Guide Microsoft AI/SI · Chapter 7

    Which levers move the AI bill at Microsoft.

    Copilot runs on licences per person, agents on credits, Foundry on tokens and an index. Here you see which lever moves which meter; the amounts are in the articles.

    As of: October 2026 · Editor: Marcel Haas

    • Copilot licence
    • Copilot Credits
    • Azure OpenAI
    • Microsoft Foundry
    • Azure AI Search
    • Azure Cost Management

    Cost levers

    Which meter, which lever.

    For every meter that moves the bill: what drives it and which lever you hold. Mechanics without figures; the amounts are in the proposal.

    1. 01

      Copilot licence per person

      What moves the meter: The number of people with a licence; Copilot Chat with web results is included in eligible subscriptions, answers from work data need the licence, Copilot Business the Business plans.

      Your lever: Assign role by role after the pilot, through groups; people who rarely work with email, documents and meetings often manage with Copilot Chat.

    2. 02

      Copilot Credits

      What moves the meter: Every answer, tool call and trigger run of an agent from Copilot Studio, advanced work in SharePoint, tools of the Dataverse MCP server outside Copilot Studio and computer use; people without a licence move the meter in full.

      Your lever: The monthly cap per agent in the Power Platform admin center, working hours for triggers, a question back that stops a loop, and internal use by licensed people within the fair-use limits.

    3. 03

      Tokens in Foundry

      What moves the meter: Input, output, cache and thinking steps per model and deployment type; the size of the context per call, reasoning models and tools stuck in a loop.

      Your lever: The model per step, a small one for classification and pre-sorting, instructions in the cache, quotas per deployment and a budget with an alert per application.

    4. 04

      Reserved capacity or pay-as-you-go

      What moves the meter: Reserved units cost per hour, used or not; pay-as-you-go counts every token, with peaks and lulls.

      Your lever: Pay-as-you-go until the pilot shows the load profile; reserved capacity only for an even, high base load, the peaks stay pay-as-you-go.

    5. 05

      Model Router

      What moves the meter: The share of simple requests that go to small models; every request may be answered by a different model from the approved set.

      Your lever: A set made only of models in the permitted place, the same test set as for a fixed model, and the router where many simple requests arrive.

    6. 06

      Search index in Azure AI Search

      What moves the meter: Tier, partitions and hours, around the clock, whether anyone asks or not; plus the tokens for embeddings with every indexing run.

      Your lever: The tier by data volume, a schedule for re-indexing, and Microsoft Graph with Copilot connectors first where the agent lives in Microsoft 365.

    7. 07

      Fabric capacity

      What moves the meter: The size of the capacity from F2, which Azure bills per second while it runs; data agents and Fabric IQ compute on it.

      Your lever: The size by the load of the data agents, measured in the pilot, and from F64 people with a free licence can read along.

    8. 08

      Traces and logs

      What moves the meter: The volume per run in Application Insights, with input, tool calls and answer, and the retention; the audit log in Purview follows the licence tier.

      Your lever: A sample rather than every run in full text, sensitive content in a table of its own, and a retention that follows your rule.

    9. 09

      Licences of the control layer

      What moves the meter: Conditional Access and Identity Protection for agents, risk signals in the agent list and the Purview coverage of Foundry agents require Microsoft 365 E7 or Agent 365; data loss prevention rules in Copilot E5 or the Purview Suite.

      Your lever: Record per agent in the risk sheet which of these features it needs; the identity in Entra Agent ID is open to every Entra customer.

    As of October 2026, derived from the dated facts of this chapter and the anatomies of the technology pages. Token meters, reserved capacity, the Model Router, the search index and the logs name the mechanism without a dated Stand; the amounts are in the dated cost articles and in the proposal, never on this page.

    01

    Licence, credits and tokens run as separate meters.

    At the end of the month the head of finance receives three documents that all have to do with AI. The Microsoft 365 invoice lists the Copilot licences, a fixed amount per person and month. The Power Platform admin center shows the agents' consumption in Copilot Credits, which grows with every answer and every tool call. And the Azure invoice lists the tokens of the models in Microsoft Foundry, next to the search index, the compute of the hosted agents and the logs. A licence is a fixed amount per person, credits and tokens move with usage, an index runs by tier and hours even when nobody asks. Whoever wants to understand the bill therefore first asks which of these meters runs for which work, and only then how high it stands.

