Salesforce Agentforce is Salesforce’s platform for building AI agents that work inside the CRM, answering customers, qualifying leads or helping employees by taking actions on Salesforce data. Implementation succeeds when data is clean, agent scope is narrow, actions are tested and humans can take over. This checklist follows the path from a first pilot to a production rollout.
What is Salesforce Agentforce?
Agentforce lets companies configure AI agents in Salesforce using topics, instructions and actions. A topic defines a job the agent handles, such as order inquiries. Instructions describe how to behave. Actions are what the agent can do: run a Flow, call Apex, query records or use a prompt template. The Atlas reasoning engine decides which topic and actions fit a request.
Agents can run in channels such as web chat, Slack or within Salesforce for employees, and they work with the Einstein Trust Layer, which handles data masking and other safeguards. Licensing and consumption pricing have changed since launch, so confirm the current model with your account executive before planning a budget.
Is your data ready for an AI agent?
An agent is only as good as the records it reads. If account owners are wrong, cases lack categories or knowledge articles are outdated, the agent will repeat those problems quickly and at scale.
- Audit duplicate accounts and contacts, and run deduplication rules
- Check required fields for the objects the agent will use
- Review Knowledge articles for accuracy, owners and last review dates
- Confirm field-level security so the agent sees only what its user profile should see
- Map where key data lives outside Salesforce and whether Data Cloud or integrations are needed
Pilot checklist: scope and design
Pick one use case with high volume, clear rules and low risk. Service order status, appointment rescheduling and internal sales research are common first choices. Refund approvals and contract changes are poor first choices.
- Write the use case as a sentence: “The agent answers delivery status questions for web customers and hands over billing questions”
- Define topics with clear scope and explicit out-of-scope examples
- Build actions with existing Flows where possible, since they are already tested
- Write instructions in plain language with examples of correct behaviour
- Set handover rules to an Omni-Channel queue with conversation context
- Agree success measures before the pilot starts
Guardrails and trust settings
Review Einstein Trust Layer settings with your security team. Decide which fields should be masked, how prompts and responses are logged and who can read the audit trail. Agent permissions should follow least privilege: an agent that answers order status questions does not need write access to opportunities.
How should you test an Agentforce agent?
Testing needs both expected and unexpected input. Salesforce provides a testing center for agents where you can run batches of test utterances; use it, but also have real staff try to break the agent.
- Build a test set of 50 to 100 real customer phrasings per topic
- Include misspellings, mixed languages and off-topic requests
- Test what happens when data is missing, such as an order number not found
- Check that the agent refuses actions outside its scope
- Confirm handover keeps the conversation history
- Re-run the full test set after every instruction or action change
Production rollout checklist
Roll out in stages. Start with a share of traffic or one region, compare results with the human-only channel and widen only when targets are met.
Prepare the service team before launch. Agents will receive handovers with summaries, and they need to know how to read them and when to flag bad agent behaviour. A simple feedback button or Slack channel for agent errors speeds up fixes.
Set a review rhythm: weekly in the first month, then monthly. Each review looks at transcripts, failed actions, handover reasons and any new questions that need a topic or a knowledge article.
Which metrics prove Agentforce is working?
Pick metrics that link to cost and customer experience, and compare them with a baseline from before launch.
- Resolution rate without human help, per topic
- Handover rate and the top handover reasons
- Average handling time for cases that reach a human after the agent
- CSAT for agent conversations compared with human conversations
- Action failure rate from Flow or Apex errors
- Consumption cost per resolved conversation
How Alien Road applies Agentforce in CRM projects
A typical engagement starts with a data readiness review: duplicate rates, field completeness and Knowledge quality for the objects the agent will touch. The agency also exports recent cases to find the intents with the most volume.
Setup is done with the client’s Salesforce admin. The team designs topics and instructions, builds or reuses Flows for actions, configures handover to existing queues and writes the test set from real case text. The pilot runs on one channel with limited traffic.
Reporting compares resolution, handover and CSAT with the pre-launch baseline, and lists the top failure reasons with the fix applied. Expansion to a second topic happens only after the first meets its targets for a full month.
What usually goes wrong in Agentforce projects?
The most common failure is scope creep in the pilot. A team starts with order status, then adds returns, then billing, and the agent’s topics begin to overlap. When topics overlap, the reasoning engine may pick the wrong one and give confusing answers. Keep topics distinct and add new ones only after the current set performs well.
The second failure is treating instructions as a place to fix data problems. If the agent gives wrong delivery dates because shipment records are late, no instruction will solve it. Fix the integration first.
- Overlapping topics with vague descriptions
- Actions that fail silently when a field is empty
- No owner for Knowledge articles after launch
- Testing only in English when customers write in Turkish or German
- Measuring only deflection and ignoring customer satisfaction
Where to start this week
Export the last three months of cases or chat transcripts and count them by reason. Choose the reason with the most volume and the simplest rules as your pilot topic.
Then check the data behind it: are the fields the agent would need actually filled in, and are the related Knowledge articles current? Fixing those records now will save more time than any prompt tuning later.
Keep reading: CRM System · Crm System