Salesforce
Agentforce Implementation

Salesforce describes an AI agent as a system that can get work done proactively and autonomously. In Agentforce you define the agent's jobs to be done, instructions, actions and guardrails in natural language in Agent Builder.

Agents act through actions: Salesforce lets you turn existing workflows, prompt templates, Apex and APIs into agent actions. They can use native tools such as Data 360, Slack and MuleSoft, and they reach people on channels including CRM, WhatsApp, Messenger and websites.

Trailhead calls reasoning the brain behind every agent: Salesforce relies on the Atlas Reasoning Engine to understand what a person wants and to call the right subagents and actions as the conversation changes. Subagents group the actions that make up one job, and their natural language instructions tell the agent when to use each action.

A Salesforce Agentforce partner is a consulting firm that helps companies plan, build and run Agentforce agents. Salesforce launched the Agentforce Partner Network on 17 September 2024 as an ecosystem of partners building agents and agent actions, and said systems integrators support enterprises in defining, building and deploying AI agents and actions with Agentforce.

Techila Global Services is a Salesforce Summit Partner and a member of the Agentforce Partner Network, with 15+ Agentforce programs delivered. The table shows what an Agentforce partner should bring, and how to check each point.

Techila's Agentforce case studies are linked below; for a checklist, read how to choose an Agentforce implementation partner, or talk to our Agentforce team.

A Salesforce AI consultant decides where AI agents and generative AI fit in your Salesforce org, then designs, builds and governs them. In practice that means ranking use cases, checking the data and permissions an agent will use, writing its instructions and guardrails, turning existing Flows and Apex into agent actions, and setting up testing and monitoring. Salesforce offers an Agentforce Specialist certification for this role.

To start with a Techila Salesforce AI consultant, the Agentforce Readiness Assessment ranks your candidate use cases and ends in a plan for a first agent. For costs, see the Agentforce pricing guide.

Service Cloud suits agents that answer common questions and escalate the rest; Sales Cloud suits lead follow-up and meeting booking; Marketing Cloud suits campaign and brief drafting; Commerce Cloud suits guided shopping and product content; employee agents suit account summaries. The best first use case is the one with volume, a clear measure and clean data.

Swipe to see all columns →

Industry examples are on the Agentforce overview. If you are not sure which of these fits, the Agentforce Readiness Assessment ranks your candidates.

Salesforce notes that AI features and agent testing can lengthen an implementation, so testing is planned from the start. For costs, see the Agentforce pricing guide and the Agentforce cost calculator.

Grounding means adding the right information to the prompt so the answer is accurate. Agentforce grounds agents in Salesforce records and, through Data 360, in knowledge, files and other sources using retrieval augmented generation (RAG). Salesforce offers two routes: a quick start with Agentforce Data Library, or an advanced setup in Data 360.

Trailhead describes RAG as a form of grounding that retrieves knowledge from unstructured sources, and notes that unstructured content such as emails and documents needs preparation before it can be retrieved well. Structured data, such as Accounts and Cases, is easier for agents to use because its format is known.

For service agents, the Data Library uses RAG to index Knowledge articles and attachments, and Data 360 must be activated before you set one up. In practice, the quality of what the agent says depends on the quality of what it retrieves, so out-of-date or duplicate articles are fixed before launch. Techila's Data 360 practice handles the advanced setup.

Agents are tested in a sandbox with Salesforce's Testing Center, which runs batch tests that pair many example requests with expected responses and shows which subagents and actions the agent chooses. Manual checks in the builder's Conversation Preview complement it. Tests run in a sandbox because testing can change CRM records.

Trailhead explains why this matters: an agent can give different outputs for the same input on different runs, and manual testing does not scale. Testing Center tests are built from a CSV template with an utterances column, can be generated with AI from the agent's knowledge or its subagents and actions, and allow up to 1,000 test cases per test and up to 10 test jobs in a 10-hour window. Testing in a sandbox consumes credits, as customer use does.

• Core requests: the questions and tasks the agent should handle, with the expected answer or action for each.
• Edge cases: vague, off-topic or hostile requests, which the guardrails should block or redirect.
• Hand-offs: requests the agent should pass to a person, checked end to end on each channel.
• Permissions: requests for records or fields the agent should not reach, confirming it cannot.
• Deployment checks: after moving from sandbox to production, Salesforce's steps include creating the agent user in the target org, assigning its permission sets, confirming Einstein, Agentforce and Data 360 are enabled, checking the right flow versions are active and confirming the agent is active.

