How to Build a Salesforce AI Agent (Playbook)

Salesforce is the customer system of record for many organizations, with a permission model, validation rules, and automation built over years. An AI agent that works inside Salesforce must inherit all of it. This playbook covers building a Salesforce AI agent that respects permissions, uses typed data access, acts through approvals and existing automation, and is governed through the platform, following FISTA's AI agents practice. The platform-agnostic design is in how to build an ai crm assistant; the sales workflow view is in how to build an ai sales assistant.

What does the agent do?

Capability	Salesforce mechanism	Control
Summaries and answers	Object and field reads; connected knowledge	User permissions; citations
Record updates	Typed update tools	Validation rules; approval gates
Case handling	Case reads, comments, routing, response drafts	Service permissions; approvals
Email drafting	Drafts into the user's workspace	User sends
Automation	Invoke existing flows as tools	Flow permissions
Audit	Field history, event monitoring, agent logs	Governance

Step 1: Specify with sales and service operations

Define objects, fields, and record types in scope; summary formats; update types and their rules; case-handling scope; flows the agent may invoke; prohibited actions (no deletions, no ownership or sharing changes, no external sends); and quality thresholds. Operations teams own the rules. See how to write an ai spec.

Step 2: Design delegated access

Authenticate users through the identity provider and call Salesforce with each user's delegated authorization so object permissions, field-level security, sharing rules, and record types apply. Reserve any integration user for background jobs that produce flags, never for user-facing reads or writes. Access design is in ai access control.

Step 3: Build typed data tools

Expose reads as typed tools per object with defined fields and filters rather than free-form query generation, which is hard to validate and easy to abuse. Include related-record retrieval for context. Connect knowledge sources through the retrieval layer with citations. Tool design is in how to build tool use for llm agents.

Step 4: Build actions behind validation and approvals

Implement update tools that propose changes with validation against field types, picklists, and required fields, apply on user approval, and let the org's validation rules and approval processes fire as they would for manual changes. Classify consequence: stage, amount, and close-date changes are user-approved; activity logging may become automatic with sampling. Gate design is in what is a human approval gate.

Step 5: Integrate with flows and automation

Expose selected flows as tools so the agent reuses tested automation for multi-step operations, and ensure agent-initiated changes trigger existing automation consistently. Avoid duplicating logic that already lives in flows. Integration principles are in ai integration services.

Step 6: Handle service cases

For service scope, the agent summarizes cases with history, drafts responses grounded in knowledge with citations for agent approval, proposes routing under the org's rules, and never closes cases or contacts customers without approval. Patterns are in how to build an ai customer service agent and how to build an ai ticket routing system.

Step 7: Govern through the platform

Use field history tracking, event monitoring, and the agent's own logs to audit agent actions; manage the agent's access through permission sets; version prompts and tools through the prompt management system; and review permissions and usage on a schedule. Governance framework is in the agentic AI governance whitepaper.

Step 8: Evaluate and pilot

Build a golden set of records, cases, and scenarios verified by operations. Measure summary accuracy, update correctness against validation rules, permission compliance (no data beyond the user's rights), approval compliance, routing accuracy, draft quality, and injection resistance through case descriptions and emails. Pilot with one team, measure behaviorally, and expand. Harness design is in how to build an agent evaluation harness.

Worked example: a software company's sales and service org

A software company deploys the agent for account executives and support agents. Account executives get opportunity summaries citing activities and proposed stage and next-step updates from meeting notes, validated against picklists, applied on approval, with the org's approval process firing for discount changes. Support agents get case summaries and grounded response drafts with citations, and proposed routing under existing assignment rules. A renewal-preparation flow is exposed as a tool so the agent triggers tested automation rather than reimplementing it. Delegated authorization means a sales user cannot see cases outside their sharing rules, confirmed by evaluation. Event monitoring and field history provide audit. After a pilot, activity logging becomes automatic with sampling while stage changes remain approved, and forecast hygiene metrics improve.

What does it cost to run?

Cost scales with users and interactions and is modest; build cost is driven by scope of objects, flows, and knowledge sources. Value is measured in time saved, data quality, and case handling metrics. Drivers are in crm ai integration cost.

What are the common mistakes?

• An administrative integration user for user-facing operations.
• Free-form query generation instead of typed tools.
• Bypassing validation rules and approval processes.
• Reimplementing logic that already lives in flows.
• No use of platform audit features.
• Evaluating without operations verification.

Who owns the agent?

Sales and service operations own scope and rules, the Salesforce admin team owns permissions and platform governance, and engineering owns tools, evaluation, and the gateway.

How do you roll out across a large sales organization?

Start with one team whose managers are engaged, one object family, and read-only actions; publish what the agent did each week; add write actions behind approvals once accuracy is proven; then expand team by team with the same measurements. Sales organizations adopt what their peers visibly benefit from, and they reject anything that surprises a rep in front of a customer.

How FISTA Solutions builds Salesforce agents

FISTA Solutions builds Salesforce AI agents to this playbook: operations-owned specifications, delegated user authorization, typed data tools, actions behind validation and approvals, reuse of flows, service case handling with grounded drafts, platform-based governance, and operations-verified evaluation. The AI agents practice delivers the agent, AI enablement the retrieval and tool platform, and forward deployed engineers embed with your Salesforce admins and operations teams. The record behind the work is 150+ projects with 99.9% uptime.

To scope a Salesforce agent, message FISTA on WhatsApp, or read ai for sales teams for the functional context.