Agent Builder or Copilot Studio: Where Should Your First Agent Live?

When a team starts building internal AI assistants in Microsoft 365, they immediately run into two different places to do the work: Agent Builder in Microsoft 365 Copilot (Microsoft's in-app builder, not to be confused with the similarly named agent builders from OpenAI or Google Vertex AI) and Microsoft Copilot Studio. Because both sit inside the same Microsoft ecosystem, it is not always obvious where your first agent belongs.

Start in Agent Builder when your agent only needs instructions and read-only access to content people can already view in Microsoft 365, such as a plan-my-day briefing over your Outlook inbox and calendar or help with an open-ended task like writing and meeting prep. We lean into Copilot Studio when the agent has strict guardrails it must stick to, needs automated evaluations to prove its output is reliable, belongs inside a managed Power Platform Solution, or runs automation logic like looking up and adding records to a SharePoint list. The split comes down to capabilities rather than headcount: Agent Builder can share with specific people or groups, support multiple co-owners, and publish to your organization catalog, and Microsoft notes that serving users across several departments does not by itself require Copilot Studio.

How the two Microsoft 365 agent tools compare at a glance

Capabilities, limits, and credit rules below were checked on November 13, 2026, against Microsoft's architecture guide for choosing between the two experiences and Microsoft's official studio pricing page.

Where Agent Builder fits (and where it stops)

You open Agent Builder by selecting New agent in the left pane of the Microsoft 365 Copilot app at microsoft365.com/chat, office.com/chat, or inside the Microsoft Teams desktop and web clients. It is available on both the Work and Web tabs on desktop and web, though you cannot author agents from mobile apps, and Microsoft formerly called this experience Copilot Studio lite.

On the Describe tab, you can explain what you want in plain English and let the tool draft the instructions, starter prompts, and knowledge connections, or you can click Skip to configure to set the name (up to 30 characters), description (up to 1,000 characters), instructions (up to 8,000 characters), and default response mode (Auto, Quick response, or Think deeper) by hand. Because agents built here do not consume your tenant's Dataverse storage entitlement and take minutes to stand up, the tool works well for personal productivity helpers and read-only team reference agents.

Before you treat it as your permanent platform, document six hard technical boundaries from Microsoft's extensibility documentation:

• One read-only SharePoint list: An agent here can read at most one SharePoint list of up to 20,000 items (rows) and 50 MB of raw text, ignores files in the list's Attachments column, and cannot add or update list rows. Also note that adding a SharePoint site URL does not include its lists; you must paste the direct URL to the specific list.
• No external write actions: You can toggle on built-in code interpreter (Create documents, charts, and code) and image generation (Create images), or add natural-language skills, but you cannot author actions or workflows that write to external services.
• No switch to block general AI knowledge: Turning on Only use specified sources tells the agent to prioritize your uploaded files or SharePoint links for search-based questions, yet it still uses general world knowledge for open-ended requests.
• No automated evaluation suite: You test the agent one prompt at a time in the Try it tab rather than running repeatable batch evaluations.
• No Teams group or 1:1 chat support: Users interact with the agent inside the Microsoft 365 Copilot app and Agent Store, and cannot pull it into a multi-person Teams group chat or 1:1 thread.
• No Solutions or Customer Managed Keys: Agents live outside Power Platform Solutions, have no separate development and production environments, and do not support Customer Managed Keys (CMK) or Purview Information Barriers on embedded uploaded files.

When to build in Copilot Studio instead

Microsoft Copilot Studio is a standalone low-code platform for building, testing, and governing custom agents, multi-step workflows, and system integrations on top of Microsoft Dataverse and the Power Platform. While Agent Builder stays inside the boundaries of a single declarative prompt and read-only knowledge, Copilot Studio gives makers a full authoring canvas for topics, generative orchestration, connectors, and automated flows.

We build in Copilot Studio whenever a workflow crosses from answering questions into following strict rules or changing records. First, you can turn off general AI knowledge completely so an HR, finance, or compliance agent only answers from your approved SharePoint libraries or returns a controlled fallback message. Second, the agent can run actions across external systems or Microsoft 365, such as looking up a vendor record, routing an approval in Teams, and writing a new row into a SharePoint list.

Third, Copilot Studio includes automated Evaluations (introduced in preview), which let you run a saved test set of realistic questions and grade the agent's accuracy, relevance, and safety scores every time you edit its instructions. Finally, every agent is stored in Dataverse and can be packaged inside a Power Platform Solution, giving IT proper development, test, and production environments alongside Power Platform data policies and connector governance.

What each option costs in seats and Copilot Credits

We have not seen a client surprised by a credit bill, because consumption should always be modeled up front with Microsoft's Copilot Studio agent usage estimator. That web tool lets you plug in your expected user counts, knowledge sources, and flow actions to forecast monthly credit usage before you turn an agent on.

For Agent Builder, authoring and using agents is included with paid Microsoft 365 Copilot licenses at no additional credit cost. If your organization has fewer than 300 seats, Microsoft Copilot Business provides the exact same Copilot capabilities as the enterprise add-on; see our Copilot Chat vs Microsoft 365 Copilot guide for a breakdown of how those seat licenses work.

Copilot Studio runs on Copilot Credits, which you can buy as a $200 per month capacity pack for 25,000 credits (pooled across the tenant with no monthly rollover) or on a pay-as-you-go meter at $0.01 per credit linked to an Azure subscription (with one-year pre-purchase Commit Units also available). To build in the studio, each maker needs either a $0 Copilot Studio user license assigned in the Microsoft 365 admin center or a Microsoft 365 Copilot license, plus a Dataverse entitlement in the target Power Platform environment.

