Key takeaways
- Agentforce is strongest when the work lives in Salesforce: it uses your Salesforce data, permissions, flows and Apex as agent actions.
- Custom agents built on AI model APIs or automation platforms offer more flexibility across systems and models.
- Compare pricing models: Agentforce offers per-conversation, Flex Credit and per-user add-on options; custom agents pay for model usage plus build and operations.
- Many organizations use both, with clear boundaries between them.
Once a business decides to deploy AI agents, the next question is where to build them. Salesforce customers can use Agentforce, which is built into the platform. Alternatively, agents can be built with AI model APIs, agent frameworks or automation platforms. Both paths are valid.
What Agentforce brings
- Native access to Salesforce data, respecting users’ sharing and permissions.
- Existing flows, Apex and prompt templates usable as agent actions.
- Channels for customer-facing agents and assistance for employees inside Salesforce.
- The Einstein Trust Layer for data masking and zero-retention agreements with model providers.
- Testing, monitoring and governance tools inside the platform.
What custom agents bring
- Freedom to choose and switch models from different providers.
- Easier orchestration across many systems outside Salesforce.
- Full control over prompts, tools, memory and user experience.
- Deployment wherever the users are, such as a website, app, Slack or Teams.
Pricing models compared
| Agentforce | Custom agent | |
|---|---|---|
| Usage pricing | Per conversation, or Flex Credits per action (Salesforce lists $500 per 100,000 credits, 20 credits per standard action) | Model provider pricing, usually per token |
| Per-user options | Add-ons for unmetered employee use, listed from $125 per user per month | None inherent |
| Build and operations | Configuration plus Salesforce development for actions | Engineering, hosting, monitoring and security |
List prices at the time of writing; Salesforce pricing changes and contracts vary. Model API prices also change frequently.
A decision framework
| Question | Points towards |
|---|---|
| Does the work mainly read and update Salesforce records? | Agentforce |
| Do agents need many systems outside Salesforce? | Custom, or Agentforce with integrations |
| Is there strong Salesforce admin and developer capacity? | Agentforce |
| Do you need a specific model or full control of behavior? | Custom |
| Is governance inside one platform a priority? | Agentforce |
Using both
A common pattern is Agentforce for service and sales agents working on Salesforce records, and custom agents for tasks centered elsewhere, such as document processing pipelines or internal knowledge assistants spanning many tools. Define which agent owns which process, and share knowledge sources where possible.
Whichever you choose
- Start with one narrow, high-volume use case.
- Clean the data and knowledge the agent will use.
- Apply least privilege and approvals for sensitive actions.
- Test with real examples and measure resolution and accuracy.
We build agents on Agentforce and on custom stacks; see our AI solutions.
A worked example
A services company wants an agent to answer customer questions about appointments and invoices, and another to process incoming documents from partners. Appointments and invoices live in Salesforce, so the customer agent is built with Agentforce, using existing flows as actions. Document processing spans email, a file store and an ERP, so it is built as a custom agent in an automation platform, writing results to Salesforce through the API.
Frequently asked questions
Is Agentforce only for Salesforce data?
It works best with Salesforce data, but actions can call external systems through integrations, flows and Apex.
Are custom agents cheaper?
Model usage can be inexpensive, but build, hosting, monitoring and maintenance add up. Compare total cost for your volumes.
Can custom agents work with Salesforce?
Yes, through Salesforce APIs or protocols such as MCP, with an integration user scoped to the needed permissions.
