How do you build an agent in Copilot Studio, step by step?
To build an agent, go to copilotstudio.microsoft.com, describe it in natural language, add knowledge and tools, test it and publish it to Teams or your website; with the trial licence you can build and test, but not publish. This is the detail of the official quickstart (22 September 2026), which includes screenshots of every screen, for an agent on the standard harness:
- Sign in at copilotstudio.microsoft.com with an account in your tenant. You need a Copilot Studio licence, a Microsoft 365 Copilot licence or the trial licence.
- Describe the agent in natural language on the home page, up to 1,024 characters. For example: "Answer staff questions about holidays and leave according to the collective agreement and internal policies". Mind the primary language: according to Microsoft Learn's language page, it cannot be changed after the agent is created.
- Review what the AI generates: Copilot Studio proposes a name, description and instructions, and suggests knowledge sources, tools, triggers and channels. Accept or discard each suggestion and refine the instructions on the Overview page, where you also pick the model.
- Add knowledge in the Knowledge section (Add): a SharePoint site, a public website or documents. Connect only sources that are current and have an owner.
- Add tools if the agent needs to act: a connector (Dynamics 365, SAP, ServiceNow), an agent flow or an MCP server.
- Test in the "Test your agent" pane: ask the questions real users will ask, change the instructions and repeat. Edit the welcome message in the Conversation Start topic. Agent flow runs started from the test chat do not consume capacity.
- Publish with Publish. The first destination can be the demo website Copilot Studio generates, handy for showing the pilot.
- Add channels and authentication: Teams and Microsoft 365 Copilot for internal use, or your website and WhatsApp for customers. Review authentication before opening it to anyone (see the warning under use cases).
If your company only has the Copilot Studio for Teams plan included in some Microsoft 365 subscriptions, Microsoft has a Teams-specific quickstart with fewer capabilities: no generative orchestration, no premium connectors and publishing to Teams only.
What are Copilot Studio harnesses and how do they affect cost?
A harness is the runtime that executes what you build in Copilot Studio (it decides when to call the model, what context to send and which tools to use), and the one you choose determines how you are billed. According to the harness documentation (19 October 2026), there are three:
| Harness |
What it is for |
How it is billed |
| Standard |
Rule-based agents and structured, repeatable conversations (for example, an internal help desk) and agent flows |
Rate per event: classic answer, generative answer, agent action |
| GitHub Copilot |
Complex, multi-step business processes: reasons towards a goal, uses several tools, creates and edits Word, Excel, PowerPoint and PDF files |
Credits for tokens, tools and the harness itself, from the moment you start building |
| Customer Conversation |
Customer service over chat or voice, with specialised sub-agents and escalation to a person |
Credits by complexity (light, medium, heavy) and modality (digital, GenAI voice, premium voice) |
Source: Microsoft Learn (19 October 2026). Last verified: November 2026.
Three details that shape the decision, according to the usage-based billing page (6 October 2026):
- The GitHub Copilot harness charges from the moment you start building: unlike the standard harness, which bills after publishing, creating the solution in natural language, testing it and generating evaluations already consume credits. Each agent's Monitor page shows what has been used. Budget for the development phase, not just production. Despite the name, this harness does not send customer data to the GitHub Copilot service.
- A Microsoft 365 Copilot licence gives access, not free consumption: the licensing guide says a Microsoft 365 Copilot licence lets you use Copilot Studio to build agents on the GitHub Copilot harness. But the credit exemption for licensed users in the rates table refers to the standard harness, and the usage-based billing page provides no exemption on the GitHub Copilot harness. In practice: with a licence you can build on it, and you should still budget credits.
- Customer Conversation is a contact centre product: Microsoft ties it to the Self-Service/Customer Assist Agent in Microsoft Contact Center Intelligence, the name used on the billing page itself. If you want a voice agent for your service centre, the conversation is also about contact centre licensing, not just Copilot Studio. If what you need is a customer chat assistant without setting up a contact centre, compare options in our enterprise chatbots service.
