction. The customer asks. The agent retrieves and responds. The call ends or escalates. There's no complex negotiation, no edge case requiring discretion, no transaction that needs verification before it can close. When the knowledge base is well-maintained and the agent is scoped to the use case, containment comes in high. An AI voice agent also means your call center never has to close. After-hours coverage is one of the easiest ROI cases to make, since every call that gets answered is a call that previously wasn't.

Order status, booking changes, cancellations and billing questions — these are the types of calls where the customer already knows exactly what they want and doesn't want to explain themselves to get it. They're also among the most well-suited for voice AI because the intent is narrow, the backend action is well-defined and the data exists in a system the agent can reach.

This use case is particularly compelling at scale because these calls tend to be the ones that spike during peak periods. An agent that handles these calls autonomously doesn't just reduce cost per interaction — it absorbs volume the team could have only handled through additional headcount.

But this deployment hinges on how natively your agent is integrated into your backend. If the agent can only look up an order but not make updates, it might deflect some calls but it will also frustrate many customers. The highest-performing deployments in this category are the ones where the agent has real-time read and write access to the systems that matter — order management, CRM, case records — so it can close the loop on the call without escalating.

Scheduling is one of the most voice-native use cases in customer service. It’s the kind of interaction people have always done over the phone because it's usually a back-and-forth rather than a simple information retrieval request. It also follows a predictable pattern: a customer needs a technician on Thursday afternoon; the agent checks availability, confirms the slot, and closes the call. The business case writes itself: shorter handle times, fewer no-shows and every call that used to require a human to check a calendar now gets handled autonomously.

Native integration is essential here. The agent needs real-time access to calendar availability. It needs to write a confirmed appointment back to the scheduling system. And it needs a confirmation loop that closes the call with a clear, spoken summary. When that confirmation lands in real time over the phone, no-show rates drop because the customer has an explicit verbal commitment rather than an email they may or may not open.

Authentication, password resets, account navigation, guided self-service — these make up a significant share of inbound call volume at most large service organizations, and they share a common trait: they follow a predictable sequence that doesn't require human judgment. Voice adds something chat-based support often lacks: the ability to verify identity conversationally, walk a customer through a process step by step, and confirm each action before the next one executes.

The escalation design matters more here than in most other use cases. A customer locked out of their account is already frustrated — a voice agent that can't resolve the issue and routes them to a hold queue exacerbates that frustration. The strongest deployments in this category treat escalation as a resolution path rather than a failure state. The agent gathers every piece of information a human representative will need before the transfer, so the representative can pick up mid-resolution rather than starting over.

Not every voice use case aims for full containment. Some of the strongest deployments focus not on resolving the call autonomously, but on understanding why the customer is calling and making sure they reach the right place, with the right context already in hand.

Legacy IVR does a version of this badly. It presents a menu, the customer picks the closest option, and the representative who picks up often starts from scratch anyway. An agentic router replaces that menu with a conversation: the agent asks an open question, interprets the response, and routes to the right team based on actual intent, not the nearest menu option. Even when full deflection isn't possible, the handoff arrives richer. The representative knows what the customer needs before saying a word.

In regulated industries, this pattern delivers value even when full automation isn't possible. The agent can gather intake information, verify identity, and scope the issue before a human representative is ever involved, compressing handle time on the representative's end even when the interaction itself requires human judgment to resolve.

The agentic routing pattern also works in use cases where the underlying task requires a human from start to finish. A legal firm can't let an AI resolve a client matter. A hospital can't let an AI prescribe medication. But both can use an AI to determine why someone is calling, verify their identity, and route them to the right person with the right context already attached.

The deployments covered in this guide share a pattern: well-defined use cases, a clear baseline to measure against and a backend integration deep enough to close the loop on a call. Adecco started with candidate prescreening. Florida Prepaid started with policy and eligibility questions. Canada Goose started with customer service. None of them started with everything.

Pick the use case with the highest call volume, the most predictable intent and the clearest resolution path. Prove that it works. Then expand. The organizations moving fastest on voice AI aren't the ones with the most ambitious roadmaps. They're the ones that started narrow, proved it early, and kept building.