Vercel Enterprise AI Governance Platform: Safety Guardrails for Citizen Developers

AI adoption inside enterprises is spreading rapidly, and it has long since moved beyond engineering teams. According to data shared by Vercel, enterprise token usage on AI Gateway quadrupled over the past three months; OpenAI also reported that from August 2025 to June 2026, non-developer Codex users grew 189-fold.

This group is known as "citizen developers." They understand the business functions they are responsible for better than anyone else, and with AI, they can build the software they need themselves, deliver work faster, and improve customer experiences.

But as building becomes easier, control and governance become harder. Employees are using AI tools whether or not they have approval; if a citizen developer vibe-codes a login page incorrectly, private data could leak to the public internet. Enterprise leaders must answer a question: how do you let everyone participate in building without giving up security and control?

Enterprise Apps and Agents: A Unified Control Plane

Vercel's Enterprise Apps and Agents is a one-stop solution for exactly these challenges. It is a central control plane for governing every builder, app, agent, token, and expenditure.

The platform brings engineers and citizen developers together into the same deployment destination, and centrally manages access permissions, providers, models, data retention policies, token budgets, integrations, and policies through the enterprise's existing IdP (identity provider). Vercel centralizes AI governance so company policies can be enforced consistently across every control domain, covering five dimensions:

• Deployment governance: Enterprise Managed Users and RBAC
• Access governance: Vercel Passport, Deployment Protection, and Vercel Firewall
• Integration and data governance: Vercel Connect and Enterprise-Managed Authorization
• Model and token governance: AI Gateway, including US-only inference and budget controls
• Observability and security monitoring: via analytics, audit logs, and Security Dashboard

Secure Compute provides private connectivity to enterprise systems, while BYOC on AWS lets application functions run in your own AWS account, following your own permissions and network controls.

Deployment Governance: Different Tools, One Exit

Citizen developers and engineers use different tools. A marketer might use ChatGPT to build a landing page, an analyst might use Claude Code to run queries and create reports, and an engineer might use Codex, OpenCode, or fx.

No matter what toolchain is used, Vercel provides a unified deployment target and connects to the enterprise's existing IdP, thereby controlling who can create deployments.

Integration and Data Governance: Minting Tokens per Task

Agents are only valuable when they can access systems and data, but that usually means using long-lived environment variables with broad permissions over systems like CRMs, databases, and Slack. Vercel Connect addresses this by implementing scoped connectors and minting tokens for each agent task, so an agent only accesses the specific context it needs for the work.

The newly added Enterprise-Managed Authorization feature lets agents automatically use connected services when employees sign in to Vercel through their IdP, with no separate authorization prompt. Connect also integrates with Passport, so an employee's individual permissions also apply to the integration and data layers—if someone cannot see a piece of data in the source system, they cannot access it through an app or agent either.

Ephemeral credentials and token minting are core practices in a zero-trust security architecture. The traditional approach is to inject long-lived API keys or environment variables into an application; once leaked, attackers can abuse them for a long time. Scoped, short-lived tokens are the opposite: each time an agent performs a task, the system dynamically issues a credential based on least privilege that is valid only for this task and has a precise permission scope, and it expires as soon as the task ends. Even if the credential is intercepted, its exploitable window and impact are minimized. This is exactly what Vercel Connect means by "minting tokens for each agent task"—the agent only receives the specific context needed to complete the current work, rather than global read/write access to CRMs and databases.

Zero Data Retention (ZDR) means that when a model provider processes inference requests, it does not store prompts and responses and does not use them for training. This is especially critical for regulated industries such as finance, healthcare, and law, because sending sensitive data to third-party models can trigger compliance risks. By choosing gateways or providers that support ZDR, enterprises can enjoy large-model capabilities while avoiding data lingering in external systems—a common hard compliance requirement in AI procurement.

Model and Token Governance: One API for Hundreds of Models

Enterprises using Vercel AI Gateway route to an average of 17 different models. Without a model router, they would need to integrate multiple independent providers, with API keys and budget management scattered across platforms.

AI Gateway provides access to hundreds of models through a single API, with no markup, and Zero Data Retention ensures data is not stored or used for training. Enterprises can also choose to route only through US providers to meet compliance requirements.

IdP (identity provider) is enterprise-grade identity authentication infrastructure, such as Okta, Microsoft Entra ID, Google Workspace, etc., responsible for centrally managing employee accounts, login authentication, and permission groups. Capabilities such as SSO (single sign-on), MFA (multi-factor authentication), and SCIM automated user synchronization are all provided by the IdP. When Vercel says "connect to the enterprise's existing IdP," it means employees sign in using the company's unified account system instead of registering separately with Vercel. The value of this is: when an employee leaves and their account is disabled in the IdP, their access to Vercel and all downstream applications is invalidated at the same time; permission changes also only need to be maintained in one place. For enterprise security teams, the IdP is the "single source of truth" for identity governance and the handle for bringing scattered AI tools under unified control.

RBAC (role-based access control) is a permission model: the system does not grant permissions directly to individuals, but first defines roles (such as "read-only user" or "deployment administrator") and then assigns roles to users. By decoupling roles from permissions, permission management shifts from "configuring person by person" to "configuring in batches by role," greatly reducing the chance of error. ABAC (attribute-based access control) is often mentioned alongside RBAC; the latter dynamically determines access based on environmental attributes such as time, IP, and data sensitivity level. Vercel's Enterprise Managed Users combined with RBAC enables enterprises to specify by role who can create deployments and which integrations they can access, thereby maintaining critical boundaries while opening up build permissions.