How much does Copilot Studio cost? Pricing and licences 2026
Copilot Studio is paid for in Copilot Credits, and there are three ways to buy them: a monthly pack, pay-as-you-go through Azure or a one-year pre-purchase plan. Prices come from Microsoft's official Copilot Studio pricing page (identical in the Spanish, German, French and Italian versions) and the Power Platform pay-as-you-go meters table. Credits replaced "messages" on 1 September 2025 without changing the quantity per pack or the rate, according to Microsoft's licensing guide.
| Credit purchase option |
Price (excl. VAT) |
Where you buy it |
When it fits |
| Credit pack (tenant licence) |
€173.30 per month for 25,000 credits (~€0.0069 per credit) |
Microsoft 365 admin centre |
Stable, predictable usage |
| Pay-as-you-go |
$0.01 per credit, billed monthly in arrears |
Azure subscription linked to the environment |
Pilots and variable or seasonal usage |
| Pre-purchase plan (one year) |
Savings of "up to 20%", according to the pricing page |
Azure portal |
High, known volume |
Sources: Copilot Studio pricing page and Microsoft Learn licensing guide (3 August 2026). Prices last verified: 2 November 2026. Microsoft revises these tables often; check the official page before budgeting.
Separate from credits is the Microsoft 365 Copilot licence, which does not buy credits but is paid per user: €26.00 per user/month with annual billing on Enterprise plans, according to the Copilot Studio pricing page, and it includes building agents and using them internally without consuming credits (more below). Business plans have Microsoft 365 Copilot Business: according to the Microsoft 365 Copilot pricing page for Spain, it costs €18.20 per user/month with annual billing or €21.84 with monthly billing, with a promotional price of €15.60 (annual billing) between 1 July and 31 December 2026. If you are budgeting for 2027, do not count on the promotion.
Packs or pay-as-you-go? The 77% break-even
Pack capacity applies monthly and unused credits do not carry over. Pay-as-you-go is priced in dollars and Azure bills it in your subscription currency. At the ECB reference rate of 30 October 2026 (€1 = $1.1172, i.e. €0.895 per dollar), each pay-as-you-go credit costs about €0.0090, against €0.0069 with a pack.
The result: a pack only pays off if you use more than roughly 77% of its credits each month; below that, pay-as-you-go is cheaper. If the exchange rate moves, recalculate with this formula: break-even = 0.6932 ÷ (euros per dollar). At €0.92 per dollar, for example, the break-even drops to 75%.
What happens when Copilot Studio credits run out?
If the tenant reaches 125% of its prepaid capacity and has no linked pay-as-you-go plan, the environment's agents are blocked, even mid-conversation, and most users see "This agent is currently unavailable". That is what Microsoft's billing and consumption policy says, with two nuances: users signed in with a Microsoft 365 Copilot licence might not be blocked, and agent flows are blocked separately (the agent can keep answering whatever does not need a flow).
Linked pay-as-you-go prevents the cut-off, with exceptions. If the environment bills overage to a pay-as-you-go plan, running out of prepaid credits does not block agents and the overage is billed in Azure. But, according to the same page, these are still blocked:
- agents an admin has turned off;
- agents that hit a monthly limit set to stop usage.
The pricing page makes the same point, recommending automatic pay-as-you-go coverage to avoid interruptions, and states that an Azure subscription is required to use agents: even if you buy packs in the Microsoft 365 admin centre, plan on having Azure. For a critical agent, the sensible combination is a pack + pay-as-you-go as cover + a monthly limit that alerts rather than stops.
How many credits does each Copilot Studio interaction consume?
Every agent event has a fixed credit rate, and cost always follows the same chain: event → credits → euros. The official rates table (updated 30 October 2026) sets these for the standard harness:
| Agent event |
Credits |
| Classic answer (predefined text) |
1 |
| Generative answer |
2 (free for Agent Builder agents without Graph grounding) |
| Agent action (triggers, reasoning, topic transitions, computer-using agents) |
5 |
| Tenant Microsoft Graph grounding |
10 |
| Agent flow actions |
13 per 100 actions |
| Basic AI tools |
1 per 10 responses |
| Standard AI tools |
15 per 10 responses |
| Premium AI tools (reasoning models) |
10 per 1,000 tokens |
| Content processing |
8 per page |
If the agent handles phone calls, a per-minute voice rate is added that already includes basic answers and actions:
| Voice tier |
Credits per minute |
Approximate cost with packs |
| Classic Voice (classic orchestration) |
10 |
~€0.07/min |
| GenAI Voice |
35 |
~€0.24/min |
| Premium GenAI Voice |
75 |
~€0.52/min |
Source: Microsoft Learn (30 October 2026); euro conversion based on the pack price. Last verified: November 2026. These rates exclude your own models deployed in Microsoft Foundry, which are billed separately.
A real calculation using Microsoft's examples
Microsoft documents, as an illustrative example, a customer service agent on a website: four classic answers and two generative answers per conversation, 900 customers a day. That is (4×1 + 2×2) × 900 = 7,200 credits a day, roughly 216,000 a month. With packs, that means 9 packs (225,000 credits): €1,559.70 a month excluding VAT. On pay-as-you-go it would be $2,160, about €1,933 at the ECB rate: here the pack wins because 96% of its capacity is used. Approximate cost per conversation with packs: around €0.06 (8 credits).
