What an agent task really costs

The most uncomfortable question in any agent workshop does not come from IT. It comes from finance, and it goes: what does this thing cost to run?

Until recently the answer was simple, if imprecise: whoever has a Microsoft 365 Copilot licence can use agents. With the GitHub Copilot harness that sentence no longer holds. The licence buys you the right to build. The right to operate you buy separately.

This dispatch is a reference. I keep it current, and the date of the last review is at the top.

What changed

Three sentences from the Microsoft documentation that together describe the whole model:

First: all usage is billed by consumption and measured in Copilot Credits.

Second: credits are incurred for LLM tokens, for tools including knowledge and MCP servers, and for the harness itself. Any experience that uses one of those consumes credits.

Third, and this is the sentence that upsets most plans: billing starts when you build. Unlike the standard harness, which only bills after publishing, the GitHub Copilot harness counts from the moment you start. Creating a solution in natural language, testing in preview, generating and running evaluations: all of it consumes credits.

This applies exclusively to agents, workflows and apps on the GitHub Copilot harness. The standard harness and agent flows continue to bill under the existing licence model.

The four cost drivers

Every task an agent completes assembles its cost from the same four components:

1. Model tokens. Input and output. A skill loaded on every request even though it rarely applies is directly visible here.
2. Retrieved organisational context. Knowledge sources, grounding, Microsoft IQ. Context is why agents are useful, and at the same time a cost item per request.
3. Tool calls. Connectors, APIs, MCP servers. An agent that calls the same tool three times out of uncertainty costs three times.
4. Orchestration runtime. The harness itself. Reasoning about the sequence of steps is not a by-product, it is a billed operation.

From that follows a rule of thumb I give customers: cost depends less on volume than on design. A badly scoped agent with fifty requests a day can be more expensive than a cleanly scoped one with five hundred.

Where you get real numbers

Do not estimate, measure. There are four places:

• The agent's monitor page in Copilot Studio. Consumption per agent. The first place an outlier shows up.
• Power Platform Admin Center, under Licensing and Copilot Studio. Consumption across an environment or the whole tenant.
• Agent usage estimator. Available since April 2026, it forecasts credit consumption across Copilot Studio and Dynamics 365 agents before you go broad.
• Enforcement policy. Describes what happens when the allocated credits run out. You want to have read that before the rollout, not after.

Credits are managed in the Power Platform Admin Center and allocated to environments. That allocation is your most important control: it caps the damage when an agent gets stuck in a loop or a maker spends a night generating variants.

The small print

Two points where I have to disclose the source situation, because they come from secondary sources and I could not verify them in the official documentation.

The first concerns the conversion rate. On the pay-as-you-go plan, one Copilot Credit is said to correspond to one US cent. That figure circulates in several analyses of the August 2026 licensing guide. It is the only figure that lets you turn a credit estimate into an amount a CFO can read. Check it against the current licensing guide before you calculate with it.

The second concerns Work IQ. Since July 2026, use of the Work IQ API by Copilot Studio agents is said to no longer be included in the Microsoft 365 Copilot licence. If that holds, it is expensive, because Work IQ is precisely the context layer that makes agents useful. Same rule here: verify against the licensing guide before you build an architecture on it.

That I could not confirm either point on Microsoft Learn is worth noting. Commercial detail is migrating into PDFs that get updated more often than the documentation does.

Checklist before rollout

1. Choose the harness deliberately. Does the agent genuinely need multi-step reasoning? A FAQ bot does not, and is cheaper and more predictable on the standard harness.
2. Allocate credits per environment before the first maker gets access. Not after.
3. Read the enforcement policy and communicate internally what happens on exhaustion.
4. Run a forecast with the estimator, using realistic task volumes rather than the best case.
5. Measure two weeks of pilot operation and hold the forecast against actual consumption. The ratio between them is the number you need for every subsequent agent.
6. Scope skills and knowledge cleanly. What is rarely needed does not belong in the system instruction but in a skill with a precise description.
7. Check tool calls in the monitor. Repeated identical calls are a design fault with a price tag.
8. Tell the makers that building costs. That is not a technical measure, it is the most important one. Nothing damages trust in a platform as reliably as a surprise invoice.

My assessment

Consumption-based billing is more honest than a flat rate. It reflects the fact that an agent working through invoice intake consumes compute, and a lot of it. Hiding that only moves the problem.

The sore point is billing from the first build step. Low code lives on people trying things out. If trying costs money, fewer people try, and the ideas that eventually reach production come from IT again. That was not the point of the exercise.

My practical answer: a sandbox environment with a small fixed credit allocation, where makers can experiment without anyone getting nervous. That costs a manageable amount per month and is cheaper than a platform nobody touches any more.

Sources

• Overview of billing for agents powered by the GitHub Copilot harness, Microsoft Learn, as of 3 August 2026
• Choose a harness, Microsoft Learn, as of 3 August 2026
• What's new in Copilot Studio, Microsoft Learn, on the agent usage estimator
• Copilot Studio: GitHub Copilot harness GA and licensing, August 2026. Secondary source for the credit rate and the Work IQ exception, not verified on Microsoft Learn.