Kochava on October 22, 2026 added a set of cost controls to the Enterprise tier of StationOne, its desktop hub for running AI models and agents in marketing work. The release, issued from Sandpoint, Idaho, describes an automatic router that sends each AI request to the cheapest suitable model, token and dollar budgets that administrators can assign to individuals and teams, usage reporting tied to business tasks, two plugins that take work away from large models altogether, and the option to run compatible models on a user's own computer. Kochava did not publish prices, availability dates or any measured savings.

In Short

Kochava makes a program that lets marketers use several AI chatbots and agents in one place, and it has added tools for big companies to keep track of how much that AI use costs. You pay AI providers by the amount of text a model reads and writes, so a team running many automated tasks can run up a bill without anyone noticing until the invoice arrives. The new controls send easy jobs to cheaper models, let managers set spending limits by person or team with alerts, and allow some work to run on a computer's own hardware, though Kochava has not said how much money any of this saves.

What Kochava released

According to Kochava, the update is "a major token efficiency expansion within the Enterprise Tier of StationOne," and it is meant to give enterprises "fine-grained visibility and control over AI usage costs." The company frames the problem plainly. As marketing teams lean on large language models to automate campaign workflows, audience management and reporting, "the cost of running those models at scale has become a growing concern for enterprise budget owners," the release says. Kochava calls unpredictable and opaque token spend "one of the biggest barriers to scaling agentic AI across large organizations."

The release does not quantify that barrier. It contains no figure for how much a typical StationOne team spends on model calls, no estimate of how much the new features cut, and no named customer. The summary line on the GlobeNewswire page refers to "runaway AI token costs," but that phrase is not backed by any number in the text.

The design principle Kochava describes is a split between two kinds of work. Some tasks need the reasoning of a large model. Others are mechanical and can be done by ordinary code at a fraction of the cost. The Enterprise features pair "deterministic, code-based execution with AI orchestration," so teams get the benefits of agentic AI "without paying LLM-level costs for tasks that don't require LLM execution," according to Kochava.

Charles Manning, Kochava's founder and chief executive, tied the change to how adoption grows inside a company. "We built StationOne so marketers could put agentic AI to work on real campaigns, not just experiment with it," Manning said in the release. "As adoption scales inside an organization, cost visibility becomes just as important as capability. Teams need to know not just what AI can do, but what it's costing them to do it, task by task."

The seven capabilities

Kochava lists seven features in the Enterprise offering. They fall into three groups: routing, budgeting and reporting, and moving work away from paid models.

Routing

The Auto Router sends each AI engagement to "the most cost-appropriate model based on task complexity," with the behaviour managed through enterprise account settings. The stated aim is to stop teams "paying premium model pricing for simple tasks." The release does not say how StationOne judges complexity, which models sit in the routing pool, whether users can override a routing decision, or whether the router can be switched off for specific workflows.

Pricing gaps between models make the logic easy to follow. AI providers charge per token, the small fragments of text a model reads and writes, and the per-token rate for a provider's largest model is typically many times the rate for its smallest. A routine reformatting job sent to a top-tier model pays the top-tier rate.

Budgets and reporting

Per-User and Per-Team Budget Allocation lets enterprise admins set budgets measured either in tokens or in dollars, by user or by team. According to Kochava, the feature includes "configurable overrides and automatic reporting when usage approaches or exceeds set thresholds." The wording matters. The release describes alerts and reports when a threshold is reached or passed; it does not say that usage stops at the limit. Whether a budget acts as a hard ceiling or a warning line is not stated.

The Personal Analytics UI & Alerting feature gives individual users a view of their own consumption: which tasks they are running AI against and what each one costs in tokens or dollars, "both current and projected." How projections are calculated is not explained.

