.gif){kind=spacer}
Agentforce Voice is Salesforce's voice agent capability inside the Agentforce platform, working from CRM data held in Salesforce. A purpose-built voice agent runs the workflow itself, executing tasks across the systems that hold the work. The distinction decides whether a call ends in a resolution or a routed ticket.
Voice is the channel where that difference is most visible. A caller will not wait through a handoff, and containment collapses the moment the agent has to ask a human to complete the task. This guide compares the two architectures on the dimensions that decide production outcomes: latency, containment, backend execution, governance, and cost.
What is Agentforce Voice vs a Purpose-Built Voice Agent?
Agentforce Voice acts as an agent layer over a CRM system. It reads records, answers from Salesforce data, and routes what it cannot handle. The system of record stays a database designed for human agents.
A purpose-built voice agent inverts that. The agent is the execution layer. It completes the task in whichever system holds it, an order management platform, a policy administration system, or a billing system, and escalates only genuine exceptions.
The architectural difference compounds over time. CRM extensions depend on manual updates when business rules change. Purpose-built platforms run feedback loops that diagnose failures and propose fixes, with a human approving each change before it ships.
The Five Dimensions That Decide Voice Outcomes
Here is how the two architectures compare on the criteria that matter once calls are live.
Latency
Voice punishes delay in a way chat does not. Every hop the agent makes to fetch or write a record eats the turn-taking budget, and callers interpret the silence as a broken system. Architecture that keeps retrieval close to execution has more room in that budget than one that crosses an ecosystem boundary for each step.
Containment
Containment measures the share of calls finished without a human. It is easy to inflate by counting deflection, so it is worth defining precisely during evaluation. A call that ends with a ticket is not contained; it is deferred.
Backend execution
This is the dividing line. Answering a question about an order is a retrieval problem. Changing that order is an execution problem, and it requires the agent to write to the system that owns it.
Governance
Neither architecture should ship changes unattended. An approve-to-promote model, where a person signs off before any change reaches callers, is what makes voice automation defensible when a regulator or a customer asks what happened on a call.
Cost
Per-resolution pricing looks clean and creates an unhelpful incentive at the margin, because a call that escalates costs both the resolution fee and the human minutes. Model the escalation rate, not just the headline rate.
Where Voice Fits in a Workflow Platform
Voice is one execution channel, not a category of company. Within NuPlay, voice and chat agents are part of NuPro, the execution layer, while NuStack makes the underlying systems agent-ready and orchestrates the workflow the call belongs to.
That matters for scoping. A voice deployment that cannot reach into the order, policy, or billing system is a deflection project. One that can is a workflow project that happens to start on the phone.
NuLoop is the closed feedback loop behind the improvement claim. It watches each run and routes validated fixes to whichever layer holds the problem, through Report, Diagnose, Propose, Try, Ship, and ships only after a human approves. Every task or decision runs as Human, Automated, or Agent, and the aim is moving the right work into agent mode rather than replacing people.
Voice at volume is a real test of that model. Working with the International Cricket Council, agent-led conversations handled more than 100,000 interactions in five days, the kind of scale where governance has to hold.
How to Choose
Start from the call types, not the vendor list. If most calls are answerable from CRM data and the team already runs Salesforce Service Cloud, a CRM-native voice agent is a reasonable fit with a short path to production.
If calls require action in systems outside that ecosystem, the evaluation changes. Ask which systems the agent can write to, what happens when a write fails mid-call, and who approves a change to the agent's behaviour.
Then model the cost honestly. Take the expected escalation rate, apply it to both the per-resolution fee and the human handling time behind it, and compare that against a contracted alternative.
Getting Started
Audit call reasons first and separate them into answerable, actionable, and genuinely exceptional. That split, more than any vendor feature list, determines which architecture fits.
Define execution modes for each call reason. Decide which need a human, which run on fixed rules, and which need an agent to reason. Pilot on a narrow, high-volume call reason with strict oversight.
Instrument containment and escalation from day one, and review what failed rather than only what succeeded. A closed feedback loop is only useful if someone reads what it reports.
Conclusion
Agentforce Voice and a purpose-built voice agent solve different problems. One extends a CRM into the voice channel; the other runs the workflow the call belongs to. The right choice follows from whether calls need answering or completing. NuPlay runs enterprise workflows in production and improves them after every run through NuLoop, with agents, the systems they operate, and the context they draw on under one platform. To see how voice fits inside a governed enterprise workflow, book a demo with the NuPlay AI team.