Now the warning. In the same document, an agent using Claude Opus 4.7 as its reasoning model consumes 141 credits in tokens alone to answer a follow-up question, and 461 credits if the question opens a new conversation and the cache barely helps. The event rate is added on top (2 credits per generative answer, 5 per action). In the example's worst case (a new conversation) that is more than $4.61, about €4.13, for a single message; in the follow-up case, more than $1.41 (about €1.26). The model you choose weighs more on the bill than the number of users: use reasoning only where it is needed, keep instructions and sources focused, and set monthly limits per agent from the Power Platform admin centre.
How to measure real consumption in your pilot
Rates give you the order of magnitude; design gives you the bill. Before choosing the model and purchase option, compare the estimate with observed consumption over two or three weeks of pilot using this template:
| Data point |
Where to get it |
What it decides |
| Estimated credits per conversation |
Official agent usage estimator |
Initial budget and monthly limit per agent |
| Credits consumed per agent per day |
The agent's Monitor page |
Whether the real design deviates from the estimate |
| Tenant consumption against purchased capacity |
Licensing > Copilot Studio in the Power Platform admin centre |
Distance to the 77% break-even and the 125% cut-off |
| Real credits per conversation (consumption ÷ sessions) |
The agent's Monitor page + Analytics page |
Whether to switch model or trim sources |
If real credits per conversation stray far from the estimate, first check the three levers that move consumption most according to the rates table: a reasoning model (10 credits per 1,000 tokens), Graph grounding (10 credits per query) and chained actions (5 credits each).
Do I need to pay for Copilot Studio if I already have Microsoft 365 Copilot?
No, for internal agents on the standard harness used by employees with a Microsoft 365 Copilot licence in Copilot Chat, Teams or SharePoint: their classic and generative answers, Microsoft Graph grounding and agent actions consume no credits, according to Microsoft's rates table. Two conditions in the small print (footnote 1 of that table): the agent must run under the licensed user's identity, and usage is subject to fair usage limits that Microsoft may update. Agent flows are also included, but only when the agent launches them with the "When an agent calls the flow" trigger.
You do pay credits, even if every employee is licensed, in these cases:
- Autonomous agents triggered by an event (an incoming email, a new order) rather than by a licensed user. Their agent flows consume credits too.
- External agents: customers on your website, WhatsApp or phone.
- Internal users without a Microsoft 365 Copilot licence. In Microsoft's example, a sales agent used by 50 licensed and 100 unlicensed users only bills for those 100: 4,800 credits a day.
- Computer-Using Agents, expressly excluded from the inclusion.
- Agents on the GitHub Copilot harness, billed by tokens and tools, as covered above.
Practical takeaway: for an SME already paying for Microsoft 365 Copilot for key roles, the first internal agents can come at no extra consumption cost. A credit budget appears when the agent goes external, works on its own, uses the GitHub Copilot harness or is used by unlicensed employees.
What are Copilot Studio's best use cases in business?
The use cases that perform best are repetitive, backed by documented knowledge and have a system behind them to act on. These are the patterns Microsoft's own documentation uses as reference, adapted to a mid-sized European company:
| Use case |
Harness / tool |
Concrete example |
| Internal IT or HR help desk |
Standard |
Answers questions about holidays, payroll or VPN from the collective agreement and internal policies, and opens a ticket when it does not know |
| Customer service on the web or by phone |
Customer Conversation (Contact Center Intelligence) |
Returns, order status and product manuals, with escalation to a human agent |
| Order processing |
Standard, autonomous |
When an order comes in, checks stock and lead times, approves it and sends the confirmation to the customer (about 20 credits per order in the official example) |
| Accounts payable |
GitHub Copilot |
Reads invoices, matches them against purchase orders and routes exceptions for approval |
| Sales assistant |
Standard + Graph |
Answers sales reps with sales data connected to Microsoft Graph |
| Employee onboarding |
Agent Builder (or Studio if actions are needed) |
Guides new hires through approved documentation and, in Studio, provisions them in systems |
Source: Microsoft Learn examples (October 2026) adapted by Technova. Last verified: November 2026.
For customer chat, the choice between Copilot Studio and other platforms depends on volume and the systems behind it; we cover it in our enterprise chatbots service.
