It’s safe to run your AI agents (they’re mostly chatbots, case summaries, or simple text generators today) on the Agentforce chassis — as long as you run them inside your Salesforce application domain. We spent two days in San Francisco at Salesforce’s TDX developer conference. Together with 5,000 Salesforce developers and administrators (Trailblazers, rebranding as Agentblazers), we touched software, attended classes, and spoke with executives, including President and CMO, Ariel Kelman, EVP of AI engineering, Jayesh Govindarajan, and senior vice president, strategic partnerships & business development, Nick Johnston.

We come away impressed with the CRM giant’s commitment to agent-powered workflows and cautiously optimistic that the no-software company can host many, if not all, the AI agents running in the Salesforce ecosystem. We liked the empowerment angle. We don’t love the ham-handed labeling of Agentforce as “the digital labor platform” because what we saw were agents doing mundane work that empower people, not automate away jobs. In subsequent sessions with chief AI officers in retail, banking, and hospitality, we learned that they too believe generative AI (genAI) is a power tool, not a digital worker.

Here’s what CIOs and other technology executives need to know:

  • Salesforce is massively interested in hosting your AI agents. If you’re a Salesforce shop, we think you should try it out and see how the platform works for you. Salesforce is using a monthly release cycle to rapidly improve the product. Features like choose-your-own language model, PII data masking, prompt templates, zero-copy data, vector embeddings, agent benchmarking tools to build a business case, and an agent lifecycle toolkit are available today in the Agentforce platform.
  • Most agents are repaved task paths, not automated workflows. Boston believes its crazy roads are paved over cow paths. It turns out that paved over cow paths are better than paths knee-deep in mud. It turns out agents today repave existing manual processes. That’s OK. There are dozens of scenarios in the Salesforce ecosystem where an AI agent can empower an employee, do the grunt work, and maybe give some advice in a text response. Of course, if that advice is through a customer chatbot, then fewer calls may flow into the contact center, thus reducing the number of reps needed. But does that make the agent digital labor? Nah. It’s just an application to help customers serve themselves. Don’t let the $2/call pricing model confuse you — that’s just a value-oriented negotiating stance on the cost of the “equipment.”
  • If you build on Agentforce, you’re committing to Data Cloud. This is the biggest strategic play we see Salesforce making — and your biggest risk of agent and knowledge asset lock-in. Salesforce, along with ServiceNow, Microsoft, Oracle, SAP, Workday, Deloitte (shockingly), and others, want your proprietary knowledge assets as well as your AI agents. Salesforce could already have your front-office data or could get it through zero-copy retrieval from an Amazon, Databricks, Google, Microsoft, or Snowflake database. But genAI and knowledge graphs have a symbiotic relationship. That means Agentforce also needs your proprietary sales manuals, product literature, marketing materials, process methods, and more to generate. That makes Salesforce Data Cloud a vital component of AI agents, hence a strategic and expensive commitment for you to make.

What Technology Executives Should Do

If you use Salesforce, then ask a small team to investigate the boundaries of common sense for building and operating AI agents on Agentforce. For ideas, check out Salesforce’s AI library, or try out one of the agent templates for sales coaching or case summaries, for example. Test these out:

  • Build a prompt template for common retrieval patterns. For example, if your team is constantly asking for the same data, give them a chat interface but prepopulated with context and prompt suggestions. That’s a prompt template.
  • Build a simple agent to do something not yet in the product and make it available with a button. More and better summary tools; personalized emails; or simulated coaching for the next sales call are good candidates.
  • Load documents into Data Cloud so they’re available to an agent through vector embeddings. If you load all your sales training material, for example, this could power your sales coach agent. One executive at a healthcare provider network we spoke with is using an agent like this so that a clinician facing a tough patient conversation can get some coaching in the context of the diagnosis.

If you want to dig deeper into the CIO’s role in AI agent success, please reach out to me by scheduling a guidance session or an inquiry via email: inquiries@forrester.com.