There’s a particular kind of exhaustion that sets in when a software vendor promises to change everything. Salesforce has made that promise before, with the cloud, with mobile, with Einstein AI. When Marc Benioff took the stage at Dreamforce 2024 and declared that Agentforce would “empower one billion agents,” most enterprise tech leaders had heard something like it before. And yet, something is different this time. The numbers are hard to ignore, the frustrations are real, and the companies getting it right are doing so in ways that don’t fit the marketing brochure.
The numbers that matter (and the ones that don’t)
Agentforce reached an ARR of $500 million on its own, and combined with Data 360, hit $1.4 billion in annual recurring revenue, growing 114% year-over-year. Salesforce has closed 18,500 Agentforce deals in total, with over 9,500 of those paid, and has processed 3.2 trillion tokens.
Those are impressive headline figures, but the more instructive metric is the gap between total deals and paid deals. For much of 2025, that gap was wide. Salesforce’s Q4 2025 results showed adoption was fairly slow out of the gate, 5,000 deals secured, with just 3,000 of those paid. The pattern told a story that anyone who has been through enterprise software cycles knows well: lots of organizations kicking the tires, fewer actually writing checks.
That said, momentum did build. Active customer accounts grew 70% quarter-over-quarter, with many enterprises moving beyond pilot testing into live deployment. The question was never really whether Agentforce would find a market. It was whether the adoption curve would be smooth enough for enterprises to trust it with real, revenue-critical workflows.
Where is it actually working?
The clearest wins have come in customer service, not because the technology is most powerful there, but because the feedback loop is fastest. The average volume of customer service conversations led by an agent grew 22 times in the first half of 2025, and agent creation among early-adopter companies surged 119% between January and June.
The companies seeing real returns share a few traits. They started narrow, they had clean data, and they resisted the urge to automate everything at once.
Travel startup Engine built, developed, and launched its first agent in just 12 days, then focused on continuous iteration, tailoring actions and building on early learnings, which was critical in achieving a 15% reduction in average handle time.
Grupo Globo, the Latin American media giant, integrated Agentforce into customer interaction channels and improved retention rates by 22% in under three months, proactively addressing viewer concerns before they became churn risks.
Adecco, the global staffing firm, implemented Agentforce to automate interview scheduling, candidate follow-ups, and document collection, giving recruiters back valuable hours each week to focus on relationship-building.
At tax services company 1-800Accountant, Agentforce now resolves up to 60% of incoming requests, including routine inquiries like tax return statuses, freeing the team for more complex cases.
What these examples have in common: a well-defined problem, an existing Salesforce data foundation, and a willingness to let the technology be imperfect at the start. None of them tried to automate a sprawling, cross-departmental process on day one.
The pitfalls nobody talks about at the keynote
The honest version of the Agentforce story includes a significant amount of friction that Salesforce’s own marketing tends to gloss over.
Pricing confusion was a genuine adoption killer. When Agentforce first launched, Salesforce priced it at $2 per conversation. In reality, conversations ran long, branched into side threads, and any conversation would count, even if it wasn’t meaningful. Costs became unpredictable and hard to budget for. This sent finance teams and procurement officers back to the drawing board just when IT teams were getting excited.
Salesforce responded in May 2025 with a new flexible pricing model called Flex Credits, charging by action rather than conversation. Customers now buy bundles, 100,000 credits for $500, with each action costing 20 credits, roughly $0.10 per action. A Digital Wallet dashboard was also launched to track usage and forecast spend. The repricing helped, but the damage to early enthusiasm was real.
Messy data is the invisible wall. No AI agent performs well if the underlying CRM data is incomplete, inconsistently structured, or fragmented across system boundaries. This is an older problem than Agentforce itself, but the platform brings it into sharp relief. Organizations that had invested in Data Cloud and data hygiene over the prior two to three years were positioned to move fast. Everyone else discovered, sometimes painfully, that deploying an agent is easy; getting it to produce reliable, accurate outputs requires a data foundation that many enterprises simply don’t have yet.
Early documentation and tooling were thin. In early 2025, documentation was still sparse — the Agentblazer trails were thinner and best practices were still being worked out. By late in the year, everything had advanced significantly. That’s a normal maturation curve for any new platform, but it meant early adopters were navigating without a proper map, which drove up implementation costs and extended timelines.
Governance concerns haven’t gone away. Despite early wins, enterprise leaders are still grappling with governance and data readiness before fully committing to agentic AI.The escalation data is telling: escalations from AI agents back to human agents actually increased from 22% in Q1 2025 to 32% in Q2 2025. This isn’t necessarily failure — it may reflect better calibration of when agents should hand off, but it signals that many organizations are still working out where the human-machine boundary should sit.
The speed advantage nobody expected
One data point consistently surprises people who are used to enterprise software deployment timelines. According to a 2025 Valoir study, organizations using Agentforce took an average of 4.8 months from strategy to full deployment, compared to 75.5 months for companies building a custom agentic AI stack from scratch.
That number deserves some scrutiny, “full deployment” can mean different things, and a narrowly scoped first agent is easier to ship than a complex multi-channel implementation. But the directional reality holds. The fact that Agentforce sits inside an ecosystem that enterprises already rely on daily removes entire categories of integration work, data migration, and user adoption friction that would exist with a standalone AI tool.
A Gartner analyst noted that Salesforce may have an adoption advantage precisely because Agentforce sits inside workflows customers already rely on daily, unlike standalone AI tools.
What does the next phase look like?
Agentforce 3, introduced in June 2025, added a Command Center for real-time visibility into agent activity, Model Context Protocol support for cleaner integrations with third-party tools, global scale improvements including lower latency and model failover, and new per-user pricing for employee-facing agents. These weren’t flashy announcements — they were the kind of operational improvements that mature a platform from “interesting pilot” to “production-ready infrastructure.”
Travel and hospitality saw AI and agent actions grow at a monthly average rate of 133% in the first half of 2025, the fastest of any industry tracked in Salesforce’s Agentic Enterprise Index. Salesforce Retail was close behind at 128%. The industries seeing the fastest gains tend to be those with high transaction volume, relatively standardized customer inquiries, and existing Salesforce penetration.
The harder question is what happens in industries where the use cases are more complex: healthcare, financial services, legal, government. Salesforce has begun showcasing adoption stories in education and government, and has launched industry-specific solutions in healthcare, finance, and public sector. But these sectors move slowly for good reasons, and the proof points are still thin.
A realistic take
The enterprises getting the most from Agentforce right now are not the ones that boil the ocean. They pick one process where data is clean, the problem is well-defined, and there’s a clear metric for success. They ship, they learn, they expand. They treat the first agent like a team member in a probationary period, useful but watched closely.
The enterprises struggling are those that built a business case on projected AI savings before understanding what it would actually take to make the agents reliable. They over-automated too early, hit governance walls, or discovered that their Salesforce data was in worse shape than anyone had admitted.
The real lesson from 2025’s Agentforce story is one of steady, measurable progress rather than overnight transformation. The roadblocks haven’t vanished, they’ve evolved, and so has the ecosystem’s understanding of what adoption actually means. Salesforce Ben
Benioff’s billion-agent vision may yet materialize. But the companies laying the actual groundwork are doing it one well-scoped agent at a time, and that, it turns out, is exactly the right approach.
👉 Want to learn how to work with Agentforce in practice? Join our Modern Salesforce Administration course and get hands-on with Agentforce through real-world exercises.
