Salesforce Agentforce is a platform for building agents that use business data and defined actions to answer questions or execute CRM-connected workflows. It can automate bounded work, but it is not an unsupervised “digital employee.” Results depend on instructions, permissions, grounding data, action logic, testing, monitoring, and human escalation.
This guide was updated on September 8, 2026 from current Salesforce product, help, testing, and pricing documentation. We did not deploy Agentforce in a Salesforce organization or run a controlled cost, accuracy, or productivity benchmark. Vendor examples are not independent evidence of return on investment.
How Agentforce is structured
| Layer | Purpose | Main failure to test |
|---|---|---|
| Subagents | Define bounded jobs and the instructions for handling them | Wrong intent classification, conflicting instructions, or tasks outside the approved scope |
| Knowledge and data | Ground responses in CRM records, Data 360, knowledge, or connected sources | Stale, incomplete, unauthorized, duplicated, or poorly permissioned information |
| Actions | Call flows, prompts, Apex, APIs, or standard functions to read or change systems | Incorrect inputs, excessive permissions, irreversible side effects, replay, and partial failure |
| Reasoning and response | Select a subagent and actions, then compose an answer or plan | Hallucination, skipped policy, unsupported conclusions, and inconsistent multi-turn behavior |
| Channels and monitoring | Expose the agent to customers or employees and review usage | Authentication gaps, missing escalation, poor audit evidence, and uncontrolled cost |
Suitable first use cases
Start with high-volume, reversible tasks that have clean data, a clear correct answer, and an established human fallback. Examples include retrieving an authenticated order status, summarizing a case for an employee, classifying a request, or drafting—but not automatically sending—a routine follow-up.
Avoid autonomous deployment first in refunds, credit, hiring, health, legal advice, safety, account termination, identity decisions, or other workflows where one wrong action can materially harm a person or create an irreversible transaction.
Current pricing is usage-dependent
Salesforce currently documents several Agentforce buying models, including Flex Credits, conversation-based pricing, and per-user access. Its public pricing page lists Flex Credits at $500 per 100,000 credits and conversation pricing at $2 per conversation, while individual action and voice-action multipliers are defined in a separate rate card. These prices can change and may exclude Data 360, Salesforce licenses, implementation, integration, model, support, and human-review costs.
Do not estimate cost from conversation count alone. A single user goal may trigger several actions, prompts, data operations, retries, or voice steps. Use Digital Wallet or equivalent usage records and calculate cost per successfully completed, audited business outcome.
Testing before deployment
Salesforce's Agentforce Testing Center documentation describes controlled testing of topic classification, action sequences, multi-turn conversations, responses, and predefined ground truth. LLM-as-a-judge evaluation can assist review but should not be the only judge for security, financial, regulatory, or factual correctness.
- Create a representative test set from authorized, de-identified historical cases.
- Define the expected subagent, permitted actions, response facts, refusal behavior, and escalation for each case.
- Add adversarial inputs: prompt injection, conflicting instructions, missing records, ambiguous identity, excessive requests, and integration outages.
- Test least-privilege access with several user roles and confirm the agent cannot reveal or modify records the user cannot access.
- Run write actions in a sandbox and verify idempotency, approval gates, rollback, retry, and audit logging.
- Measure false resolution, incorrect action, escalation, abandonment, correction time, latency, and credits consumed.
How to calculate ROI
| Measure | Recommended calculation |
|---|---|
| Verified completion rate | Correct completed outcomes divided by eligible requests; do not count fluent but wrong responses |
| Net time saved | Baseline handling time minus agent runtime, review, correction, escalation, and incident work |
| Total operating cost | Credits, licenses, data, models, integrations, implementation, monitoring, support, and human review |
| Cost per correct outcome | Total operating cost divided by independently verified successful outcomes |
| Risk-adjusted benefit | Gross labor or revenue benefit minus error, remediation, compliance, and customer-impact costs |
Use an A/B or phased rollout where possible, compare like-for-like case types, and publish the sample period and exclusions. The old claims on this page about 15–30% conversion gains, 45% savings, and 10–20% Salesforce revenue growth had no adequate supporting evidence and have been removed.
Bottom line
Agentforce is strongest when Salesforce data and workflows are already well governed and the agent has a narrow job, explicit actions, measurable ground truth, and a reliable human fallback. Build the financial case from verified outcomes and full operating cost, not vendor examples or speculative market-size projections.