Salesforce published the second edition of its Agentic Enterprise Index this month, and the headline number is one every business leader should sit with for a second: the average organization running AI agents in production went from 5 activated agents in February 2025 to 13 by April 2026. That's not a slow creep — it's a near-tripling in about fourteen months, growing at a 7% compound monthly rate. The same report found the time it takes a business to stand up a new agent once it's provisioned has dropped 53% over the same period, and that seven in ten customer-service sessions inside the dataset are now being handled without a human touching them at all.
Read as a productivity story, that's a good headline. Read as an operations story, it's a different one: most businesses that now run 13 agents didn't sit down and design a fleet of 13. They added one agent to handle support tickets, another to qualify leads, another to draft follow-up emails, another because a vendor bundled it into a renewal — and thirteen months later, they're running a small workforce of autonomous software that nobody explicitly built as a workforce. Speed of creation went up 53%. There's no matching number in the report for how much governance, oversight, or accountability structure grew alongside it, and that gap is the actual story.
The Number Grows Faster Than the Org Chart
When a business hires its fifth employee, it usually still knows exactly what each person does, who they report to, and what happens if one of them makes a costly mistake. By employee thirteen, most businesses have built at least the beginnings of real structure — a manager, an onboarding process, some form of performance review. AI agents don't automatically get any of that. An agent provisioned in a weekend to handle one narrow task can still be running eight months later, quietly making decisions, with the person who set it up long since moved to a different project and nobody else fully sure what it's authorized to do.
- Ownership dilutes as the count climbs. Agent one has an obvious owner — whoever built it. By agent thirteen, ownership is often split across departments, vendors, and whoever happened to be in the room when it got turned on, which means nobody owns the whole fleet.
- Permissions accumulate, they rarely get revisited. Each agent typically gets provisioned with whatever access made it work on day one. Few businesses schedule a recurring review of what each agent can still touch six or twelve months later.
- Autonomous customer-facing decisions scale invisibly. If seven in ten support sessions in Salesforce's dataset are agent-handled with no human in the loop, that's seven in ten chances for a wrong answer, a bad policy interpretation, or an overpromise to go out under your business's name before anyone reviews it.
- Faster creation means faster proliferation of the same mistake. A 53% drop in creation time is good news for speed and bad news for oversight if the underlying process for vetting a new agent didn't get 53% more rigorous alongside it.
What This Means If You're the One Who Has to Answer for Agent #14
Most businesses reading a stat like "13 agents on average" will check whether they're above or below the average and move on. That's the wrong question. The right one is narrower and more uncomfortable: if someone asked you right now to list every AI agent currently running in your business, who owns each one, what data or systems each can touch, and what it's allowed to decide without a human — could you actually produce that list today, or would you have to go find out?
- 1Build the inventory before you build the fourteenth agent. A simple spreadsheet listing every active agent, its owner, its permissions, and its escalation path takes an afternoon and closes the single biggest gap most businesses have — not knowing what they're already running.
- 2Assign an owner to every agent, not just every project. A project has a start and an end; an agent that keeps running after the project wraps needs someone whose job includes checking on it periodically, the same way a piece of infrastructure needs an owner, not just a launch date.
- 3Set a review cadence tied to access, not to performance. Reviewing whether an agent is doing a good job is a different question from reviewing whether it still needs the access it was given at launch — do the second one on a calendar, independent of whether anyone's complained.
- 4Put a floor under autonomous decisions before you raise the ceiling on volume. If an agent is going to handle customer-facing decisions without a human in the loop, decide in advance which categories of decision it's never allowed to make alone, rather than discovering the boundary after it's crossed.
- 5Treat every new agent as fleet growth, not a point solution. The question isn't just "does this agent solve the problem in front of us" — it's "what does adding a fourteenth thing to our current governance model actually cost us," and that second question gets skipped almost every time.
The part that doesn't show up in the productivity numbers
A tripling in agent count with a 53% drop in creation time is a story about how easy it's become to add capability. It says nothing about how easy it's become to know what you've added. Businesses rarely get burned by the agent they built carefully and reviewed. They get burned by the one that's been running quietly for eight months that nobody remembered to check on.
A Realistic Scenario
A Wizeb client, a regional professional-services firm, came to us after realizing — almost by accident, during an unrelated software audit — that they were running six separate AI agents across three departments: one qualifying inbound leads, one drafting client email responses, one summarizing intake calls, one routing support tickets, one pulling data from a scheduling tool, and one nobody could initially explain, left over from a trial a former employee had set up and never fully turned off. None of the six had a documented owner. Two had access to the client CRM that was broader than anyone currently on staff would have knowingly granted. We didn't need to rebuild any of the six agents — most of them were doing their jobs fine. What the firm needed was the inventory and ownership structure that should have existed from agent one: a documented list, an owner assigned to each, permissions trimmed to what each agent actually needed rather than what it happened to be given, and a quarterly access review calendared going forward. The fix took under two weeks and cost far less than the incident it prevented — the orphaned trial agent still had live access to a shared inbox, which nobody had realized until the audit surfaced it.
How Wizeb Approaches This
When we design or deploy an AI agent for a client, ownership and access review aren't an afterthought bolted on once something goes wrong — they're part of the same project as the agent itself. Every agent we build ships with a documented owner, a defined permission scope, and a review point on the calendar, because the businesses that get the most value out of agentic AI over time are the ones that can answer "what do we have running, and who's watching it" without having to go find out first. If you're not sure you could produce that list for your own business today, that's exactly the kind of gap worth closing before you add agent number fourteen. Start at wizeb.com/services/ai-agents.
Three Questions Before You Add Your Next Agent
- 1Could you list every AI agent currently running in your business, its owner, and what it's allowed to access, right now, without going to check first?
- 2Is there a standing owner for the agents you already have, separate from whoever originally built them and may no longer be involved?
- 3As your agent count grows toward the 13 organizations now average, is your review process growing with it, or is speed of creation the only number going up?