C1.ai launches LLM Gateway to route enterprise model calls by policy

C1.ai said C1 LLM Gateway is available today as the final launch of C1 Launch Week, putting governed routing and metered inference in front of enterprise model traffic so security teams decide which data may leave the boundary and finance teams can see who spent what before the invoice lands.

Enterprises are spraying prompts at a default lab with almost no record of who spent what or what data left. A gateway that treats model access like a permission — who may call, what it cost, and the right to shut it off — is what has to sit in front of agents before they touch real systems.

On Thursday, 1 October 2026, C1.ai said it launched C1 LLM Gateway. The GlobeNewswire page is titled “C1.ai launches C1 LLM Gateway to govern enterprise AI model routing.” The page stamps October 01, 2026, 09:00 ET, which is 9:00 a.m. Eastern. The dateline is San Francisco. The source line is C1. The line under the headline says C1.ai routes every model call by sensitivity and cost, attributes each call and its cost to the user, agent, or application that made it, and makes the routing decision a policy the security team owns. The same line calls this the fourth and final launch of C1 Launch Week. An agent, in that sentence, is software that can take a next step for a person, not only reply in a chat. Those lines are the wire’s.

What the opening says is going wrong. The release says enterprise AI traffic is outpacing the controls needed to govern it. Sensitive data crosses model providers with little record of which user, agent, or application made a call, or what the call cost. A model provider is a company that runs an AI model and answers what people send it. A prompt is that input: the text, or other material, a person or a program sends. Without a governing policy, the release says, model traffic follows defaults instead of explicit rules about which providers may receive which data. Spend is hard to trace to the responsible user, agent, or application until the provider invoice arrives as one number, too late to prevent an overage. An overage is a bill that runs past what the company meant to spend. The same paragraph says sensitive data may reach unintended providers, and that dependence on a single lab can deepen by default. A lab, here, is the company that built and hosts the model. Those lines are C1.ai’s.

How the gateway routes, as the release states it. A gateway, here, is the door a model call has to pass through before a provider sees it. C1.ai puts that routing decision under policy. Governed routing decides which models each identity is allowed to reach. An identity, in that sentence, is the user, the agent, or the application making the call. Sensitive work routes to infrastructure the customer controls. Everything else runs wherever it is fastest or cheapest. Certain data can be blocked from reaching certain providers outright. The release says that block is how a company protects proprietary data, meaning information it treats as its own and does not want a given provider to see. A day-one line says organizations control where their most sensitive data goes while reducing dependence on any single model provider. Those lines are the release’s. The release does not name the providers, and it does not print a price for the gateway.

Who the bill names, and how a budget can be pulled. The release says a smart routing policy determines where model traffic goes, and that metered inference makes the cost visible. Inference is the step where a trained model answers. Metered means each call is measured, so the usage can be counted. Users, agents, and applications call models through C1.ai. Each request can then be attributed to the caller, the model that answered, and the resulting cost, instead of disappearing into a single provider invoice. Alerts fire before an overage instead of after it. Model budgets are governed like access rights. A person can request an increase. An application owner approves it. Access can be time-limited. The day-one line also calls that time-boxed, meaning the permission expires. Access can be revoked. Revoking model access takes effect immediately, and the next call fails. The day-one line says the right is pulled the moment someone leaves. Those lines are the release’s.

The three lines the release prints for day one. Route by sensitivity and cost: the company decides where the most sensitive data goes, and it is less tied to one model provider. Every dollar has a name on it: spend is attributed to the user, agent, or application that made each call, with alerts before an overage instead of after the invoice. Revoke in one place: model access is a governed right, requestable when a budget runs out, approved by the application owner, time-boxed, and pulled the moment someone leaves. Those three lines are the release’s labels for the same routing, attribution, and revocation it describes above.

Alex Bovee, chief executive and co-founder of C1.ai, is quoted on the release. “Every prompt your teams send is a description of how your company works, and most companies are sending them to a single vendor by default,” he said. “That decision belongs to you. Policy should decide which models see which data, and you should know what every one of those calls costs and who made it.” A vendor, in that sentence, is the model provider. That quotation is his, in the release.

