Agentforce ROI: The Numbers Salesforce Disclosed After Dreamforce 2026

On September 24, 2026, Salesforce published a post-Dreamforce recap on its official site (salesforce.com/news) containing customer deployment numbers rarely disclosed with this level of detail. Beyond recapping already-known announcements (AIforce, Koa, Claudeforce — see our dedicated articles), it's these quantified customer cases that add genuinely new, usable information for evaluating an Agentforce investment. Here is exactly what the official source states, along with the caveats that apply.

Fulton Bank: From First Contact to 3,000 Users in Six Weeks

According to Salesforce, it took Fulton Bank one hour to go from hearing about Agentforce to setting up its first agent — Hunter, a prebuilt outbound sales agent, deployed to re-engage an at-risk pipeline targeted at $340,000. Within a week, the agent was rolled out to power users; six weeks later, 3,000 employees were using it. This timeline covers the technical setup of a standard prebuilt agent, not full integration with custom business processes — worth keeping in mind before extrapolating this timeline to your own context.

Southwest Airlines: A 7x ROI Claim, on an Undisclosed Scope

Southwest Airlines started with a focused AI assistant answering customer questions, freeing human reps for high-empathy interactions, then expanded into more complex operations (name changes, receipt requests, travel vouchers) orchestrated via subagents within a unified experience. Salesforce states "millions" in productivity gains and a 7x ROI "in a relatively short time frame" — without disclosing the amount invested, the exact duration, or the calculation method behind that ratio. This figure should be read as a case study the vendor chose to publish, not as an audited benchmark.

Koa: Performance Gains Measured Against Salesforce's Own Internal Benchmarks

Salesforce states that Koa, its reasoning model for Agentforce, is 14% better than the company's own internal benchmarks at retaining context in long, multi-turn conversations, delivers 15% more relevant answers, and executes the correct tool calls with 11% more precision. These percentages are measured against Salesforce's own internal references, not an independent third-party benchmark — an important nuance for anyone trying to compare Koa against another model on the same criteria.

What Jason Lemkin (SaaStr) Actually Admits, Beyond the Quote That Circulates

SaaStr CEO Jason Lemkin said at a Dreamforce session that he runs "a little eight-figure business with three humans and 21 agents on Salesforce for over a year." The full quote, often cut short in media coverage, continues: "Agents are great, right? But you can't trust them. You need a harness." In response, Salesforce presents its "Enterprise AI Harness," structured around six capabilities: trusted models, trusted context, trusted agency, trusted actions, trusted governance, and trusted security — which is, implicitly, an admission that deploying autonomous agents without guardrails remains risky, even in the vendor's own framing.

Before You Calculate Your Own ROI

• These numbers are customer cases disclosed by Salesforce itself, not audited by an independent third party — treat them as illustrations, not guarantees transferable to your organization.
• The very short setup time (Fulton Bank: 1 hour) covers a prebuilt agent on a standard use case, not integration with your existing business processes or your own data.
• Southwest Airlines' 7x ratio comes with no disclosed investment amount or published calculation method — it cannot serve as a quantified comparison baseline.
• Salesforce's own governance narrative (the "harness") confirms that deploying agents without a control framework (access rights, authorized actions, human oversight) carries a risk the vendor itself acknowledges.

Rather than extrapolating a ROI ratio achieved on a scope you don't know, our consulting offering audits your existing processes and quantifies an Agentforce deployment scenario based on your own data and use cases.