Governing Salesforce Agentforce Engagements with the SAISA

Salesforce Agentforce provides the runtime. Your CRM provides the data. But what governs the engagement itself? The SAISA provides the contractual layer that wraps any runtime, including Agentforce.

Runtime Agnostic by Design

The SAISA does not specify where the agent runs. Article 1 defines the Runtime Provider as "the entity whose infrastructure executes the Agent" - but the Runtime Provider is NOT a party to the agreement.

SAISA Parties:
- Buyer (you)
- Developer (agent creator)

NOT Parties:
- Platform Operator (facilitates but doesn't contract)
- Settlement Provider (holds escrow but doesn't contract)
- Runtime Provider (executes but doesn't contract)
- AI Provider (inference but doesn't contract)

This means the same Paper can govern an agent running on Salesforce Agentforce, NVIDIA NemoClaw, Docker, Kubernetes, AWS Lambda, WebAssembly, or any MCP server. Runtimes execute. The Paper governs.

Agentforce Integration Pattern

A typical SAISA integration with Agentforce:

{
  "paper": {
    "id": "paper_agentforce_001",
    "runtime": {
      "provider": "salesforce_agentforce",
      "version": "2026.1",
      "region": "us-west-2"
    },
    "executionManifest": {
      "maxCostCents": 250000,
      "timelineDays": 5,
      "completionCriteria": [
        "All leads in Exhibit A scored and ranked",
        "Top 20% flagged for immediate follow-up",
        "Scoring rationale documented per lead"
      ],
      "permissionScopes": [
        "read_salesforce_leads",
        "write_salesforce_lead_scores",
        "read_salesforce_accounts"
      ],
      "allowedEgressUrls": [
        "api.salesforce.com",
        "login.salesforce.com"
      ]
    }
  }
}

Salesforce Data Considerations

When agents access Salesforce data, additional considerations apply:

Liability Allocation

The SAISA's bipartite liability framework applies regardless of runtime:

• Developer liability - Agent Logic, prompts, configuration, output quality
• Buyer liability - Salesforce data accuracy, access grants, usage of output
• Salesforce liability - Governed by Salesforce MSA, not the SAISA

Acceptance Criteria for Salesforce Agents

Good acceptance criteria for Salesforce-integrated agents:

{
  "completionCriteria": [
    // Coverage
    "Processes all records matching SOQL filter in Exhibit A",
    "No records skipped due to agent errors",

    // Output quality
    "All updated fields conform to picklist values",
    "No data validation errors on save",

    // Business logic
    "Scoring algorithm matches specification in Exhibit B",
    "Edge cases in Exhibit C handled correctly",

    // Audit trail
    "All modifications logged with agent attribution",
    "Summary report generated in specified format"
  ]
}

Enterprise Considerations

For enterprise Salesforce deployments:

• Sandbox testing - Specify sandbox vs. production in Paper
• Change management - Reference Salesforce release calendar in timeline
• User permissions - Document required Salesforce permissions in Execution Manifest
• Rate limits - Account for Salesforce API limits in budget calculations

Example Use Cases

SAISA-governed agents on Agentforce: