Skip to main content
Salesforce Agentforce Agent counting tokens
5 min read

Agentforce Flex Credits: The Math Behind the Most Affordable AI Agent Platform

The Pricing Shift That Changed the Math

For most of its first year, Agentforce ran on a simple pricing model: $2 per conversation. It didn't matter whether the agent answered one question or executed thirty actions across a complex support case. Every conversation cost the same flat rate.

That model is gone. Salesforce replaced it with flex credits, a per-action pricing structure that charges $0.10 for each discrete action an agent performs. One action covers up to 10,000 tokens of processing. For organizations running high-volume, lower-complexity agents, the savings are immediate and significant.

The math tells the story. A typical Agentforce conversation involves five to fifteen actions. Under the old model, that's $2. Under flex credits, it's $0.50 to $1.50. A simple FAQ interaction with one action and one answer costs $0.10. That's a 95% reduction compared to the conversation model.

Old Model vs. New Model

Conversation model: $2 flat per conversation, regardless of complexity or number of actions.

Flex credits: $0.10 per action (20 credits). Each action covers up to 10,000 tokens.

Break-even: ~20 actions per conversation. Below that, flex credits are cheaper.

Typical conversation: 5-15 actions = $0.50-$1.50 on flex credits vs. $2 flat.

Real Scenarios, Real Costs

Abstract numbers only matter when applied to use cases organizations actually run. Here's how the math breaks down across four common agent types.

An FAQ or knowledge base agent handles the simplest interactions. One action per query. At $0.10 per action, an organization fielding 5,000 questions a month pays $500. Under the old conversation model, that same volume cost $10,000.

Case management agents require more work. Three actions per interaction to gather context, classify the issue, and route or resolve puts the cost at $0.30 per case. At 5,000 cases a month, the bill comes to $1,500.

Field service scheduling sits in the middle. Six actions per interaction for checking availability, matching technicians, and confirming appointments runs $0.60 each. Five thousand monthly interactions cost $3,000.

SDR outreach is the outlier. These agents perform thirty-five or more actions per sequence: researching accounts, crafting personalized messages, logging activities, and following up. At $0.10 per action, that's $3.50 or more per conversation. Here, the old $2 flat rate was actually cheaper. It's a critical distinction most pricing guides skip entirely.

Monthly Cost by Agent Type (5,000 Interactions)

$500

FAQ Agent: 1 action per query at $0.10 each

Salesforce Pricing
$1,500

Case Management: 3 actions per case at $0.30 each

Salesforce Pricing
$3,000

Field Service: 6 actions per interaction at $0.60 each

Salesforce Pricing

For context, the average human-handled customer service interaction costs roughly $4.60 in 2026. Even the most action-heavy Agentforce agents come in well under that threshold. And for organizations currently paying $1.50 to $2.00 per interaction on other AI tools or the legacy conversation model, the potential savings on high-volume, low-complexity use cases are hard to ignore.

The Hidden Multiplier No One Talks About

Every $0.10 action assumes the agent stays within a 10,000-token ceiling. That's roughly 8,000 words, more than enough for most interactions. But when an action exceeds that threshold, Salesforce bills it as multiple actions.

An action that consumes 15,000 tokens counts as two actions: $0.20 instead of $0.10. At 20,001 tokens, it counts as three: $0.30. These costs multiply silently, and most organizations don't catch it until they're reviewing a surprisingly large invoice.

Understanding token composition is critical here. Every action's token count is the sum of system instructions, user input, prompt template text, retrieved data chunks, and the generated response. A knowledge base article that runs 5,000 words, retrieved in full as grounding context, can push a single action well past the 10,000-token line and turn what should be a $0.10 interaction into $0.20 or $0.30.

The Art of Keeping Tokens Down

Token optimization isn't a one-time configuration. It's an ongoing discipline that separates cost-effective Agentforce deployments from expensive ones.

Start with your knowledge base. Long, unstructured articles are the most common culprit for token overflow. Condense them. Write summaries. Structure content so agents can retrieve the relevant section rather than the entire document. The prompt instruction "retrieve concise, summarized content rather than full documents" sounds obvious, but it's the single most impactful change most implementations can make.

Limit your retrieved chunks. If your agent pulls ten knowledge articles when three would suffice, you're burning tokens on context that doesn't improve the response. Use top-K retrieval settings and pagination to control how much data enters the prompt window.

Keep system instructions lean. Every word in a prompt template counts toward the token total. Verbose instructions that repeat guidance or include unnecessary caveats add up fast, especially at scale across thousands of daily interactions.

Test across varied data. Token counts aren't static. A case record with two comments processes very differently than one with fifty. Test your agents against real data with varying complexity levels to understand the range of token consumption before you go live.

Use the sandbox. Salesforce charges 16 credits per action in sandbox environments versus 20 in production, a built-in 20% discount for testing. There's no reason to debug token overflows in production when the sandbox gives you cheaper iterations.

Monitor with Digital Wallet. Salesforce's consumption dashboard tracks usage by agent and action type in near real time. Set threshold alerts. Review consumption weekly. The organizations that manage costs effectively are the ones watching the numbers, not estimating them.

Token Optimization Checklist

1

Groom Your Knowledge Base

Condense long articles into summaries. Structure content so agents retrieve relevant sections, not entire documents.

2

Limit Retrieved Chunks

Use top-K retrieval and pagination. Don't pull ten articles when three will do.

3

Keep Prompts Lean

Every word in system instructions counts toward the token total. Cut verbose or repeated guidance.

4

Test With Varied Data

Token counts change by record. Test with both simple and complex records to understand your range.

5

Use the Sandbox

16 credits per action vs. 20 in production. Debug overflows where it's 20% cheaper.

6

Monitor With Digital Wallet

Track consumption by agent and action type. Set threshold alerts. Review weekly.

Pick the Right Model for the Job

Flex credits aren't universally cheaper. The break-even point sits at roughly twenty actions per conversation. Below that, flex credits win. Above it, the $2 conversation model is more economical.

One thing to note: organizations choose between flex credits and conversation pricing at the org level, not per agent. Flex credits themselves are agnostic — they power any agent type, from a simple FAQ bot to a complex SDR workflow. But the billing model applies across the board, so it's wor

[@portabletext/react] Unknown block type "sanityImage", specify a component for it in the `components.types` prop

th running the math on your full mix of use cases before committing.

For organizations still evaluating, the Foundations edition includes 200,000 free flex credits. That's enough to run a genuine pilot across real use cases and gather actual consumption data before committing to a pricing model.

Here's the bottom line: Agentforce is one of the most affordable AI agent platforms available when deployed with intention. Organizations paying $0.10 per interaction and those paying $0.30 are often running the same agent. What separates them is how they built it.

Hunter, AgencyQ team member

Hunter Savage

VP, Salesforce Practice

Stay Informed

Get industry-leading insights delivered to your inbox.