AI & CopilotCopilot Studio pricing explained for UK organisations

Copilot Studio is billed on consumption rather than per user: each time an agent answers a question or takes an action, it uses Copilot Credits, which you buy in prepaid monthly capacity packs or pay for as you go through Azure. The monthly cost therefore depends on how many conversations your agent handles and what kind of work each one involves, not on how many staff you have. This article explains the moving parts and how to estimate a budget before you build.

At the time of writing (October 2026). Microsoft has renamed and repriced Copilot Studio several times, including renaming messages to Copilot Credits. Rates and packs below are indicative and quoted from Microsoft list pricing where available. GBP prices depend on your agreement. Always check current Microsoft pricing and the official billing rates table before committing.

How Copilot Studio billing works

There are three ideas to understand.

Copilot Credits are the unit of consumption. Every agent response or action consumes a number of credits. Different types of work cost different amounts.

Capacity is bought at tenant level. You buy credits for the organisation, then allocate them to environments where agents run. Individual users do not need a Copilot Studio licence to chat with a published agent.

Makers need access to build. People who build agents need appropriate access, which Microsoft provides through Copilot Studio user licences or related entitlements. The licence for makers is not the main cost driver; consumption is.

The purchasing options

Unused prepaid credits do not usually roll over month to month, so buying too much capacity wastes money. Pay-as-you-go avoids that but is harder to budget. Many organisations start on pay-as-you-go during a pilot, then move to packs once usage is known.

What uses more credits and what uses fewer

Microsoft publishes a billing rates table that sets how many credits each kind of event consumes. The exact weights change, so treat this as relative guidance:

• Scripted replies from a topic you authored use the fewest credits.
• Generative answers grounded in documents or websites use more.
• Grounding on organisational data through Microsoft Graph can use more again.
• Actions such as calling a connector, running a flow or updating a record add credits.
• Autonomous agents that run on triggers, for example when an email arrives, consume credits for every run even though no one is chatting.

The common surprise: an agent that runs automatically on every incoming email or new record can consume more credits than a busy chat agent, because it runs whether or not anyone needs it. Filter triggers carefully.

How it relates to Microsoft 365 Copilot

These are separate products that are often confused.

Agents built in Copilot Studio can be published into Microsoft 365 Copilot, and at the time of writing some usage by Microsoft 365 Copilot licence holders is covered by their licence. The detail changes, so check the current terms before assuming coverage. If you are deciding whether you need one, both or neither, our Microsoft 365 Copilot page sets out the differences.

Estimating monthly cost

A rough estimate needs five numbers.

• Users: how many people are likely to use the agent in a month.
• Conversations per user: how often each of them asks something.
• Turns per conversation: questions and answers in a typical exchange.
• Response mix: the share that are scripted, generative or involve actions.
• Autonomous runs: how many trigger-based runs happen per month.

Multiply users by conversations by turns, weight by the credit rate for each response type, and add autonomous runs. Then add a margin of 25 to 50 percent for the first few months while behaviour settles.

An illustrative example

These numbers are assumptions, not data from a real deployment. A UK organisation of 250 staff publishes an HR policy agent in Teams. Suppose 100 people use it in a month, asking an average of three questions each, mostly answered generatively from a SharePoint policy library. That is about 300 generative answers a month, which at the credit rates published at the time of writing sits comfortably inside a single capacity pack. Add an action that logs a ticket for unanswered questions and the figure rises, but usually not dramatically.

The point of the exercise is not precision. It is to find out whether you are talking about one pack or ten before you build. Our AI agent for policies use case describes this kind of agent.

Other costs to include

• Build and testing. A useful agent needs well-structured content, tested topics and clear escalation routes. Our Copilot Studio service covers scoping and build.
• Content preparation. Out-of-date or contradictory policies produce bad answers. Budget time to tidy the source content.
• Connectors. Agents that act in systems through premium connectors can bring Power Platform licensing into scope.
• Governance. Decide who may build agents, which data sources are allowed and how agents are reviewed before publishing. See governance.
• AI Builder or document processing, if the agent reads invoices or forms. Our invoice and document processing use case explains the pattern.

How to keep Copilot Studio costs under control

• Start with a pilot on pay-as-you-go and measure real consumption for a month.
• Use scripted topics for the most frequent, simple questions.
• Restrict generative answers to curated knowledge sources.
• Filter autonomous triggers so the agent only runs when it needs to.
• Set capacity allocation per environment so one experiment cannot drain the tenant's credits.
• Review analytics monthly: unanswered questions, abandoned conversations and cost per resolved query.

Checklist before you buy capacity

• Have we defined the one job this agent does?
• Do we know roughly how many users and conversations per month?
• Have we identified scripted, generative and action-based responses?
• Are there any autonomous triggers, and how often will they fire?
• Do the knowledge sources have correct SharePoint permissions?
• Do any actions use premium connectors?
• Have we checked current credit rates and GBP pack prices?
• Is there a named owner for content and monthly review?

Frequently asked questions

Do all staff need a licence to use a Copilot Studio agent?

Generally no. Users chatting with a published agent consume tenant credits rather than needing their own Copilot Studio licence. Makers need access to build, and actions in premium systems can bring other licensing into scope.

Is Copilot Studio included with Microsoft 365 Copilot?

Some use is included for Microsoft 365 Copilot licence holders in specific scenarios at the time of writing. Organisation-wide agents, autonomous agents and external audiences usually need Copilot Studio capacity. Check current terms.

What is the difference between messages and Copilot Credits?

Microsoft renamed the billing unit. Older documentation and quotes refer to messages; current pricing refers to Copilot Credits. Check the conversion when comparing old and new quotes.

Can unused credits roll over?

Prepaid monthly capacity generally does not roll over. That is why measuring usage during a pay-as-you-go pilot before buying packs is worthwhile.

How accurate can a cost estimate be before launch?

Within a reasonable range, if you know users, conversation patterns and response types. Expect to adjust after the first month of real use.

Key takeaways

• Copilot Studio is priced on consumption through Copilot Credits, not per user.
• Generative answers, actions and autonomous runs cost more than scripted replies.
• Microsoft 365 Copilot and Copilot Studio are different products with different licensing.
• Pilot on pay-as-you-go, measure, then buy capacity packs.
• Content quality and permissions matter as much as price. If you want help scoping an agent, see our packages or book a call below.