Enterprise Token Reporting Linked to Tasks rolls usage up across individuals and teams. Kochava stresses that the reporting is tied "specifically to the business tasks being performed - not just raw chat conversation logs," which it says gives leadership "clearer ROI visibility." That is a meaningful distinction for a finance team. A log of conversations shows who spent tokens; a task-level rollup shows what the tokens were spent on.

Taking work away from paid models

Two plugins aim to shrink the amount of work that reaches a large model at all.

The Audience Management Plugin addresses a common inefficiency: pushing large audience files through an LLM. According to Kochava, the plugin "removes the need to pass high-volume audience data through LLMs at all." Instead, the StationOne harness runs "deterministic, low-token code for audience deduplication, appends, reductions, and segmentation," and the model only orchestrates the steps rather than processing the records itself. Audience lists can run to millions of rows, and every row passed through a model is billed. Keeping the data in code and letting the model issue instructions is the cost logic here, and it also means the audience records themselves are not sent to the model provider for those operations, although the release does not make that privacy point explicitly.

The Calibrated Decision Plugin routes "calibrated decision-making tasks" through a new native plugin as "a lower-cost plugin alternative to full model usage for calibrated decisions and scoring," reserving LLM calls "for the scenarios that truly need them." The release names the first supported option as "TypeSafe - Jev" and gives no further description of what it is, who builds it, or how its scoring is calibrated.

Embedded Model Support is the seventh feature and the one with the widest reach. StationOne now supports "the optional ability to run compatible AI models directly on a user's own machine," according to Kochava. Users choose a model that suits their hardware and performance needs, while administrators keep the flexibility to support a broader model strategy across the organization. Kochava lists three benefits: less dependence on externally hosted inference, lower token consumption for routine tasks, and "an additional execution path" for privacy-sensitive workflows. The release names no compatible models and sets out no minimum hardware.

Media-spend discipline, applied to AI

Manning's second quote draws the comparison Kochava wants buyers to make. "With the token efficiency features made available in the new Enterprise release, we're giving enterprises the same financial discipline over AI spend that they already apply to media spend: route intelligently, set budgets, see usage in real time, and only pay premium model costs when the task actually demands it," he said.

The comparison is apt for an audience of performance marketers. Media budgets are allocated by channel and team, paced against targets and reconciled against outcomes. Model usage, by contrast, often sits on a single API account or a handful of seat licences, with little breakdown by task. Kochava's pitch is that the measurement company that already reports on media spend can report on AI spend in the same way.

What the release leaves open is just as relevant for procurement. It does not say what the Enterprise tier costs, whether the controls are included in existing contracts, whether model usage is billed by Kochava or passed through from the model providers, or when the features become generally available. The StationOne website, which offers a free account and download, does not list tier pricing.

Where StationOne came from

StationOne has moved quickly from a free desktop client to an enterprise product. Kochava introduced StationOne on November 25, 2025 as a free app for Windows, macOS and Linux that connected Anthropic's Claude, OpenAI's GPT, Meta's Llama and custom models through one interface. At launch, Kochava said data stayed on local devices, and the product already offered some local model processing and offline functions. PPC Land noted at the time that the company did not explain how that claim applied when cloud-hosted models were used. The new embedded model feature, pitched as an extra path for privacy-sensitive work, implicitly concedes that the cloud models in StationOne are a separate case.

Four months later, Kochava removed the invitation code requirement and opened StationOne to public beta with an IAB Tech Lab AAMP workspace on March 25, 2026, carrying 19 skills across eight functional areas and running on IAB Tech Lab's reference MCP server. The Q2 2026 product bulletin added a Kochava AI Workspace that connects to a customer's Kochava account at no extra cost, plus a Yahoo DSP workspace.

On August 13, 2026, StationOne gained workspaces for Google, Meta, Reddit, Snap and TikTok for Business, covering campaign builds, bulk changes, audience and creative operations, and reporting. Manning said then that "StationOne is built for enterprise-grade governance and privacy objectives." Two more releases followed. On September 3, 2026, Kochava announced creative workspaces with Higgsfield, Luma and Spaceback by Rembrand. On September 30, 2026, it added workspaces for Adjust and AppsFlyer and MCP connectors for Adjust, AppsFlyer, Branch and Singular, meaning a Kochava product now connects directly to its main attribution rivals.

