AI usage inside enterprises is spreading well past engineering. Vercel reports that token usage by enterprise companies on AI Gateway quadrupled over the last three months, and OpenAI's data shows non-developer Codex users growing 189x between August 2025 and June 2026. Vercel calls this population citizen developers: people who understand their own part of the business and can now build the software they need. The same capability that makes building easy makes governance hard — employees adopt AI tools whether or not they are approved, and a login page vibe coded incorrectly can expose private data to the public internet.
Enterprise Apps and Agents is Vercel's answer: a central control plane for every builder, app, agent, token, and dollar. Two capabilities are new:
- Bring Your Own Cloud (BYOC) on AWS is generally available, keeping compute and data inside your AWS boundary while preserving the full Vercel developer experience.
- Enterprise-Managed Authorization through Vercel Connect enters public beta, letting agents reach connected services on behalf of employees through your IdP.
Governance across every control domain
Apps and agents deploy to one destination shared by engineers and citizen developers, governed centrally for access, providers, models, data retention, token budgets, integrations, and policies, all driven through the IdP you already run. Company policy applies consistently rather than per-tool.
Control domain | Vercel controls |
|---|---|
Deployment governance | |
Access governance | |
Integration and data governance | |
Model and token governance | AI Gateway, including US-only inference and budget controls |
Workloads land on the infrastructure behind Vercel's production deployments. A single surface covers observability and security monitoring across projects via analytics, audit logs, and the Security Dashboard. Secure Compute supplies private connectivity to enterprise systems, and BYOC on AWS runs application functions inside your own AWS account under your permissions and network controls.
Who deploys what
The tooling is fragmented by role: a marketer on ChatGPT building a landing page, an analyst on Claude Code running queries, an engineer on Codex, OpenCode, or fx. Vercel serves as a single deployment target regardless of harness and connects to your existing IdP to decide who may create deployments.
Klaviyo previously spent two weeks of IT tickets and platform-team follow-ups to ship one simple internal app. It built an internal app-building platform on Enterprise Apps and Agents; ideas now deploy in three minutes. In its first two weeks on Vercel, 512 employees shipped 356 apps.
"Vercel let us engineer guardrails instead of walls."
Mohamed Ali, Sr. Lead AI Architect at Klaviyo
Deployment control alone does not protect what reaches production. Deployments can be configured private by default, and Vercel Passport requires employees to authenticate through your IdP before touching internal apps.
Connectors, tokens, and data scope
Agent usefulness depends on access to systems and data, which traditionally means long-lived environment variables with broad permissions to CRM, databases, Slack, and similar systems.
Vercel Connect builds connectors with scoped access and mints a token per agent task, so an agent reaches only the context that job requires. Enterprise-Managed Authorization removes the separate consent prompt: once an employee signs in through your IdP, agents can use connected services automatically. Because Connect integrates with Passport, employee permissions cascade to the integration and data layer — no visibility in the source system means no access through an app or agent.
Models, tokens, and spend
The average enterprise on Vercel AI Gateway routes across 17 models. Without a router that means separate provider integrations and API keys and budgets scattered across platforms.
AI Gateway exposes hundreds of models through one API with no markup on cost, Zero Data Retention so data is neither stored nor trained on, and optional routing through US-only providers for compliance.
Mercor, which connects AI labs with subject-matter experts who train and evaluate models, needed engineers to reach hundreds of models. Direct integrations and third-party routers had made stolen credentials a major risk and token spend unmanageable. After moving routing to AI Gateway, engineers use a single expiring secret with per-person and per-project budgets. Mercor processed 63 billion tokens in its first 10 days on the gateway without hitting rate limits.
"Not all gateways have the capability to do time-expiration keys or limits at the key level. As an admin, being able to provision keys with guardrails, and not having fears of stolen keys or runaway spend months later, has been particularly valuable."
Abhi Kottamasu, Engineering Manager at Mercor
Visibility and security findings
Vercel Observability collects app and agent telemetry from every project in one view, covering internal tools and customer-facing products alike. The Security Dashboard surfaces risks across teams and apps — missing MFA, unprotected preview deployments, long-lived credentials, framework vulnerabilities — and each finding links to the setting that resolves it.
Compute inside your own AWS account
With BYOC on AWS, now generally available, application functions execute in your AWS account within your security and network boundaries. Meta Superintelligence Labs runs Vercel this way, letting its AI teams test and deploy inside their own AWS environment; Meta leaders said the first week on Vercel shipped a month of work.
From idea to global scale
Vercel's own Vertex support agent illustrates the path from a citizen-developer idea to production. A vendor product could not answer every customer question, so the support team built agents and skills to close the gaps and improve response times, laying the foundation for a customer-facing experience. The GTM engineering team then added evals and model failover. Vertex now resolves 91% of support cases autonomously across thousands of tickets per week, on the same platform where the work began.
