Agent Compute Units, or ACUs, are a way of measuring how much work an AI agent does. Instead of charging by the hour like a human employee, or by a flat monthly fee like traditional software, an ACU-based model meters the actual computational effort an agent spends completing tasks. One ACU represents a standard chunk of that effort, so the more work an agent performs, the more ACUs it consumes.

A useful analogy is the electricity meter in your office. You do not pay a fixed amount regardless of usage, and you do not pay per employee. You pay for the kilowatt-hours you actually use. ACUs work the same way for AI agents. A quiet week where the agent processes fifty invoices costs less than a busy week where it processes five thousand. Usage and cost move together.

For a business, ACUs make AI spending predictable and fair. You are billed for outcomes and effort rather than for seats you may not fully use. It also makes it easy to compare the cost of automating a process against doing it manually, because you can see exactly how much compute a given workflow consumes. The main thing to understand is what drives ACU consumption, such as task complexity, the number of steps involved, and how often the agent runs, so you can forecast and manage your costs with confidence.

Frequently Asked Questions

How are ACUs different from a per-seat software subscription?

A per-seat subscription charges a fixed price for each user, whether they use the software heavily or barely at all. ACUs charge based on actual work done. If your agent has a light month, you spend less, which aligns cost directly with the value you receive.

What actually consumes an ACU?

Consumption is driven by the computational effort behind a task. A simple, one-step action uses fewer ACUs than a complex task that involves reading documents, reasoning across many steps, calling several systems, and handling exceptions. More work means more ACUs.

How do ACUs compare to the cost of a human doing the same work?

ACUs let you make a direct comparison. For example, if processing one invoice manually takes an employee ten minutes, you can compare that labor cost against the ACUs the agent consumes to do the same task. This makes the return on automation easy to quantify.

Can ACU costs be predicted in advance?

Yes, with reasonable accuracy. Because consumption tracks task type and volume, you can estimate monthly costs from your expected workload. A process that runs a known number of times per day produces a fairly stable ACU forecast.

What if costs run higher than expected?

This is a common concern with usage-based pricing, since a spike in volume can mean a spike in spend.

Zamp addresses this with dashboard visibility that shows process activity and health at a glance, and activity logs that record every action an agent takes, so you can see exactly where ACUs are being spent. Structured processes keep each agent focused on its specific job rather than wandering into unnecessary work, and approval checkpoints let you gate high-volume or sensitive runs, so usage stays aligned with what you actually intended.

Do unused ACUs roll over or expire?

This depends on the specific plan and agreement. Some arrangements include a monthly allotment, others are pure pay-as-you-go. The key is to confirm the terms up front so your finance team knows how budgeting and any rollover work.