What the data shows
Salesforce released usage data drawn from its Agentforce customer base alongside a survey of roughly 5,000 respondents across seven countries, covering the period from February 2025 through April 2026. The headline figure is that the number of active AI agents nearly tripled over those 15 months. That is a usage curve, not a marketing claim, built from actual account activity rather than stated intentions, which makes it a more reliable signal than most adoption surveys published this year.
The speed metrics are the more consequential number for IT operations. Time to deploy a new agent, from creation to activation, fell 53 percent and now sits under two days for most customers. Actions completed per account grew at a compound monthly rate of 31 percent, and total agentic work output grew at a 15 percent compound monthly rate. Put together, the picture is an adoption curve accelerating faster than most enterprise software categories have moved historically, including the early cloud and mobile transitions IT leaders remember managing.
Where the growth is concentrated
Retail stands out sharply in the data. Retail-sector agents generated 22 percent of total agentic work output across the entire dataset and grew 18-fold during the analysis window, a disproportionate share for a single vertical. Caila Schwartz, Salesforce's director of consumer strategy, described the pattern directly: agents can do more than most teams expect, and they are increasingly taking on secondary functions that used to require dedicated staff time, from order status handling to inventory reconciliation tasks that sit adjacent to a core workflow.
That concentration matters for anyone running commerce or operations at scale. The functions agents are absorbing first are the high-volume, rules-based secondary tasks that never individually justified a dedicated automation project but collectively consumed significant operational headcount across a retail organization, from order status handling to inventory reconciliation to return processing. That is a useful reframe for where to look first when scoping an agent pilot: the highest-volume tedious workflow available, rather than the most technically ambitious one your executive sponsor wants to showcase.
The gap between speed and proven value
The data includes an important counterweight to the adoption curve. Fewer than half of organizations surveyed by Aptean, cited alongside the Salesforce findings, considered AI essential to their core operations. That is a real gap between how fast agents are being deployed and how confidently leadership can point to them as load-bearing infrastructure. Shree Reddy, CIO at PenFed, framed the goal as using AI in a trusted, practical, and meaningful way, which is a modest ambition compared to the deployment velocity the broader data shows.
That gap describes the honest state of enterprise agent adoption right now, where velocity has outrun conviction. Organizations are standing up agents quickly because the tooling makes it easy and the incremental cost of trying is low, often ahead of validating each deployment against a clear business case. That description points directly at where the discipline gap sits, and it is exactly the gap a CIO's governance process needs to close before agent sprawl accumulates into agent debt that takes a dedicated cleanup project to unwind.
Why deployment speed breaks quarterly governance
Most enterprise AI governance frameworks were designed around a review cadence that assumed weeks or months between a proposed deployment and its approval. A two-day deployment cycle makes that cadence structurally incompatible with how fast agents are actually being created inside business units. By the time a quarterly governance committee reviews an agent, dozens more may already be live, built by teams who found the barrier to entry low enough to skip the formal process entirely.
The redesign that actually works matches governance to the deployment speed the tooling now permits. That means pre-approved agent templates with baked-in guardrails, automated logging that flags anomalous agent behavior without requiring a human to review every deployment, and clear escalation triggers rather than blanket pre-approval requirements for every new build. Governance that cannot operate at two-day speed gets bypassed in practice, and the resulting shadow agent sprawl is harder to unwind later than the review friction it was originally meant to prevent.
The roadmap decision
The practical takeaway for an operations or engineering leader is to treat the retail-style secondary-function pattern as the template for where to start, rather than the ambitious flagship use case most executive sponsors want to lead with. High-volume, rules-based, low-risk tasks are where the data shows agents delivering measurable output growth fastest, and they are also the lowest-risk place to build institutional confidence in agent reliability before extending scope into customer-facing or judgment-heavy workflows where a mistake is far costlier to unwind and far more visible to the business when it happens.
The second decision is structural: build the governance model now, while agent counts are still in the hundreds rather than thousands per organization. Waiting until the Aptean confidence gap closes on its own, once more organizations decide AI is essential to core operations, means governance gets built retroactively on top of sprawl that already exists and is far more expensive to inventory and correct after the fact than to design alongside adoption from the start. The organizations that get this right will treat governance as infrastructure built in parallel with rollout, sized for two-day deployment cycles rather than quarterly review boards, and staffed accordingly before agent counts multiply further.