Where the product sits. The release says C1 LLM Gateway is available today and is the final launch of C1 Launch Week. It is part of C1 Run, which the release calls the runtime layer for secure agents. A runtime layer is software a program passes through while it is running, rather than a report written after the fact. Alongside C1 MCP Gateway, the release says, the two products bring identity and policy to model calls and to tool calls, so a company can put AI to work without giving up control. A tool call is the step where an agent uses a company system, not only the step where a model writes an answer. The release says to see the product at c1.ai and at C1 Transform in San Francisco on October 6. It calls this the fourth and final launch. It does not name the earlier three launches in this announcement. Those lines are C1.ai’s.

Who the about box says C1.ai is. C1.ai calls itself the identity platform for the AI era. The box says the platform connects humans and agents to enterprise systems with scoped access, enforced policies, human approval, and a record of every action. Scoped access means a person or an agent gets only the slice of a system the company allowed. The box says the fastest path to adoption at scale is the governed path. It names Ramp, Zscaler, Qualtrics, and DoorDash as companies that trust C1.ai to govern identity, secure agents, and let their teams put AI to work. Those four names sit in the about box, as companies on the platform. The release does not say they are case studies for LLM Gateway, and it does not say what each of them bought. The media address on the release is press@c1.ai.

What the company product page adds. The page at c1.ai/products/llm-gateway says the gateway routes inference across supported public, private, and customer-controlled model endpoints under company policy. An endpoint is the address a program calls. The page says the customer’s team owns the models and the inference deployments, and that C1.ai governs the route an application uses to reach them. It says the gateway does not host the models. Routing on that page can use identity, the workflow, data-handling requirements, region, and cost. Inside the allowed set, the page says the choice can also use capability, health, latency, and cost. Latency is how long the answer takes to come back. Applications can call one C1.ai endpoint, so a provider key does not have to sit inside the application code for every route. The page separates the two gateways. C1 LLM Gateway owns inference routing. C1 MCP Gateway governs tool and API actions after an agent or a client chooses to act. An API is a door one program uses to ask another program to do something. The page says spend can be rolled up across providers, models, and callers, and that a budget can stop unplanned use before it becomes an overage. The page offers a demo. It does not print a dollar price.

The picture is the C1 Gateway product screen, on the Spend overview tab, with black bars at the sides. The header reads C1 Gateway. The tab row includes Connected clients, Settings, Overview, Attribution, Controls, Blocks, MCP traffic, MCP findings, LLM usage, and LLM attribution. Overview is the selected tab. The line under Spend overview says the screen tracks organization AI spend, the remaining limit, and expected spend for the current period. A control on the screen reads Last 30 days. The cards cover spend this period, with a projected figure beside it, money in flight, what remains, an organization limit, and spend for the selected dates. Two smaller cards show a count of blocks and a count of apps. A chart is labeled Spend and forecast. Its caption says the bars are daily spend for the selected dates, followed by an expected range through the end of the current limit period. Darker bars cover the selected dates. Lighter bars continue as the forecast. A lower row is labeled Billable tokens, Calls, Cost per million tokens, and Cache savings. A token, on that label, is the small unit many model providers use when they bill. The left rail lists Home, Requests, Go links, Reviews, My resources, My AI spend, Apps, Governance, Identities, Security, Vaults, and Platform. The dollar figures, the token count, the call count, and the bar heights are labels on that product screen. The GlobeNewswire release does not print them. They are not a price for the gateway, and they are not a named customer’s bill. The screen does not print a calendar date.

In plain terms, C1.ai said on Thursday that C1 LLM Gateway is available, as the last launch of C1 Launch Week and as part of C1 Run. A security team’s policy decides which model provider may see a prompt. Sensitive work can stay on machines the customer controls. Other work can go to the fastest or cheapest allowed provider. Some data can be blocked from some providers. Each call is supposed to be tied to the user, agent, or application that made it, to the model that answered, and to the cost, with an alert before the spend runs past the budget. A budget can be requested, approved, set to expire, and pulled so the next call fails. The about box names Ramp, Zscaler, Qualtrics, and DoorDash as companies on the C1.ai platform. It does not present them as studies of this gateway. The product page does not print a price. Alex Bovee’s line is that a prompt describes how a company works, and that the choice of who sees it should be a policy.