    02

    The Copilot licence applies per person and decides what Copilot fetches itself.

    Copilot Chat with internet results comes at no extra cost with eligible Microsoft 365 subscriptions; whoever wants answers from emails, files and meetings needs the Copilot licence, which fetches work data through Microsoft Graph. For the Business plans Microsoft offers the Copilot Business variant, and Microsoft 365 E7 includes Copilot. In SharePoint the everyday features belong to the licence, large-scale work such as reviewing whole libraries is billed in Copilot Credits and needs a spending rule in the admin center. The lever lies in the assignment: licences apply per person, you grant them through groups, and so it stays traceable who has worked with which tier since when. People who rarely work with emails, documents and meetings often manage with Copilot Chat; people who spend the whole day in them get the value of the licence out day after day. For the pilot the group is enough, after that you decide role by role with the figures from the Copilot Dashboard.

    03

    Agents are billed in Copilot Credits, with a cap per agent.

    An agent from Copilot Studio consumes credits for every answer, every tool call and every trigger run, through prepaid packs or pay-as-you-go through an Azure subscription. Internal use by people with a Copilot licence is included within the fair-use limits, computer use excepted; people without a licence, visitors on the website and agents that work on a trigger at night move the meter in full. The tools of the Dataverse MCP server are also charged in credits when agents outside Copilot Studio use them. What drives the meter: the number of people without a licence, triggers that check every few minutes whether something is pending, tools stuck in a loop, and long answers from large knowledge sources. What you hold in your hand: a monthly cap per agent in the Power Platform admin center, working hours for triggers, a question back that stops a loop, and the decision who gets a licence and who works through credits. In the pilot we read the consumption per agent before it is opened to everyone.

    04

    In Foundry every token counts, by model and deployment type.

    A model in Microsoft Foundry is billed by tokens: input with instruction, history and documents, output with the answer, for reasoning models also the thinking steps, and repeated input from the cache more cheaply. The rate depends on the model and on the deployment type with which the models chapter sets the place: regional, in the data zone or global, plus batch for jobs without hurry. Pay-as-you-go you pay per token; reserved capacity you pay per unit and hour, used or not, and it pays off only with an even, high load that the pilot has measured. The Model Router sends simple requests to small models and hard ones to large models and saves tokens where many simple requests arrive. Quotas per deployment cap consumption, the Control Plane shows cost and token usage per agent and model, and Azure Cost Management keeps a budget with an alert per application. The levers: the model per step, a small one for classification and pre-sorting, the size of the context per call, instructions in the cache, the router for the volume, and the deployment type.

    05

    The search index and the logs are the items that are easy to forget.

    An application that answers with references needs an index in Azure AI Search, and this service runs around the clock: it is billed by tier, partitions and hours, whether someone asks on a Sunday or not, and every indexing run consumes tokens for the embeddings, again with every re-index. Traces in Application Insights grow with every run in which input, tool calls and answer are stored; retention and sampling decide the size of the item. Data agents in Fabric need a capacity from F2, which Azure bills per second while it runs. And some features of the control layer from chapter 8 hang on licences: data loss prevention rules in Copilot require E5 or the Purview Suite, Conditional Access and Identity Protection for agents and the Purview coverage of Foundry agents a licence for Microsoft 365 E7 or Agent 365. The levers: an index tier sized to the data volume, Microsoft Graph and Copilot connectors first where the agent lives in Microsoft 365, a schedule for re-indexing, a sample for traces, and a list of which agents need the licensed security features.

    06

    Cost per case is the figure that counts, and the pilot delivers it.

    One task, one pilot group, real cases, and at the end a sheet with one figure per case: licence share, credits, tokens, index, compute and the hours of the person who maintains the agent, divided by the cases it has handled. You set this figure against what the case costs today, and from that comes the decision whether the agent gets more people. Before that come the caps: a budget with an alert per application, a cap per agent, quotas per deployment. The proposal names the frame for build and operation, the dated cost articles name the list prices with their Stand, and we check both again every quarter because Microsoft changes prices and counting rules. What the sheet per application records:

    • Meters: which licence, which credits, which tokens, which index and which compute run for this application, with the Stand at the vendor.
    • Drivers: what moves the meter, from the number of people through triggers and context size to re-indexing.
    • Cap: the amount per month, the alert before it and the person who receives it and acts.
    • Cost per case: the value from the pilot, compared with today's workflow, and the date of the next review.

