--- title: "AI Agent Governance Is the New Job. Whose Is It?" description: "Vercel's CEO calls agent governance \"literally our new job.\" Schellman: 74% of practitioners say they're AI audit-ready, 27% are. Here's whose job it is." image: "https://storage.googleapis.com/promptmetrics-uploads/website/posts/1786721274624-922605681.jpeg" author: "Izzy" category: "Governance & the Human Gate" publishedAt: "2026-08-14T15:27:55.126Z" updatedAt: "2026-08-14T15:27:55.130Z" canonical: "https://www.promptmetrics.dev/blog/governance-is-the-new-job" --- # AI Agent Governance Is the New Job. Whose Is It? "You have to be really really really thoughtful about it. That's literally our new job." That's Guillermo Rauch, CEO of Vercel, on the Agent Native podcast in August 2026. He's describing what it takes to run a company agent that touches customer data. Not the model. Not the prompts. AI agent governance: deciding what the agent may read, what it may write, and who says yes before anything ships. Now the part that should bother you. Rauch runs a company where almost 1,000 people use one internal agent, and he says governance is the new job. At most companies, nobody holds that job. The AI mandate landed. The tools landed. The oversight work got scoped by exactly no one: no owner, no budget line, no job description. So I wrote the job description. This post covers what AI agent governance actually consists of, who should hold it, and why it isn't a coding job. > **The short version:** Vercel's CEO says the real work of company agents is governance: data access, review gates, oversight. Most companies haven't staffed it. Schellman surveyed 525 US enterprise practitioners in July 2026. 74% think their organization is audit-ready for AI. 27% run mature governance. The job exists. Almost nobody owns it. ## What did Vercel's CEO actually say about agent governance? He said the job is deciding what the agent can touch. And that it belongs to whoever builds it. Here's the full quote, from around the 20:43 mark: > "Someone that's building an agent can very carefully define governance, data access control in the security model (...) because agents are interacting with customer data. So you can't just be like, I don't know man, it runs a computer and it has access to (...) all of the databases of everything. You have to be really really really thoughtful about it. That's literally our new job." ( [Agent Native](https://podcasts.apple.com/us/podcast/they-built-an-al-god-agent-for-1-000-employees/id1826127218?i=1000780278017) , Aug 2026, ~20:43) He's not theorizing. Vercel's internal agent, V, serves almost 1,000 people. Rauch says "hundreds of millions of dollars of revenue are dependent on the well-being of this agent" (~41:12). Sales reps depend on it. Support depends on it. He depends on it. And earlier in the conversation (~7:37): "the new job of someone that runs a company is to actually enable their workforce with agents and to work on the agent." Read those two quotes together. A CEO with real revenue riding on one governed system says the new work is enablement plus oversight. The hiring is already happening. IBM's Institute for Business Value surveyed 2,000 CEOs across 33 countries between February and April 2025. 54% were hiring for AI roles that didn't exist a year earlier ( [IBM IBV](https://newsroom.ibm.com/2025-05-06-ibm-study-ceos-double-down-on-ai-while-navigating-enterprise-hurdles) , 2025). Rauch just named the most neglected one: the person who governs the company agent. The job is being invented in real time, at companies that can afford to invent it. Rauch also talked about whether you need one agent or a team of many. Different question, different post: I cover it in [one god agent or a team of agents](/blog/god-agent-vs-team-of-agents) . This one stays on the job. ## Why does the AI agent governance job go unstaffed? In July 2026, Schellman surveyed 525 US professionals who evaluate, deploy, secure, or govern AI. All of them work at companies with 500+ employees and $100 million or more in revenue. 74% believe their organization could pass an AI compliance audit today. Only 27% describe their governance program as fully mature ( [Schellman](https://www.globenewswire.com/news-release/2026/07/29/3335281/0/en/New-Schellman-Research-74-of-Enterprises-Say-They-Are-Audit-Ready-for-AI-Only-27-Actually-Are.html) , 2026). That's a 47-point gap. It's the unstaffed job, measured. The gap exists because everyone assumes someone else holds the role. The same Schellman survey found only 37% of respondents identify the CIO or Head of IT as accountable for AI-related risk, and only 44% maintain AI-specific incident response procedures. The mandate carrier got told "you're smart, figure it out." Nobody attached an oversight budget to the sentence. Quick test: who at your company answers for it when an agent writes something wrong into a customer record? If the answer takes longer than five seconds, you just found the gap. ![](https://storage.googleapis.com/promptmetrics-uploads/website/content-images/1786720459654-521697259.png) The cost of leaving the job empty is already priced. In June 2025, Gartner predicted that over 40% of agentic AI projects will be canceled by the end of 2027, and named inadequate risk controls as a driver alongside escalating costs and unclear business value ( [Gartner](https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027) , 2025). Not model