How many AI agents has your Series B actually shipped?

The setup

I read the Seoul Economic Daily piece on the Dataiku CIO survey this morning and had to put my coffee down. The number that did it was buried three paragraphs in. One enterprise assumed employees were spinning up about 500 AI agents per week. When the company ran an Agent Management review against its production environment, the real count was 2,000. Four times what the executives thought.

That is not an AI hype story. That is the new operating reality for any company past Series A.

The Dataiku global survey published this week ran across 600 chief information officers. Ninety-five percent of them already report AI performance to their boards. Only one in four can actually monitor every agent their teams have shipped. If a board asks for an AI scorecard, most CIOs can produce a deck. Very few can produce a real count.

For Series B founders right now, that gap is the operating-model problem of the quarter.

What they tried

The companies in the Dataiku survey are not negligent. Most of them did the recommended things. They wrote an AI policy. They named a head of AI. They signed off on a vendor stack. They built an AI committee that meets every other Tuesday and produces presentable slides.

That worked when there were ten agents. It does not work at two thousand.

What changed in the past two weeks is the pace of new agent creation, on two fronts at once.

The first front is the platform layer. On May 22, SiliconANGLE covered IBM Think 2026 and the line that landed for me was a CIO at the event saying “the proliferation of agents is real in enterprises. Everybody is starting to build.” That same day at Dell Technologies World, Bud Ecosystem rolled out an Enterprise AI Operating System aimed squarely at tool sprawl. Their founder told theCUBE that the enterprises he is working with have “50 different tools, noise and errors and all of that.” That is not vendor talk. That is what governance fragmentation looks like when every team self-serves.

The second front is the consumer-grade surface area. On May 21, Marc Benioff shipped Agentforce Coworker into every Salesforce search bar, available immediately to all Enterprise, Unlimited, and Agentforce 1 customers. A working Salesforce admin posted on launch day that the agent did in seconds what used to take 45 to 60 minutes of swivel chairing between screens. That is genuinely good. It also means every salesperson in the company is now creating agent-driven workflows whether the VP of IT signed off or not.

Real production deployments are landing in the same window. The MarketingProfs roundup on May 22 covered Hershey running Mutinex (Claude plus Gemini) and Tracer across more than $2 billion in annual marketing spend, expecting a four to five percent revenue lift. Graebel is using Microsoft Copilot Studio’s computer-use agents to handle relocation services through actual UI navigation. Bayer is running spend caps and data anonymization on programmatic media buying with agent-driven guardrails. These are not pilots. They are live and on someone’s P&L.

Where it broke

The break is not technical. Models work. Vendor platforms work. The break is that the org chart never caught up with the number.

Versa’s senior director of AI and machine learning Sridhar Iyer said it cleanly at the Verbo launch on May 21: “Enterprise AI is at an inflection point. AI in production can turn into a liability, not an advantage.” Versa’s whole pitch is that every agent action gets validated against identity and policy before it executes. The reason that pitch is landing this quarter is that thousands of agents acting on enterprise data without a control plane is no longer a productivity story.

Chris Farris published a piece on May 23 that updated the 1979 IBM principle for the agentic era. The original was: a computer can never be held accountable, therefore a computer must never make a management decision. His updated version, paraphrased: an agent must never be the final decision-maker for a consequential decision, because accountability still has to land somewhere with a human and a job title. When it does not land cleanly, it gets displaced upward, usually to whoever can least afford it. That is the executive who signed the platform purchase order and now owns an incident she never saw coming.

For Series B founders specifically, the place this breaks first is enterprise renewals. The procurement team at the largest customer account is now asking for an agent register, an incident response plan, and a tested kill switch. Without those three artifacts, the renewal stalls.

The pattern

The pattern across the data this week is consistent enough to name. The first wave of enterprise AI was a model selection problem. The second wave was a use case problem. The third wave, which we are in right now, is a count problem.

Production-grade AI for a Series B company in 2026 means owning the agent count, the kill switch, and the accountability chain, not picking the right model.

IBM launched Sovereign Core at Think 2026 this week. Bud Ecosystem launched its Enterprise AI Operating System at Dell Tech World. Both are betting that the next dollar of enterprise AI spend goes to unification, not new agent capability. That is also why the smartest Series B founders I have talked to this month are pulling one person off their FinOps or developer-productivity team and giving them a new title: agent owner.

The agent owner does three things. They keep a real-time register of every agent in production. They own the kill protocol. They sit in the room when the board asks about AI risk. That is not a new department. It is one person with a one-page document, a Slack channel with the authority to switch things off, and a standing slot on the Q3 board pack.

The companies further along this curve are already running the count alongside their AI policy. The Dataiku survey notes that 95 percent of CIOs report AI performance to their boards, and one of the most useful Q3 board metrics is the simplest one in the report: how many agents shipped this quarter, how many were retired, and how many fired at least once without human review. That is a credible operating metric. “We have a policy” is not.

What I’d tell you over coffee

If we were sitting at the coffee shop near the office and the Q3 board meeting was three weeks away, here is what I would ask before letting anyone order a second cup.

First, what is the actual number of AI agents your team has shipped to production this quarter? Not the number on the deck. The real number. If the answer is “I would have to check,” that is the project for next week.

Second, who owns the failure mode? Pick one name. If the answer is “the AI committee” or “we are working on it,” there is no ownership yet, and the procurement conversation at the largest enterprise account will land harder than expected.

Third, what is the production-grade reliability bar the top three enterprise customers are now asking about in renewal conversations? That bar moved this quarter. Better to find out from a friendly account call than from a procurement questionnaire that arrives the week before close.

The reassuring part is that none of this requires a new platform or a six-month strategy review. The companies getting this right this week are the ones doing the boring count first. The four-times gap between perception and reality is figure-out-able. It just has to be somebody’s actual job.