Companies now run 13 AI agents each, up from 5

The number of AI agents a company runs has stopped being a signal of sophistication. Salesforce's second Agentic Enterprise Index puts the average activated agent count at 13 per organization as of April 2026, up from 5 in February 2025. The more interesting number is buried under it: the average time to create an agent is now 1.9 days, down 53% over the same period. When building an agent takes two days, thirteen agents is not a strategy. It is what happens when nobody stops you.

What actually happened

Salesforce published the second edition of its index in August 2026, combining aggregated Agentforce platform telemetry across thousands of businesses with a survey of 4,689 respondents fielded in May 2026. The headline findings: agent production grew nearly 3x at a 7% compound monthly growth rate; average actions per account grew at a 31% compound monthly rate over 15 months; and the number of distinct skills each agent can act on rose from an average of 2 at the start of 2025 to 6.

Salesforce also reports a split in how agents get deployed. Consumer-facing sectors run high-volume, task-specific agents. Regulated and operationally complex fields — manufacturing, public sector — run fewer agents that handle longer multistep work.

One caveat worth stating plainly: this is a vendor measuring its own platform. The telemetry describes companies that bought Agentforce and turned agents on, not the market at large. The direction is credible. Treat the absolute numbers as Agentforce numbers.

Why agent sprawl matters for your business

Thirteen agents is thirteen sets of credentials. Every agent you activate holds an identity, a permission scope, and a set of tool connections. At two days per agent, permission design is the step that gets skipped. The question to ask your team is not how many agents are live — it is which of them could read a customer record they have no business reading, and who would notice.

Skills per agent is the number that predicts value. Salesforce's 2-to-6 jump in skills per agent is the real story in that dataset. An agent with one skill is a script with a chat interface. An agent with six is doing the thing you actually wanted: holding context across a task that used to require a person to switch tools. If your agent count is climbing and your skills-per-agent is flat, you are building a directory of scripts.

Fast to build is also fast to abandon. A 1.9-day build time means the marginal cost of an experiment is nearly zero, and so is the marginal cost of leaving one running after the person who built it moved on. Put an owner and a review date on every agent at creation. Nothing else scales.

The bifurcation is a hint about your own shape. If your work is high-volume and repetitive — order status, returns, appointment changes — many narrow agents is a reasonable architecture. If your work is multistep and consequential — quoting, compliance, anything that touches money — you want fewer agents with deeper skills and a real audit trail. Copying the wrong pattern from the wrong industry is how automation projects stall.

Sources: Salesforce Agentic Enterprise Index, Second Edition, Salesforce News: Agentic Enterprise Index insights.