quality. Risk controls. The projects die from the missing job, which is also [why the failure is rarely the tool](/blog/it-s-not-the-tools-it-s-the-shape) . One more reason nobody staffs it: the market sells governance as everything except a job. Vendors sell it as software you configure. Consultancies sell it as a framework PDF. Rauch described a job, with tasks and an owner. Companies believe they're governed because audits haven't disagreed yet; Schellman's 74%-versus-27% split shows what happens when someone finally checks ( [Schellman](https://www.globenewswire.com/news-release/2026/07/29/3335281/0/en/New-Schellman-Research-74-of-Enterprises-Say-They-Are-Audit-Ready-for-AI-Only-27-Actually-Are.html) , 2026). ## What does the governance job actually consist of? In operator terms, the job is four decisions: 1. **Read access.** Which systems and fields the agent may see. Your RevOps manager decides the agent reads Acme Corp's deal stage and last activity date, and never reads payroll. 2. **Write access.** Which fields it may change, and which it may only draft against. 3. **The gate.** Which actions need a human yes before they ship. Write actions to customer records: all of them. 4. **The record.** What the yes leaves behind. Who approved what, when, and what exactly changed. ![](https://storage.googleapis.com/promptmetrics-uploads/website/content-images/1786720530372-42552064.png) ### Who should be making those four decisions Notice what's missing from the four decisions: code. Rauch wants people who "really understand data flows, threat models and architecture of systems design." Then he adds: "I don't think that the code-writing part is as important" (~41:56). He says it straight, and I'll say it too. The person who knows the workflow is more qualified for this job than the person who knows the code. Vercel splits it the same way. "Everyone can contribute to the agent: information, skills, critique, feedback," while engineers hold "the core system loop, the access to data, the governance, security" (~41:30). The decisions sit with the operator. The plumbing sits with the engineer. ### Why a human approval isn't automatically governance There's a trap inside decision three. The stakes aren't abstract. In early 2024, an employee at UK engineering firm Arup approved 15 transfers worth US$25.6 million. Every other person on that video call was a deepfake ( [Gulf News](https://gulfnews.com/world/asia/deepfake-scam-fake-video-meeting-with-cloned-cfo-costs-firm-256-million-1.1707219298784) , 2024; [CNN](https://www.cnn.com/2024/05/16/tech/arup-deepfake-scam-loss-hong-kong-intl-hnk) , 2024). An approval happened. The gate still failed, because the reviewer approved something they couldn't inspect. Every fake face looked real. Governance isn't "a human said yes." Governance is a human saying yes to the real change, with enough context to catch the wrong one. IBM's Phaedra Boinodiris and Jamie Mackenzie named the failure mode in June 2026: liability laundering. When a reviewer rubber-stamps output they can't inspect, "the accountability that should flow toward the design and deployment of the system is quietly redirected to the person who clicked 'approve'" ( [IBM](https://www.ibm.com/think/insights/liability-laundering-problem-human-in-the-loop-not-governance-strategy) , 2026). A gate that shows a vague summary isn't governance. It's blame routing. The reviewer has to see the actual change, field by field. I've written up [what a human gate looks like inside a production CRM](/blog/the-gate-belongs-inside-the-tool-giving-an-ai-agent-write-access-to-a-production-crm) if you want the mechanics. Field note from rebuilding PromptMetrics' own stack this year. A draft sync was set to update 14 contact records. The agent's summary said exactly that: 14 records, contact details refreshed. If that summary had been the gate, I'd have approved it. The diff said something else. The contact owner field was changing on every one of those records, because a field mapping in the sync config pointed at the wrong column. Nothing was broken, so no error would ever have fired. The sync was valid. It was just wrong. Reading the actual field changes took about forty seconds, and the sync died there. That's the difference between a gate and a rubber stamp. Every gate like that is an error that never reaches a customer record. Every scoped permission is an incident nobody bills cleanup hours for. ## Governed teams ship more agents, not fewer Schellman's 2026 data settles the "governance slows us down" argument. The governed teams ship more, not fewer: | Governance maturity | Has AI agents in production | | --- | --- | | Mature governance program | 78% | | Developing governance program | 22% | | All surveyed organizations | 46% | Source: [Schellman](https://www.globenewswire.com/news-release/2026/07/29/3335281/0/en/New-Schellman-Research-74-of-Enterprises-Say-They-Are-Audit-Ready-for-AI-Only-27-Actually-Are.html) , survey of 525 US enterprise practitioners, July 2026. That tracks with the wider adoption numbers, which are ugly. In September 2025, BCG surveyed 1,250 senior executives across nine industries. 