Speakeasy launches tool to track AI agent costs across the workforce

Speakeasy Development Inc. has launched AI Cost Control, a service for tracking what employees spend across AI coding agents and related tools, with pricing for the product not disclosed. The release targets a growing blind spot for enterprises: AI usage that is spread across vendor dashboards, desktop apps, subscriptions and personal accounts used for work.

The service is part of Speakeasy’s AI Control Plane, a layer the company says sits between workers and the AI tools they use. AI Cost Control gathers usage telemetry from services including Anthropic PBC’s Claude Code, Anysphere Inc.’s Cursor, Claude Cowork and OpenAI LLC’s Codex. Speakeasy said it records tokens used, model selection, cache behavior and estimated cost for individual interactions, then rolls that data up by employee, team, tool and account type.

Chief Executive Sagar Batchu said enterprises lack a single operating view because each model provider reports only its own activity. Anthropic, for example, can show spend attached to an Anthropic key, he said, but not usage across rival models and tools.

Personal subscriptions create a reporting gap

The product is aimed at companies where AI procurement has moved faster than finance and IT controls. Centralized admin consoles can show managed accounts, but Speakeasy argues they miss work done through employee-purchased subscriptions. That matters for companies trying to understand whether AI agent adoption is reducing work, shifting costs into expense reports or creating unmanaged data exposure.

Speakeasy’s approach uses a software agent installed on employee work machines. The company said the agent observes AI session activity on enrolled corporate devices and maps that activity to the employee’s work identity, including cases where the worker accesses an AI service through a personal account. Batchu said the company does not log into personal accounts or collect AI activity from personal devices.

The data collected can include session transcripts, which raises the same privacy and governance issues that surround other employee observability products. Batchu described the service as workforce observability and said Speakeasy has customer agreements covering storage and use of customer data. The company also said it can provide dedicated data residency for some enterprises and deployments inside customer virtual private clouds.

What the service measures, and what it misses

Speakeasy said AI Cost Control uses the OpenTelemetry Semantic Conventions for generative AI, a shared format for reporting details such as token counts, model names, costs and tool-call metadata. The company normalizes those records so customers can compare usage across providers rather than read separate billing pages.

Current features include cost estimates for each turn in a session, cross-agent reporting and breakdowns by employee, team, model, session and source. Customers can inspect individual sessions to see which tools, skills and Model Context Protocol servers contributed to consumption, according to the company.

Speakeasy is positioning the product against AI gateways and observability tools by arguing that those systems do not cover enough employee behavior. Batchu said gateways track requests that pass through them, while observability platforms typically measure model calls made by a company’s own applications. Speakeasy’s claim is narrower: it is trying to measure workforce AI spend, including usage from desktop clients and subscription tools that do not route through a gateway.

The roadmap includes enforceable budgets for teams and individual workers. Speakeasy said customers will be able to send notifications, require approvals for continued high-cost use or limit activity that passes through Speakeasy’s gateway after a threshold is reached.

Planned additions also include categorizing spend by tasks such as code review, research and ticket triage, separately measuring tokens consumed when agents load system prompts, skills and MCP servers, and reconciling flat-rate subscriptions with metered API charges. Speakeasy said the service does not yet cover work done on unmanaged personal devices or AI features embedded inside software-as-a-service products, though embedded AI coverage is planned.

This story draws on original reporting from SiliconANGLE.