What is Agentforce? From Coworker to a customized agent

Agentforce is Salesforce’s AI agent platform. In practical terms, it provides assistants capable of understanding a request, determining the appropriate actions to take, and executing them directly within your CRM.

Since early September, Salesforce has been enabling this feature by default in eligible organizations. There is no cost involved, and no agents are assigned to it, but the tools are now available. Our clients are quick to ask: Where do we start? What kinds of projects can we implement?

We address this by distinguishing three levels, each of which is aimed at a different audience and requires a different level of commitment.

Coworker replaces the Salesforce search bar. It is designed for your employees—not your customers—and draws on data from your CRM. Unlike an agent, which must be configured step by step, it natively understands your org’s metadata, including your custom objects and fields. As a result, it generally adapts quite well to highly customized orgs.

Where a standard search takes you to an account’s details page, Coworker can explain what’s happening there. A sales representative preparing for a meeting can ask Coworker what has changed with their client since their last visit. Depending on the account, they might see a drop in orders for a particular product line, pending complaints, or an internal discussion about a discount—or even receive a suggested agenda.

To estimate its value, start with your numbers: How much time do your sales reps spend putting together a client profile before each visit?

Before opening it, review your sharing rules. Coworker only shows what the user is authorized to see. If those permissions are too broad, information that no one would have looked for can be summarized in a matter of seconds.

An agent is defined as a role, with a scope, instructions, authorized actions, and potentially escalation rules. When faced with a request, the reasoning engine interprets it, selects the appropriate actions, and executes them sequentially.

A chatbot, on the other hand, follows a pre-written dialogue tree, whereasan agent constructs a response in real time and acts directly on your data.

Salesforce provides templates for common use cases, including seven new ones introduced just before Dreamforce. On the service side, for example, an agent can track an order, verify that a return complies with your sales policy, open a case, and then forward any issues that fall outside the scope to a specialist. On the sales side, they can qualify incoming leads and suggest an appointment time. Internally, they can handle common HR or IT questions. It all depends on the processes you assign to them.

These templates can be adapted in a matter of weeks, provided the necessary information is available. A service representative is only as good as your knowledge base. If your return policy is circulating in three different versions, the representative may end up using one of them—with no guarantee that it’s the correct one.

That leaves what is unique to your company. Depending on your business lines, this could be an assistant who helps a field sales rep create a quote, compare two products, or move an opportunity forward from their phone. No standard template covers this type of need. The value of these agents is also evident in Salesforce organizations where a large number of custom objects or processes have been implemented.

This is the type of agent we primarily develop at 2PACE, and two key lessons emerge from this.

The first issue concerns categorization. Contacts, accounts, quotes, and opportunities don’t follow the same rules, and a single agent responsible for everything tends to mix them up. For a field sales representative, we’ve therefore created a sub-agent for each business area, plus one final sub-agent that redirects off-topic questions. Their design is similar to drafting job descriptions and involves both business and technical considerations.

The second issue concerns the boundary between AI and code. When faced with two “Martin Duponts,” an agent might confidently choose one—and be wrong half the time—regardless of the model. It’s better to take that decision out of the agent’s hands: an Apex query detects the ambiguity, and the agent then requests clarification. The same logic applies to lists of values, which an Apex action verifies before each save.

An agent does not compensate for the weaknesses of the system it connects to; rather, it exposes them—and on a large scale. A Flow designed to process one record at a time will pass testing without difficulty, but risks failing the day an agent calls it repeatedly.

To gauge the value without launching a project, Coworker remains the simplest way to get started, with an Agentforce license and verified sharing rules.

For your first agent, choose a single use case—one with high volume and low risk—and clean the data within that scope only. Limit yourself to three to five actions, then test it before rolling it out to users. Do this on mobile as well: an agent doesn’t always display the same way on mobile as it does on desktop, and that’s where your field sales reps spend most of their time.

The classic mistake is to look for an employee who can do everything right from the start. We prefer to approach it as a hiring process, with a clearly defined position, a probationary period, and then gradually expanding responsibilities.

The best candidates are recurring processes that rely on reliable data and have rules that can be codified. Depending on your teams, these might include order tracking, return management, lead qualification, quote preparation, or routine HR and IT requests.

Conversely, a decision that has implications for the company is best left in the hands of an employee, with the agent then taking charge of preparing the necessary documentation.

Dreamforce took place in San Francisco from September 15 to 17, and three announcements are worth noting.

The most important one is called AIforce, an interface layer that makes your Salesforce context (data, processes, permissions) available to other tools. It is available in Salesforce as Coworker, in Slack as Slackforce, and in Claude, Anthropic’s AI assistant, as Claudeforce, which is currently in open beta with no published pricing.

Salesforce also unveiled Koa, its first reasoning model designed specifically for CRM, built in collaboration with NVIDIA. It is still in the pilot phase, with availability planned for this winter in the United States—for organizations hosted in Europe.

Finally, according to Salesforce, Coworker surpassed 100,000 users in 35 days. However, it’s important to note that the tool has been automatically activated for some customers since August, and an activated user isn’t necessarily a satisfied one.

Our take: Your teams will increasingly be working with Salesforce data without opening Salesforce itself—which doesn’t change the substance of the work, but does require special attention.

At 2PACE, we tailor these projects to your current situation rather than starting with a concept. If you’re wondering where your organization stands, let’s talk about it.