Under the Standard harness, Microsoft bills consumption at these published rates:

Notice the right-hand column: when an authenticated employee with a Microsoft 365 Copilot license uses an internal agent, classic answers, generative answers, tenant graph grounding, agent actions, and agent flows called directly by the agent carry no extra credit charge (subject to fair use). In practice, Copilot Credits matter mainly when unlicensed employees use the agent, when you publish to external or customer channels, when an agent or flow runs autonomously on a schedule or event trigger, when you run Computer-Using Agents, or when metered runs consume premium reasoning tokens. Microsoft states that usage estimator outputs are informational rather than a guaranteed quote and recommends adding a 10% to 20% buffer above your estimate.

The Lilac Agent Build Path

When a 50- to 300-person company asks where to start, we advise pointing the first agent at a targeted operational bottleneck that eats team hours every week, such as document analysis or finding answers in standard operating procedures on a SharePoint site. Avoid kicking off with dozens of a-la-carte personal agents that summarize random threads without saving measurable hours or moving revenue.

Instead, we walk teams through the six steps of the Lilac Agent Build Path:

1. Pick one bottleneck with a repeatable question set. Choose a single process where staff lose hours digging through procedures, onboarding docs, or policy binders, and write down 15 to 25 real questions the team asks every week.
2. Build a read-only proof of concept in Agent Builder. Connect one to three SharePoint folders or sites after verifying that view permissions on those locations are locked down to the right groups.
3. Pilot with two to five users on live work. Share chat access (Can chat) with a small group of subject-matter experts for two weeks so they can test the agent against real daily questions and flag missing context.
4. Note the first capability ceiling the agent hits. Watch for the exact moment the team needs the agent to write a record to a SharePoint list, trigger an approval flow, block general AI knowledge, appear in a Teams group chat, or pass a repeatable batch evaluation.
5. Copy the agent to Copilot Studio and put it in a Solution. Open the ... menu in Agent Builder and select Copy to Copilot Studio, which creates a separate one-time snapshot in your chosen Power Platform environment. Immediately add that copied agent to a Power Platform Solution, and reconfigure anything that does not copy over automatically (uploaded embedded files must be re-uploaded, code interpreter must be re-enabled in settings, and Outlook or Teams knowledge must be reconnected via Power Platform connectors).
6. Run evaluations, assign co-owners, and estimate credits before rollout. Build a repeatable test set in Copilot Studio Evaluations, assign a named owner plus co-owners, and run the Microsoft usage estimator with a 10% to 20% buffer before publishing the agent to the organization catalog and retiring the original proof of concept.

Starter agents for workshops (plus one that belongs in Studio)

Non-technical staff can build a working Agent Builder agent in a single hands-on session when they do not start from a blank page. In our AI training workshops, we have participants start from Microsoft's 21 built-in Agent Builder templates and combine two complementary templates into one practical helper:

• Plan My Day + Meeting Coach: Compiles a morning briefing from the maker's own Outlook inbox and calendar, then shifts into drafting agendas, talking points, and prep notes for the day's meetings.
• My Company Policy + Learning Coach: Grounds answers in authoritative HR or operations procedure documents on SharePoint, then breaks complex policies into layered summaries and practice questions for new hires.
• Project Delta Digest + Status Update Agent: Scans project activity to summarize what shipped, what is in progress, and what is stuck, then turns those notes into audience-ready weekly status updates.

When a workshop idea needs to write data or enforce an approval gate, we move it to Copilot Studio as part of our AI implementation work. A classic example is an IT or operations request agent that writes to a SharePoint list: an employee asks for software access or replacement gear, the agent checks the policy PDF on SharePoint, collects the cost center and justification, triggers an agent flow for manager approval, and writes the approved row into a SharePoint tracking list. Because Agent Builder can only read one SharePoint list and cannot author write actions or approval flows, that workflow belongs in Copilot Studio from the moment it moves past read-only Q&A.

Five common mistakes we fix before rollout

Most agent headaches come from governance shortcuts taken during the first week of building. Check for these five issues before you publish an agent to a department or the company store:

• Leaving SharePoint permissions too wide: An agent surfaces whatever files the signed-in user can already open, and Agent Builder's automatic file-sharing option only works when sharing with specific security groups rather than the whole organization. Audit and tighten folder permissions before connecting a SharePoint source.
• Leaving a Copilot Studio agent outside a Solution: Makers often build directly in the tenant's Default environment without creating a Power Platform Solution. Put every Copilot Studio agent inside a Solution in a dedicated environment so you can version changes and promote them cleanly from development to test and production.
• No ownership plan for when the builder leaves: Never build agents under generic shared accounts. Follow Microsoft's ownership model: assign a named owner plus individual co-owners, use Solution-based ALM with a service principal for automated pipeline deployments, and remember that admins can reassign orphaned agents in the admin center when someone changes roles.
• Skipping automated evaluations before launch: Testing three friendly prompts in the chat pane does not tell you how an agent handles ambiguous policy questions after an instruction edit. Use Copilot Studio Evaluations to run a repeatable test set and safety check before every release.
• Skipping a two-to-five-user pilot: Rolling an agent out to the whole company on day one exposes broken SharePoint links and missing instructions to everyone at once. Run a two-week pilot with two to five people who know the underlying process cold before enabling org-wide discovery.