Content reviewed: 13 October 2026

Separate chat access, data access and the billing of each feature. A Copilot label in the interface does not establish all three.

Microsoft is transitioning documentation to Microsoft Copilot and Microsoft Copilot Chat. Some interfaces and licences still use Microsoft 365 Copilot. Identify the exact product and SKU when ordering.

What basic chat provides

Copilot Chat supports conversation with AI and web context. Organizational content access depends on the experience and licence. Supplying a document is different from access to the full work context.

The Copilot Chat FAQ distinguishes chat from the licensed work experience. The privacy and protections guide describes additional ways to supply organizational content.

In the pilot, test an actual task: summarize an attached document, retrieve information from team sources or perform another required operation. Record which app, identity and source produced the result.

Depending on the experience, you can supply a specific file or use open content; Microsoft also documents limited personal context in Outlook. This is not automatically full Work IQ context with semantic indexing. Record the application, how content is supplied and the entitlement used in your test. Microsoft distinguishes these access paths. An internal document appearing in a response does not establish a paid entitlement or unrestricted access.

Data protection and feature scope

Work login and enterprise data protection are important. But they do not mean that every available AI function has the same access to data or the same cost.

Microsoft documents enterprise data protection, with controls that depend on the underlying subscription. Verify the work identity and intended usage.

Prepare a clear rule for users: what account they use, what information they are allowed to enter and who checks the result. The license approval itself does not replace the rules of working with internal data.

The agent model matters

An agent can reply from public pages, read shared company resources, or perform other actions. Each case needs its own assessment.

Copilot Chat agent guidance distinguishes declarative agents using instructions and public websites from metered agents using shared tenant data. Apply the rules to the actual agent, audience and data source.

Do not automatically transfer the conditions of a simple internal agent to autonomous processing or an external audience. Also note the engine, extensions used and services outside of Copilot itself.

Deployment checklist

Who uses the agent and what is billed
SituationLicensing and billing conditionPilot evidence
Declarative Copilot Chat agent using only instructions and public websitesNo additional cost in the documented modelActual agent type, sources and availability settings
Copilot Chat without Copilot USL, agent using shared tenant dataMetered usage and supported subscription/billing setupTarget users, linked service and consumption
Licensed internal user, standard agent in Microsoft 365Usage under the authenticated user identity can be included, subject to conditions and fair useChannel, user licence and actual execution identity
Agent flow called by a licensed user through an agentInclusion requires When an agent calls the flowFlow type and actual trigger, not just its owner
Computer use or GitHub Copilot harness agentSeparate consumption not covered by ordinary Copilot USLEngine, tools and consumption during building/testing

Use the table to classify the deployment. It does not establish a price or an entitlement independently of its stated prerequisites.

The first two cases are described in Copilot Chat agent documentation. The inclusion for standard agents in Microsoft 365 and computer-use exceptions are specified in the current Copilot Studio overview. Agent flows also depend on the trigger and identity footnote. Do not apply this table to anonymous website visitors or another channel without checking its conditions.

Licences and consumption billing

The current USL and UBB overview separates everyday tasks from selected advanced AI experiences. Its UBB experiences require a USL for access and a billing policy for consumption. This differs from a PAYG Copilot Chat agent for someone without the paid Copilot add-on. Credits are not a universal licence replacement. Continuing with UBB after certain USL limits is labelled coming soon; do not assume it is already available in every tenant.

Budget against the proposed task and feature. Do not assume that one paid Copilot subscription includes every engine, trigger and downstream service.

The GitHub Copilot harness can consume credits during creation, preview, testing and evaluation generation. An ordinary Copilot user licence does not cover these activities. Its separate billing overview defines the scope. Any statement that testing is free must identify the specific feature and environment, rather than the whole platform.

Separate user subscriptions, billed consumption and related services. Assign an owner to each cost line so that usage in another billing account is not missed.

Licensing does not grant source access

For SharePoint agents, distinguish agent availability, .agent file permissions and source-data permissions. PAYG access without a Copilot licence also requires membership of the relevant billing-policy group. Joining that group does not grant access to a restricted document. Microsoft documents these access layers. If an agent works for its maker but not an employee, check every layer rather than sharing an entire library as a fix.

Before approving another tool, review the terms of the particular agent and downstream services. Enterprise data protection for chat does not prove that every extension has identical terms. Microsoft requires checking agent terms.

Verification of the assistant's answer according to the work document

Pilot for internal policies

A hypothetical company builds an agent to answer employee questions from approved internal policies. Its purpose is to find the applicable rule faster.

Before deployment, the team selects approved resources, prepares common questions, and verifies permissions. It divides the tests according to job roles. The agent must not derive the correct answer from a document to which the employee is not supposed to have access.

