What counts as production-grade AI for a Series B board?

The setup

On Monday May 11, 2026, OpenAI announced a new entity called the OpenAI Deployment Company. Four billion dollars plus committed from OpenAI and nineteen outside investors. TPG leading the group. Bain Capital, Brookfield Asset Management, and Advent International as co-leads. Capgemini, McKinsey, and Bain & Company on the systems-integrator bench. The first move on day one was the acquisition of Tomoro, a London applied-AI consulting firm with 150 forward-deployed engineers.

Read that lineup once more. The largest frontier model lab in the world just told the market that the moat is not the model. It is the deployment layer. Three days later at IBM Think 2026, IBM CFO Jim Kavanaugh said a line I have not stopped thinking about: “Great CFOs co-architect with the CEO the AI vision, strategy and business model.” That is the voice you want in your Q3 board pack. The question that follows it is the one this piece is about.

What they tried

Here is the optimistic version of a Series B AI stack as of May 2026. A model layer from OpenAI or Anthropic, sometimes both for failover. An orchestration tool, often LangGraph or the newly shipped Claude Managed Agents. A retrieval layer over the company data lake. A vector index. Three to five named agents running named workflows: revenue ops triage, customer success summarization, post-sales onboarding handoff, finance reconciliation, maybe contract review.

The headcount-to-agent ratio looks lean. The dashboards look healthy. Adoption charts trend up and to the right. The board deck calls this production.

The press around it in the past week reinforces that framing. SiliconANGLE’s two-day coverage of the AI Agent Conference profiled production-stage operators: Thomson Reuters, Appen, Bright Data, Catena Labs, OutSystems, WalkMe, AG2ai, Monte Carlo Data. At Red Hat Summit on May 12, Google Cloud Marketplace MD Dai Vu announced 2,000-plus agents from ServiceNow, Oracle, and Red Hat live on the Gemini Enterprise Marketplace, with $750M committed to deployment-partner enablement. Google Cloud’s annual run rate just crossed $80B with a $462B committed backlog. Anthropic shipped Claude Managed Agents the same week. Sean Neville’s Catena Labs is building what the company calls know-your-agent, an identity and audit framework for financial agent transactions.

The picture looks like enterprise procurement has caught up with founder ambition. So far, so good. The interesting part is what the same week’s sources said about what is actually running underneath that picture.

Where it broke

HPCwire ran a piece on May 13 with a title that does not need decoding: Why Enterprise AI Keeps Failing, and It’s Not the Model’s Fault. The data inside is the uncomfortable kind.

The piece names the mechanic. Agentic pipelines rarely fail because one component breaks. They fail because the sequence of interactions, retrieval into inference into tool use into downstream action, starts to diverge under real-world load. The system degrades behaviorally before it degrades operationally. The first signal is not an incident ticket. It is user mistrust.

Stack the next number next to the 25% reliability-confidence figure and the math gets harder: 88% of organizations reported a confirmed or suspected AI agent security incident in the past year. So most companies have agents in production, almost all of them have had an incident, and three out of four cannot confidently say whether the system is reliable.

Barr Moses, CEO of Monte Carlo Data, named the pattern at the AI Agent Conference in plain English: “The gap between promising proof of concept and trustworthy production deployment is where most initiatives break down.” Tai Carmi, WalkMe’s CIO, added the human side from the same stage. 80% of executives believe they are providing excellent AI tools to their workforce. Only a fraction of employees agree. That is the silent failure outside the system itself. CloudBees ran a separate survey of 200-plus enterprise tech leaders for its State of Code Abundance 2026 report and named the new line item the rest of us are quietly burning: token anxiety. The pipelines work. The bills are surprising. The visibility is thin.

The pattern

Seven validated sources from one week, three angles. Capital is saying the deployment layer is the moat. Operators are saying production is where the breakage lives. Researchers are naming the failure pattern as silent and behavioral. They converge on the same thesis. Production-grade is not a stage gate the engineering team passes. It is an operating discipline the company has to learn.

IBM CFO Jim Kavanaugh laid out the six imperatives at IBM Think 2026: auditability, cost controls, trusted data governance, operating discipline across fragmented environments, observability for agentic systems, and human accountability mechanisms. For a Series B, that list is one page too long for a board read. So compress it.

A Series B CFO can run that audit in one afternoon. A Series B board can read the result in five minutes. The number of agents that clear all five lines is the real production count. Everything else is a pilot wearing production clothing.

Production is not a stage gate engineering passes. It is an operating discipline the company learns.

What I’d tell you over coffee

If I were sitting with your Series B CFO this week and the AI deck landed on the table, I would not ask how many agents are live. I would ask how many would clear the five-line audit. My honest guess, gently offered, is that most decks underline two and quietly delete three. That is normal. The companies figuring this out before Q3 are the ones taking Kavanaugh’s line literally. Co-architect with the CEO. Treat AI like a P&L line, not a science project. Trust the operating layer, not the demo.

The deployment layer is the moat now. OpenAI just bet four billion dollars on that idea. The audit fits on one page. You can run it Monday.