AI agent cost Module

Vaks PM · Integration guide · Agent cost & billing · August 2026

The principle — an agent is measured in usage, not hours

An agent has no timesheet: its capacity is never hours, and its consumption is not counted like a human's. It is measured in model usage — the volume of text processed by the large language model that drives it. Vaks PM turns that usage into a money amount on two distinct counts: what the agent costs you (internal cost) and, where relevant, what you charge a client (selling price). Both feed the same profit & loss view as human work.

Tokens or credits — two units depending on the provider

The unit depends on how the model's provider bills, and Vaks PM picks it automatically:

Internal cost — what the model costs you

An agent's internal cost is derived from configurable rates, set on the agent's profile (Cost tab):

• Two rates for a token-billed agent: one per 1,000 input tokens and one per 1,000 output tokens — because an LLM's output costs several times its input. A run then costs input_tokens/1000 × in_rate + output_tokens/1000 × out_rate.
• One credit price for a Copilot agent: an amount per credit (set globally in policies, see estimate), multiplied by the number of credits measured.

These rates need not be typed in model by model: the price catalog pre-fills them for known models.

Billing a client — a separate selling price

The internal cost (what you pay) and the billed price (what you sell) are two independent values, deliberately decoupled. The selling price is a single rate per 1,000 units (tokens or credits, input and output combined) — a commercial price, unrelated to the provider's cost structure. An agent with no billing rate is simply non-billable: its internal cost is still tracked, but nothing goes to revenue.

Rate resolution — from most specific to default

For a given project, the rate applied to an agent resolves in priority order:

1. The agent's rate on this project — an override set project by project (under the project's finance), when a client negotiates a specific price.
2. The agent's default rate — the one on its profile, applied everywhere else.
3. Non-billable — with neither, the agent's work generates no revenue (the cost is still always tracked).

Where the number comes from — three sources

Knowing a run's real cost means knowing its real usage. Vaks PM accepts that figure in three ways, from simplest to most reliable. They do not exclude each other — a real measured or declared figure is never overwritten by an estimate.

Declared — the agent reports its usage

After working, an agent declares its run cost: the number of input and output tokens (and optionally an already-computed cost). If the agent supplies only tokens, Vaks PM derives the cost from its rates — and always keeps that derived amount as an anti-cheat reference: a declared cost that strays too far from what the tokens imply is flagged as an anomaly. The declared cost wins, the reference is the guardrail.

Measured — the LLM gateway

For an exact measure that does not depend on the agent's declaration, Vaks PM provides an optional LLM gateway: a lightweight relay, deployed on the client side as close to the agents as possible, that the model calls (Anthropic, OpenAI and compatibles) flow through. The gateway reads the real usage from the provider's response and reports it to Vaks PM, marked as measured. The agent declares nothing.

Estimated — credit-billed agents (Copilot)

Some agents are not measurable in tokens: a Microsoft Copilot Studio agent is billed by Microsoft in credits, and the per-call detail is not exposed in tokens. So this cost does not escape the P&L, Vaks PM estimates it: it counts the agent's activity — a baseline per task, plus an increment per tool call — in credits, then converts them to money via the credit price you set.

The model price catalog

Converting tokens into money means knowing each model's price. Vaks PM keeps a price catalog that serves two purposes: offering the list of models when creating an agent, and pre-filling its internal cost rates.

Refresh — optional and opt-in

Online refresh is off by default. An administrator turns it on with a toggle in policies; it then runs about once a day, adds or updates prices, and flags drift when a price moves significantly. Two guardrails:

• A hand-set price is never overwritten. If you set a model's rate yourself, the feed respects it — it only touches what it set itself.
• The network must be open. Refresh reaches openrouter.ai and LiteLLM's price repository. On an instance with no internet egress, leave it off and rely on the offline-shipped list (which can be extended by hand).

Where the cost shows — P&L, reports, exports

Once known, an agent's cost (and revenue) joins the rest of finance, alongside human work:

• Project profit & loss — agents' real cost enters the project cost, their forecast enters the estimate-at-completion, and billed agent revenue enters revenue.
• Project & portfolio reports — agent cost and revenue are broken down by project and client, with an “of which AI agents” sub-line.
• Exports — CSV, spreadsheet (XLSX) and PDF carry the agent breakdown.

Where to configure

See also: AI agents — cost & billing (overview) · Autonomous Copilot agent — the execution cost · Agent directories (rates at inbound registration) · all integrations.