Salesforce AI agents: How they work and where to use them

Salesforce built AI agents to move CRM AI from helping people understand data to helping them act on it, one CRM's version of the broader shift covered in our guide to agentic AI for sales.

These agents use CRM data and business context to update records, research accounts, flag issues, and trigger next steps, while keeping people involved when judgment or approval is needed. It's one implementation of the broader category covered in our guide to AI agents for CRM.

But how do Salesforce AI agents work? What can they do with your CRM data? And which workflows should you trust them with?

This guide answers those questions with practical examples for revenue teams.

In short:

• A Salesforce AI agent reads CRM and conversation context, decides what to do, and acts on its own. See what separates an agent from a chatbot.
• Every agent runs the same trigger, context, reasoning, action loop before it hands off. Here's how that loop works.
• Four workflows are worth starting with: CRM updates after calls, stalled-deal alerts, pre-call briefs, and missing-data checks. See all four workflows.
• Automation runs a fixed rule. An agent works out the rule itself, case by case. See how the two compare.
• An agent is only as good as the context it can reach, CRM fields plus real conversation data. More on why context matters.
• Low-risk actions can run unsupervised. High-impact ones need a person to approve first. See where to draw that line.

What is a Salesforce AI agent?

Salesforce describes an AI agent as an autonomous, proactive application that uses a large language model to understand context and reason through what to do next, rather than waiting for someone to type a request each time it needs to act.

Agentforce is the platform Salesforce built for this: Prebuilt agent roles for service, sales, and other functions, plus tools to configure custom ones on top of your own CRM and knowledge data.

Three things separate an agent from a plain automation or a chatbot.

• Goal. It's working toward an outcome, not just answering a single question.
• Reasoning. It figures out what needs to happen based on the data in front of it, not a fixed script.
• Action. It does something (update a record, draft a task, send an alert) instead of only returning text.

Here's the fastest way to feel the difference. A copilot handles "summarize this demo call." An agent handles "check every demo this week, find the questions that never got answered, and draft a follow-up for the rep to send."

Salesforce's own agent library shows the range this can cover.

• A Sales Development Representative (SDR) agent qualifies leads and books meetings around the clock.
• A Sales Coach agent runs role-play sessions grounded in real deal data.
• Agentforce for Service resolves customer cases across channels and escalates to a person when a request goes beyond its scope.

These sit alongside the wider set of jobs covered in our guide to AI sales agents. Different jobs, same shape: Read context, reason, act, hand off when it should. That shape shows up again in the four revenue workflows worth setting up first.

How AI agents work with Salesforce

Every agent runs on the same basic loop, whatever job it's built for.

Something has to start the loop. An agent doesn't sit around thinking. It needs a trigger: A meeting ending, a record changing, a schedule firing, a new lead landing.

Once triggered, it pulls context and plans. Salesforce's Atlas Reasoning Engine, the model that sits under Agentforce, breaks a request into smaller steps and works out a plan before acting. That planning step is what separates an agent from a single-prompt tool.

It can check something, decide it needs more information, check again, and only then commit to an action. Not every action earns the same trust level, and where a person should step in matters as much as the reasoning itself.

What can AI agents do in Salesforce?

Skip the list of twenty theoretical use cases. Here are four revenue workflows where agentic behavior is easy to picture and worth the setup time.

Keep CRM records updated after customer conversations

Reps finish calls, but Salesforce doesn't reflect what changed. An agent reviews the conversation, identifies what moved (next step, stakeholders, qualification detail) and updates the CRM record automatically after every call, or recommends the update.

Keep deal stage, amount, and close date behind a human approval. Let it handle the rest on its own.

Find stalled deals before pipeline review

An opportunity can look healthy in Salesforce while buyer activity has quietly stopped. An agent checks open opportunities against activity, next steps, and recent conversations, and surfaces the ones carrying real deal risk.

Cascade Robotics has had no buyer activity for twelve days, and no next meeting is on the calendar. The last call points to security approval as the blocker in procurement.

The AE reviews the evidence and decides how to re-engage. This is the workflow worth building first, since it fixes a problem every pipeline review runs into.

Brief reps before sales meetings

A rep has ten minutes before a call and needs the full picture across an account. An agent pulls CRM history, recent conversations, stakeholders, open opportunities, and outstanding commitments into one pre-call brief, instead of the rep hunting across five tabs.

Flag deals with missing qualification data

Salesforce says a deal is moving. The record is missing the economic buyer, the decision criteria, or the procurement timeline.

An agent checks the CRM record and the conversation history for these gaps and asks the owner to fill them in. This checks whether CRM reality matches buyer reality, which is a different job than simply updating fields.

AI agents vs Salesforce automation: What changes

Automation: If a demo call ends without a follow-up logged within 24 hours, remind the rep. The condition and the action are both fixed in advance.

Agent: Review every demo call, find the questions or objections that never got a response, draft one, and queue it for the rep to send.

Automation runs a rule someone already wrote. An agent works out what the rule should be, case by case, when reaching the goal takes judgment along the way.

The two aren't competing. An agent can decide what should happen, then hand execution to a Flow, an API call, or automation that already exists in your org.

Agents work better when they have the right context

Salesforce might show:

Stage: Negotiation. Amount: $120,000. Close date: September 30.

The last three customer calls might show something different: Legal hasn't approved the contract, procurement wants revised pricing, and the economic buyer still hasn't joined a call. Gaps like this widen with every extra stakeholder in the deal, which is why our guide to enterprise sales treats multithreading as a first-class problem, not a footnote.

Both pictures matter, and most agent deployments only give the agent one of them. CRM fields are structured context. Meeting notes, objections, and next-step commitments are unstructured context, and they're where most of the deal's real status lives.

An agent's action is only as reliable as the context it can reason over, and conversation intelligence turns those unstructured calls into something an agent can read alongside the CRM record. That gap between what's in the CRM and what's on the call is exactly where Avoma fits in.

Where should a person stay in the loop?

The goal isn't maximum autonomy. It's giving an agent enough room to remove repetitive work while keeping a person responsible for anything that materially affects a customer, a deal, or the forecast.

The bottom line

Salesforce holds the structured system of record for a deal. AI agents get more useful once they can combine that record with what's being said on the calls behind it, and turn the combination into a next step someone can act on. The same pattern applies to teams running HubSpot instead, covered in our guide to HubSpot AI agents.

Avoma connects customer conversations to your CRM, so the context a Salesforce agent reasons over includes what a buyer said on the call, not just what a rep remembered to log after the fact.

That's the difference between an agent working off a CRM field that hasn't been touched in weeks, and one that can see what happened on the call itself.

See how Avoma syncs conversation data with Salesforce →