5% of companies came back "future-built" on AI. 60% reported minimal gains ( [BCG](https://www.bcg.com/press/30september2025-ai-leaders-outpace-laggards-revenue-growth-cost-savings) , 2025). The 35% in the middle are the ones BCG calls Scalers. IBM's CEO study rhymes: only 25% of AI initiatives delivered their expected ROI, and 16% scaled enterprise-wide ( [IBM IBV](https://newsroom.ibm.com/2025-05-06-ibm-study-ceos-double-down-on-ai-while-navigating-enterprise-hurdles) , 2025). I've written elsewhere about [why the demo worked and the rollout died](/blog/poc-easy-integration-hard) ; the short answer is that ungoverned pilots don't survive contact with production. ![](https://storage.googleapis.com/promptmetrics-uploads/website/content-images/1786720638896-332041273.png) Does the failure rate mean you should wait? No, because the work isn't waiting. BCG's same research puts agentic systems at 17% of total AI value in 2025, projected to reach 29% by 2028 ( [BCG](https://www.bcg.com/press/30september2025-ai-leaders-outpace-laggards-revenue-growth-cost-savings) , 2025). Agents are arriving whether or not you staff the job. The Schellman 78/22 split just tells you which version of the arrival you get. ## Who should hold the job in a lean team? Not a new hire. Not your most technical person by default. Only 37% of enterprises point to the CIO or Head of IT as accountable when AI risk shows up ( [Schellman](https://www.globenewswire.com/news-release/2026/07/29/3335281/0/en/New-Schellman-Research-74-of-Enterprises-Say-They-Are-Audit-Ready-for-AI-Only-27-Actually-Are.html) , 2026). Below that, the answers scatter. Write one name down and you're ahead of most. The right name: whoever can state the workflow, the data it touches, and the failure that hurts. In a 10-person company that's usually the founder or the ops lead. The enterprise answer looks different. Writer CPO Jevan Soo Lenox proposes three new roles in [The AI Agent Owner](https://writer.com/blog/ai-agent-owner/) (Feb 2026): AI Agent Builder, AI Owner, AI Champion. Fine at Fortune 500 scale. But a RevOps manager with an AI mandate and no budget clarity can't requisition an org chart. What can you actually run this quarter? The minimum viable version of the job: * **One named owner.** A person, not a committee. * **One written access list.** What the agent reads, what it writes. A page, not a policy binder. * **One human gate on every write action.** The reviewer sees the actual change before it ships. * **One log.** Every approval recorded: who, what, when. That list isn't theory. It's the task list we scope in a First Skill Sprint, and I'm publishing it because nobody else in the search results writes the job down as tasks. An agent nobody gates isn't a productivity story. It's an unstaffed governance job with write access. If you're evaluating outside help to custom-build agents for your business, here are [the questions to ask anyone who builds this for you](/blog/ai-implementation-agencies-in-the-eu-how-to-evaluate-them-2026). ## Frequently asked questions ### What is AI agent governance? AI agent governance is the set of rules and review steps controlling what an agent can read, what it can change, and who approves its actions before they ship. Gartner named inadequate risk controls a driver of the over 40% of agentic AI projects it expects to be canceled by end of 2027 ( [Gartner](https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027) , 2025). ### Who is responsible for AI agent governance in a company? Whoever owns the workflow the agent touches: the RevOps manager for CRM agents, the CS director for support agents. Most companies haven't settled it; only 37% point to the CIO or Head of IT as ultimately accountable for AI risk ( [Schellman](https://www.globenewswire.com/news-release/2026/07/29/3335281/0/en/New-Schellman-Research-74-of-Enterprises-Say-They-Are-Audit-Ready-for-AI-Only-27-Actually-Are.html) , 2026), and no other role commands a majority. Write one name down. You're already ahead. ### Do small teams need AI agent governance? Yes, scaled down: one named owner, one written access list, one human gate on write actions, one log. The perception gap exists at every size; Schellman found 74% of surveyed enterprise practitioners believe their organization is AI audit-ready while 27% run fully mature governance ( [Schellman](https://www.globenewswire.com/news-release/2026/07/29/3335281/0/en/New-Schellman-Research-74-of-Enterprises-Say-They-Are-Audit-Ready-for-AI-Only-27-Actually-Are.html) , 2026). Small teams just close it faster. ### Is AI agent governance a technical job? Partly. Vercel's CEO says it needs people who understand data flows, threat models, and systems design, and that the code-writing part matters less (Agent Native podcast, 2026). Operators qualify: with 54% of CEOs hiring for AI roles that didn't exist a year ago ( [IBM IBV](https://newsroom.ibm.com/2025-05-06-ibm-study-ceos-double-down-on-ai-while-navigating-enterprise-hurdles) , 2025), workflow knowledge is the scarcer half. ## The job exists. Assign it. What Rauch said out loud, the numbers confirm: * Vercel's CEO named the job: governance, data access control, oversight. "Literally our new job." * The 74%-versus-27% Schellman gap is that job, unstaffed and measured. * Governed teams ship more agents to production (78% versus 22%), not fewer. * The owner is whoever knows the workflow. In a lean team, that's probably you. I've made a strong claim here: that governance is an ops job wearing a security costume. Tell me where I'm wrong. _Disclosure: I drafted this with a coding agent in the loop and reviewed every quote, number, and link myself. That's the workflow this post recommends, applied to the post._