Agentforce runs with the Einstein Trust Layer, which Salesforce describes as a set of features and guardrails that protect the privacy and security of your data. Trailhead lists secure data retrieval, dynamic grounding, data masking, zero data retention, toxic language detection and an audit trail that tracks a prompt through each step of its journey.

One detail matters in design: Trailhead notes that data masking for large language models is currently disabled for agents. So the agent's access is controlled through the permissions of the user it runs as, and sensitive fields are kept out of its actions unless they are needed.

Which user that is depends on the agent. Employee-facing agents use the permissions of the person interacting with them; external-facing agents, such as a service agent on your website, run as a single dedicated agent user. Trailhead advises giving that user only the permissions it needs, because an agent can read any data its user can.

To choose a first use case, book a discovery call.

After launch, agents are watched with Agentforce Observability, which Salesforce says provides session-level tracing across every interaction, Agentforce Analytics, near real-time alerts on key metrics and Quality Scores for relevance and helpfulness. Governance adds versioning, filters on topics and actions, customer verification and a regular review of conversations.

Techila can run this review as part of Salesforce managed services, or train your admins to own it.

Effort depends mostly on how ready your data and knowledge are, how many actions the agent needs and whether they already exist as Flows or Apex, how many channels it serves, how strict the security review is and how much testing the use case demands. A first agent on one use case and one channel keeps effort contained.

License and usage costs are separate from implementation effort: see the Agentforce pricing guide and estimate usage with the Agentforce cost calculator.

Bring a shortlist of processes you would like an agent to handle, the numbers you measure them by today, a named business owner, and access to the knowledge, records and automation involved. Add your security and compliance requirements and your Salesforce editions and licenses. With these, a partner can scope a first agent quickly and accurately.

• Candidate use cases, each with its monthly volume and current measure, such as handle time or lead response time.
• A business owner who can decide what the agent may and may not do.
• An inventory of knowledge: where articles, policies and documents live and who keeps them current.
• Existing automation: the Flows, Apex classes and APIs that already carry out the tasks the agent would take on.
• Security requirements: data the agent must never see, approval steps and who signs off.
• Your org details: Salesforce editions, licenses, sandboxes and whether Data 360 is enabled.
• Channels where customers or employees should meet the agent.

Check the partner's AppExchange consulting listing for its tier and customer reviews, verify the certifications of the consultants who will do the work with Trailhead's credential verification, read case studies with agents in production, and ask who exactly will design, build and test your agent. Claims you cannot check should carry little weight.

Techila's AppExchange listing shows the Summit tier and a 4.82 rating from 19 reviews. For a fuller checklist, read how to choose an Agentforce implementation partner, and see our Agentforce case studies; more are linked below.

An Agentforce deployment moves a tested agent from a sandbox to staging or production. Salesforce's deployment steps cover the agent user, permission sets, the features the agent depends on, active flow versions and the agent's own status. The checklist below adds the testing and governance work that should be finished before the agent meets customers or employees.

• Test set signed off: Testing Center batch tests pass in the sandbox for core requests, edge cases, hand-offs and permissions, and the business owner has approved the results.
• Agent user created: the agent user exists in the target org and has only the permission sets it needs, because an agent can read any data its user can.
• Features enabled: Einstein, Agentforce and Data 360 are enabled in the target org, as the agent requires.
• Actions ready: the right versions of the Flows behind each agent action are active, and any Apex classes and API connections the actions call are deployed.
• Knowledge current: the knowledge articles and data the agent grounds on are up to date in the target org.
• Channel configured: the channel the agent serves is set up, and hand-offs to people are checked end to end.
• Agent active: the agent is activated in the target org, with one version active at a time.
• Monitoring in place: Agentforce Observability alerts, a conversation review routine and a named owner for pausing the agent are agreed before launch.

For the full sequence, see the steps of an Agentforce implementation and how agents are tested before launch.

Salesforce publishes prebuilt agents and industry data models for several industries. Our industry pages set out what Salesforce lists for each, and what a first agent could look like:

• Agentforce for Manufacturing: sales planning, asset service, warranty claims and channel management on Manufacturing Cloud.
• Agentforce for Consumer Goods: sales planning, retail execution, account management and trade promotions on Consumer Goods Cloud.
• Agentforce for Public Sector: complaints, benefits and licensing, job matching and compliance on Public Sector Solutions.

Start with an assessment if you have several candidate use cases, uncertain data or an untested security process. Start with a pilot if you already know the use case, its baseline measure and the data it needs. Both are scoped in a short call, and an assessment usually ends in a pilot plan.

Read more about the Agentforce Readiness Assessment and the Agentforce Pilot.