Each of those additions multiplies the number of workflows that can call a model. The October 22 release is, in effect, the bill arriving for that expansion: more workspaces and more agents mean more tokens, and enterprise buyers want to see where they go.

Kochava's broader business has kept moving alongside StationOne. The company launched Atlas Performance, a supply performance system for premium publishers, on February 11, 2026, and the Federal Trade Commission closed its nearly four-year location data case against Kochava with a stipulated order in May 2026. Most recently, a co-branded guide from Jampp, Kochava and YouAppi claimed CTV comes before search in 96% of installs that touch both.

Why cost controls are arriving now

The release lands in a year when the economics of AI agents have become a board-level question. KPMG's Global AI Pulse survey for Q2 2026, covering 2,145 senior leaders, found that 49% of organizations had scaled back, narrowed, delayed or paused AI agent deployments because costs outweighed value. Only 35% of respondents in that survey said their AI operating costs were fully visible and actively monitored, and those with full visibility were five times as likely to report established ROI - 15% against 3%, by PPC Land's arithmetic. Task-level token reporting is aimed squarely at that visibility gap.

Inside marketing departments, scale has lagged ambition. A Typeface survey of more than 200 marketing leaders, released on June 22, 2026, found that 86% use AI agents in campaign execution but only 36% have scaled at least one. Moving from pilot to production is precisely the point at which usage, and therefore spend, rises sharply.

The ad tech plumbing that StationOne relies on has also been in flux. The Model Context Protocol, which underpins most of the platform connectors StationOne uses, went through its largest specification rewrite since launch on July 28, 2026. And the case for keeping some work out of language models entirely has a parallel elsewhere in the industry: PPC Land has reported that taxonomy identifiers, not natural language, are what keep multi-agent advertising workflows deterministic. Kochava's audience plugin applies the same principle to data processing.

What is still missing

For a release built around financial discipline, the missing numbers stand out. There is no price for the Enterprise tier, no general availability date, no benchmark of the Auto Router's accuracy or savings, and no statement on whether the routing decisions are logged for review. The budget feature is described in terms of alerts and reports, not enforcement. The embedded model option comes with no list of supported models or hardware. And the Calibrated Decision Plugin is introduced through a single product name, "TypeSafe - Jev," without explanation.

None of that makes the features less relevant to buyers. Model routing, per-team budgets and on-device execution are becoming standard requirements for enterprise AI software, and a marketing platform that offers them answers a question finance departments are already asking. Whether StationOne's versions deliver the savings implied will depend on figures Kochava has not yet published.

Timeline

Summary

Who: Kochava, the Sandpoint, Idaho measurement company, and its founder and chief executive Charles Manning. The features are aimed at enterprise administrators and marketing teams using StationOne.

What: Seven additions to StationOne's Enterprise tier: an Auto Router that sends tasks to the most cost-appropriate model, token or dollar budgets by user and team with threshold alerts, personal usage analytics, token reporting tied to business tasks, an Audience Management Plugin that processes audience data in code rather than through LLMs, a Calibrated Decision Plugin, and support for running compatible models on a user's own machine. No prices, availability date or savings figures were given.

When: Announced on October 22, 2026, at 09:30 ET.

Where: In StationOne, Kochava's desktop AI hub for Windows, macOS and Linux, announced via GlobeNewswire from Sandpoint, Idaho.

Why: Kochava says unpredictable and opaque token spend is one of the biggest barriers to scaling agentic AI in large organizations. The release follows a run of StationOne workspace launches that widen the number of workflows calling paid models, and survey data showing many organizations pulling back on AI agents when costs outrun value.