A model router (Model Router / Gateway) is a unified proxy layer between applications and numerous model providers. It provides a single interface externally, while internally handling request distribution, failover, key management, usage metering, and budget control. Without it, every model an enterprise adds requires maintaining a separate set of keys, SDKs, and billing, and costs and security risks grow linearly with the number of models. A model router can also enable model failover: when a provider rate-limits or goes down, it automatically switches to a backup model to ensure service continuity. Vercel AI Gateway reports that enterprises route to an average of 17 models, showing that multi-model use is now the norm—using different models for different tasks (such as a strong model for reasoning and a cheap model for simple classification) can both optimize costs and improve results, and the router is the infrastructure that makes this strategy manageable.

Two Customer Case Studies: Klaviyo and Mercor

Klaviyo: From two weeks to three minutes. Before adopting Vercel, publishing a simple internal app at Klaviyo took two weeks—chasing IT, filing tickets with the engineering team, and following up with the platform team. They built an internal app-building platform on top of Enterprise Apps and Agents, and now anyone in the company can deploy an idea in three minutes. In the first two weeks after launch, 512 employees deployed 356 apps. Mohamed Ali, senior principal AI architect at Klaviyo, said: "Vercel lets us design guardrails, not walls."

Mercor: Managing security and budgets for massive model calls. Mercor connects AI labs with subject matter experts to help train and evaluate models. Its engineers need access to hundreds of models, but before migrating to Vercel, access was scattered across direct integrations and third-party routers, creating a huge risk of credential theft and making token spending impossible to manage. After migrating to AI Gateway, engineers access hundreds of models through a single expiring key and can set budgets by person or project. In the first 10 days after launch, they processed 63 billion tokens without hitting any rate limits. Abhi Kottamasu, engineering manager at Mercor, noted: "Not all gateways have key-level time expiration or quota capabilities. As an administrator, being able to configure keys with guardrails and not worry about keys being stolen or costs spiraling out of control months later is especially valuable."

BYOC on AWS: Keep Compute and Data Within Your Own Boundary

BYOC on AWS is now generally available. Application functions execute within the enterprise's AWS account, inside its own security and network boundary.

Meta Superintelligence Labs runs on Vercel through BYOC. Meta said its AI teams can quickly test and deploy ideas in their own AWS environment, completing what amounted to a month of work in the first week after launch.

BYOC (Bring Your Own Cloud) is a deployment model in which a software vendor's control plane runs in its own cloud, but the compute functions handling actual business workloads run in the customer's own cloud account (such as AWS). This way, data does not need to leave the customer's security and network boundary, and the customer can continue using its existing IAM permissions, VPC network isolation, encryption keys, and audit logs. For large enterprises, especially those with strict requirements around data sovereignty and industry regulation, BYOC is often a prerequisite for procurement. Vercel's BYOC on AWS lets application functions execute within the enterprise's own AWS account, delivering on the demand to "keep compute and data within your own boundary," at the cost of some operational complexity returning to the enterprise itself.

From Idea to Global Scale: Vertex in Practice

Vercel aims to let everyone in a company build safely while making it easy for engineers to scale the best ideas globally. Vertex—Vercel's own customer support agent—is a product of this philosophy. It previously used a third-party vendor but could not answer all customer questions. The support team built agents and skills to cover gaps and improve response speed, then realized this laid the foundation for a direct customer-facing experience. The GTM engineering team took over and made it production-ready through evaluations and model failover. Today, Vertex autonomously handles 91% of customer support cases, processing thousands of tickets per week, running on the same platform where the idea began.

The Core Tension in Enterprise AI Governance

Vercel's approach reflects the fundamental tension in enterprise AI adoption: the barrier to building has fallen to a historic low, while governance difficulty has risen to a historic high. Employees' AI use often exists before approval processes, and security teams cannot solve it by simply "prohibiting" it. Vercel's strategy is to embed governance into the deployment pipeline—using IdP for unified identity, Gateway for model and budget control, Passport for access verification, and Connect for permission alignment. The core idea is not walls, but guardrails.

Klaviyo's 512 people deploying 356 apps, Mercor's 63 billion tokens, and Meta completing a month of work in one week—these numbers show that when security and governance are no longer blockers, the productivity unlocked from citizen developers can far exceed expectations.

One-line Summary

Vercel launches an enterprise AI governance control plane, using guardrails rather than walls to unlock citizen developer productivity.

Paragraph Summary

Enterprise AI adoption is rapidly spreading beyond engineering teams. Citizen developers use AI to build software themselves, unlocking productivity while creating security and control challenges. Vercel's Enterprise Apps and Agents unified control plane brings engineers and citizen developers to the same deployment destination and centrally governs five dimensions through the enterprise IdP: deployment, access, integrations and data, models and tokens, and observability. Supporting capabilities include BYOC on AWS and Enterprise-Managed Authorization. Case studies show Klaviyo enabled 512 employees to deploy 356 apps in two weeks, while Mercor processed 63 billion tokens in 10 days through AI Gateway. The core philosophy is "guardrails, not walls."