    What you decide.

    A licence for everyone, or a licence for some and credits for the rest?

    Options

    1. 1A Copilot licence for every person
    2. 2Licences for the roles that work with emails, documents and meetings daily, Copilot Chat for the others, agents through credits
    3. 3A pilot group, then role by role by the figures

    Our recommendation: The pilot group first, then role by role. Whoever spends the day in emails, documents and meetings gets the licence, and for these people the internal use of the agents is included. Whoever rarely works with them manages with Copilot Chat and uses agents through credits with a cap. The Copilot Dashboard and the consumption reports deliver the figures for every further group.

    What we advise against: Licences by headcount before the pilot. Equally, saving the licence on people who work with agents all day: their use then runs through credits, without the fair-use inclusion of the licence.

    Reserved capacity or pay-as-you-go in Foundry?

    Options

    1. 1Pay-as-you-go per token
    2. 2Reserved capacity per deployment
    3. 3Reserved capacity for the base load, pay-as-you-go for the peaks

    Our recommendation: Pay-as-you-go until the pilot has shown the load profile. Reserved capacity pays off with an even, high load with demands on response time, and only once the tokens per day are known from operation; then it carries the base load, and the peaks keep running pay-as-you-go. The decision is recorded in the application's sheet and reviewed against consumption every quarter.

    What we advise against: Reserving capacity before anyone has measured the load. A reserved unit costs at night and at the weekend too.

    Where do we set the cap: per agent, per application or per subscription?

    Options

    1. 1A cap per agent in the Power Platform admin center
    2. 2A budget with an alert per application in Azure Cost Management and quotas per deployment
    3. 3One budget for the whole subscription

    Our recommendation: Per agent and per application, because there a person knows the meter and can act: in Copilot Studio the monthly cap per agent, in Foundry a budget with an alert per application and quotas per deployment. A budget for the subscription comes on top as a safety net, and the alert goes to the person named in the sheet.

    What we advise against: A single budget for everything. When it fires, nobody knows which agent it was, and the alert arrives when the month is over.

    An index of your own or Microsoft Graph?

    Options

    1. 1Knowledge through Microsoft Graph and Copilot connectors, without an index of your own
    2. 2An index of your own in Azure AI Search for the application in Foundry
    3. 3Both, separated by application

    Our recommendation: Microsoft Graph first where the agent lives in Microsoft 365: synced connectors index line-of-business systems without indexing cost, and Copilot and Copilot Studio answer with the rights of the person asking. An index of your own pays off when the application runs outside Microsoft 365, needs its own chunking and ranking or bundles sources that no connector reaches; then the tier is chosen by data volume and re-indexing is scheduled.

    What we advise against: Building an index because you might need it later. It runs around the clock from the first day.

    Status and sources.

    As of: October 2026

    • Licensing as of 25 September 2026

      Web chat is included, chat with work data needs the licence.

      Copilot Chat with internet results comes at no extra cost with eligible Microsoft 365 subscriptions. Answers from your work data need a Copilot licence. For smaller organisations, Microsoft offers a «Copilot Business» variant for the Business plans.

      Source: Microsoft Learn: License options for Microsoft Copilot
    • Billing as of 14 September 2026

      Copilot Credits, included for licensed employees.

      Agents are billed in Copilot Credits, through prepaid packs or pay-as-you-go. Internal use by people with a Microsoft Copilot licence is included within fair-use limits, computer use excepted. In the Power Platform admin centre a monthly cap can be set per agent. This applies to the standard harness.

      Source: Microsoft Learn: Billing rates and management
    • From 30 September 2026

      Copilot in SharePoint becomes generally available.

      Everyday features are part of the Microsoft Copilot licence; "Advanced work" consumes Copilot Credits and needs a spending rule in the admin center. The PowerShell control from the preview applies until 1 November 2026; after that, Site AI settings and Restricted Content Discovery take over.

      Source: Microsoft Learn: Copilot in SharePoint
    • As of 1 June 2026

      Copilot connectors: synced or federated, with different licences.