Record the licence and billing model for the intended audience. Testing only the maker’s account is insufficient when employees have different licences and permissions.

Calculate usage for a mixed group

Our hypothetical model has 20 employees with a paid Copilot licence and 30 with eligible Copilot Chat and access to a metered agent. Everyone uses a standard internal agent in Microsoft 365 under their own authenticated identity and asks three questions daily. Assume each question produces exactly one generative answer and one tenant graph grounding event, without actions, a reasoning model or other billed tools.

The current rates for these two events are 2 and 10 Copilot Credits. For the 30 users without Copilot USL, that gives 30 × 3 × (2 + 10) = 1,080 credits daily. Under the stated conditions, the 20 licensed users do not generate additional billed consumption for those included features. Check the rates and inclusion conditions before applying this model.

The 1,080 figure is neither a price nor measured customer usage nor a universal per-question rate. Additional actions or another engine change the calculation. After the pilot, compare billed events with completed tasks, including retries. Our Copilot Studio article covers capacity and PAYG selection in more detail.

Verify agent responses

Prepare the right documents for known questions, conflicting documents, and out-of-scope questions. Test if the agent recognizes when it doesn't have a reliable answer.

For directives and licences, the output needs a date, a clear source and the right context. A naturally written answer may be wrong. Therefore, the verification must not depend on the impression of the fluency of the text.

For each failure, distinguish missing sources, incorrect sharing, outdated rules and misuse of evidence. Fixing the cause is more useful than another instruction to answer correctly.

Pilot measurements

  • Completed work assignments and correctness of answers
  • The need for human corrections
  • Availability and quality of cited documents
  • Results under different user roles
  • Real consumption and its allocation
  • Behavior when resource is unavailable or conflicting

The number of conversations alone does not demonstrate business benefit. What matters is whether people manage to complete a specific job correctly.

Cost boundaries

Assign a billing owner and approved usage scope before expanding. Distinguish notifications, prepaid allocation and actual stopping controls; test the control you depend on.

Cost controls in Copilot Studio
ControlWhat it limitsWhat it does not do
Environment credit allocationReserved prepaid capacityDoes not by itself stop tenant-pool or PAYG consumption
Disable Draw from tenant capacityDrawing unallocated prepaid creditsDoes not stop consumption through linked PAYG
Agent limit with Stop usageThat agent’s consumption at its limitNot an aggregate limit for every agent in the environment
Azure budget and alertCost notificationsDoes not automatically stop agents or billing

These controls apply to Copilot Studio; do not assume the same interface for every Copilot Chat or SharePoint service. In PPAC, check Licensing → Copilot Studio and the available agent settings. If a limit must stop operations, verify Stop usage and test the result. Assign a fallback process and owner for an interrupted workflow. Microsoft distinguishes allocation, PAYG and agent stopping.

A Copilot Studio agent flow is different from an ordinary Power Automate cloud flow. When prepaid capacity is exhausted, new agent-flow runs can be blocked while non-flow interactions remain available; runs already in progress finish. Included licensed-user usage has an exemption subject to the relevant trigger. A working conversation therefore does not prove completion of its business action. Microsoft explains the different enforcement behaviour.

Separate pilot and production budgets. Testing and extensions can have different consumption patterns from routine employee use.

Re-assess before any significant expansion. A new data source or autonomous trigger can also change the licensing and financial model.

Frequently asked questions

Is Copilot Chat free for every AI task?

No. Basic chat or an included declarative agent does not establish rights to every tool. Check the particular licensing and consumption conditions for shared data, computer use and other engines.

Do I need paid Copilot to upload a document manually?

Attaching a document alone does not establish a need for Copilot USL. Check the available feature and its limits; supplied content is different from full Work IQ context.

Does a user licence cover every autonomous run?

There is no universal inclusion. Check engine, channel, authenticated identity and trigger. Included agent-flow usage specifically requires When an agent calls the flow; other triggers have different treatment.

Is an agent ready if it answers its maker correctly?

No. Test target users, their licences and permissions, failure cases and billed events. The maker’s account can hide access or consumption issues for the future audience.

Is a paid Copilot licence enough when the agent controls a computer?

No. Computer use is outside ordinary Copilot USL inclusion. Check metered consumption, the engine and target application rights.

Why can an agent answer but fail to start its flow?

Copilot Studio agent-flow enforcement can block new runs when prepaid capacity is exhausted while non-flow conversations continue. Check the trigger, included entitlement, data and the actual error too.

Document the proposal

Document the task, product, users, data sources, permissions and billing model. This allows benefits and costs to be compared against the deployment rather than similar product names.