      Synced connectors index line-of-business systems into Microsoft Graph and can be used with any Microsoft 365 licence without indexing cost; federated connectors query the system live through MCP. Federated connectors and use in Copilot need the Microsoft Copilot licence or E7; for agents, a Copilot Studio licence or pay-as-you-go covers synced connectors.

      Source: Microsoft Learn: Copilot connectors prerequisites
    • As of 5 June 2026

      Agents reach Dataverse through an MCP server.

      For Copilot Studio, the Dataverse MCP server is switched on in every environment; admins enable further clients such as GitHub Copilot or Claude one by one. Since 15 December 2025, the tools are charged in Copilot Credits when agents outside Copilot Studio use them.

      Source: Microsoft Learn: Connect to Dataverse with MCP
    • As of 15 June 2026

      From F64, people with a free licence can read along.

      Fabric capacities range from F2 to F8192 and are billed through Azure per second, with a one-minute minimum. Below F64, every person reading Power BI content needs a Pro or Premium Per User licence; from F64 the free licence with the Viewer role is enough. Premium Per User does not create a Fabric capacity.

      Source: Microsoft Learn: Microsoft Fabric licenses and capacity
    • As of August 2026

      Purview for agents on Microsoft Foundry comes with E7 and Agent 365.

      According to its licensing description, Microsoft intends to cover Purview's protection and compliance capabilities for agents on Microsoft Foundry and for AI apps connected to Entra ID with Microsoft 365 E7 and the Agent 365 licences. Before every agent project, we check which licence your agents need.

      Source: Microsoft Learn: Microsoft Purview service description

    Frequently asked questions.

    Do we pay credits even though everyone has a Copilot licence?

    Not for internal use by licensed people within the fair-use limits, computer use excepted. Credits arise when people without a licence or visitors on the website use the agent, when it works on a trigger without a person, when it calls tools of the Dataverse MCP server outside Copilot Studio, or when Copilot in SharePoint takes on large-scale work. The consumption report in the admin center shows per agent where the credits come from.

    What does an agent cost that runs at night and answers nobody?

    Every trigger run consumes credits or tokens, even when nothing is pending: an agent that checks a mailbox every few minutes pays for every check. That is why every trigger gets working hours, a condition that ends the run when nothing new is there, and a cap per month. In the trace you see how many runs were empty.

    Why is the search index on the bill although nobody asked?

    An index in Azure AI Search is a service that runs as long as it exists: billed by tier, partitions and hours, on Sunday as on Monday. Added to that are the tokens for the embeddings with every indexing run. The tier is therefore chosen by data volume, re-indexing is scheduled, and where the agent lives in Microsoft 365, Microsoft Graph comes first, which needs no index of its own.

    When is reserved capacity in Foundry worth it?

    When the pilot has shown an even, high load with demands on response time and the tokens per day are known. Reserved capacity is paid per unit and hour, used or not; pay-as-you-go you pay per token and carry peaks and lulls with it. Many applications stay pay-as-you-go, and whoever switches keeps running the peaks pay-as-you-go.

    How do we set a cap before the first agent runs?

    In Copilot Studio with the monthly cap per agent in the Power Platform admin center, in Foundry with quotas per deployment and a budget with an alert per application in Azure Cost Management, deliberately tight for the first application. The alert goes to the person named in the application's sheet. After the pilot the limit is set again with the measured consumption.

    Which licence do the security features for agents need?

    The identity per agent in Entra Agent ID is available to all Microsoft Entra customers. Conditional Access and Identity Protection for agents, the risk signals in the agent list and the coverage of Foundry agents in Purview require a licence for Microsoft 365 E7 or Agent 365, data loss prevention rules in Copilot E5 or the Purview Suite. Which agents need these features is recorded in the risk sheet from chapter 8.

    Where do we find the current prices?

    In the dated cost articles below, which name the list prices with their Stand, in Microsoft's price list and in the proposal, which records the frame for your application. This page explains the mechanics and therefore stays without amounts; we check prices and counting rules every quarter, and the date of the last check is in the article.

    Next chapter · Chapter 8

    How agents and models at Microsoft stay under your control.

    How do agents and models at Microsoft stay under our control?