Cost scenario: HR help desk at a manufacturing company
To ground the rates, here is a hypothetical scenario we calculated at Technova using the official rates (not a specific client). A 400-employee manufacturing company wants an agent in Teams that answers questions about holidays, leave and payroll from the collective agreement and internal policies, and opens a case in its HR tool when it cannot answer. 100 office staff have Microsoft 365 Copilot; the 300 shop-floor workers do not.
| Assumption |
Value |
| Monthly conversations from unlicensed users |
1,500 |
| Design per conversation |
3 generative answers (6 credits) + 1 agent action (5) = 11 credits |
| Monthly consumption |
16,500 credits (66% of a pack, below the 77% break-even) |
| Pay-as-you-go cost |
$165, about €148 a month (versus €173.30 for a pack) |
| Conversations from the 100 licensed users |
0 credits |
| Same agent with a reasoning model (~141 credits per answer, per Microsoft's example) |
~434 credits per conversation: 651,000 a month, about €5,827 on pay-as-you-go or €4,679.10 with 27 packs |
Technova calculation using Microsoft Learn rates (30 October 2026) and the ECB rate (30 October 2026). Real consumption depends on instructions, sources and conversation length.
The takeaway is twofold. With a general model, the agent costs about €0.49 per shop-floor worker per month, against €26.00 for one Microsoft 365 Copilot licence per person. With a reasoning model on every answer, the same use case costs almost 40 times more. That is why the model decision should be made with pilot data, not by default.
Three more practical warnings. First: the agent answers with whatever its knowledge sources contain, so if internal policies are outdated or duplicated across three SharePoint folders, the agent will be too. Second: language support is uneven; for example, Catalan works in chat answers but not in voice, according to Microsoft's language page, so check every language your staff use. Third: an external channel means authentication and governance. The publishing documentation warns that the "no authentication" option lets anyone with the link talk to the agent, and that in this mode it cannot use tools with user credentials.
"The question we hear most is how much Copilot Studio costs, and the honest answer is that it depends on the design more than the licence. At the official rates, a help desk conversation costs 11 credits with a general model and more than 400 if it reasons on every answer: almost 40 times more for the same service. That's why we recommend measuring for two or three weeks on the Monitor page before choosing the model and the purchase option," says Alfons Marques, founder of Technova Partners.
Is Copilot Studio GDPR-compliant and where does it store data?
Copilot Studio lets you keep environment data in a European geography, but GDPR compliance depends on how you configure it (geography, models and channels), not just on the product. The Copilot Studio data location page lists the Europe geography, with West Europe (Netherlands) and North Europe (Ireland), plus France, Germany, Sweden, Switzerland, Norway and the United Kingdom; it lists no region in Spain or Italy.
Microsoft gives no selection criteria; ours has three steps:
- Align with your Microsoft 365 tenant: conversation history is stored in the region of each user's Exchange mailbox anyway, so you gain nothing by separating the environment from that geography.
- Apply what your contract or DPO requires: if data must stay in the European Union, choose Europe, France, Germany or Sweden; the United Kingdom and Switzerland are not in the EU.
- Check feature availability in Microsoft's availability-by-geography report before committing to a use case.
For an EU company with no special requirements, that usually leads to the Europe geography or its own country's geography where one exists. Microsoft keeps customer data in the chosen geography, with documented exceptions; the ones that matter most:
- External channels such as Facebook or Teams, and handoff to human agents, may move data outside the chosen geography.
- Agent activity history, including conversation logs, is stored in the region of the user's Exchange mailbox.
- If the tenant location is not in the table, data is stored in the United States. Check the tenant location before creating environments.
Which AI models does Copilot Studio use and where is data processed?
By default, standard harness agents use OpenAI's GPT Chat Latest, which is processed in-region in Europe. According to the model selection page (22 October 2026), these are the generally available models in Europe (except the UK):
| Model |
Type |
Processing in Europe |
| GPT Chat Latest (OpenAI, default) |
General |
In-region |
| GPT-4.1 and GPT-5 Chat (OpenAI) |
General |
In-region; retiring on 12 November 2026 |
| Claude Sonnet 4.6 (Anthropic) |
General |
Cross-geo |
| Claude Opus 4.6, 4.7 and 4.8 (Anthropic) |
Reasoning (deep) |
Cross-geo |
Source: Microsoft Learn (22 October 2026). Last verified: November 2026.
Cross-geo means the model may process data outside your region. The GPT-4.1 and GPT-5 retirement does not apply to environments outside the United States where data movement across regions is restricted. Besides Anthropic, Microsoft lets you add external models from xAI and Mistral; to use any external model, the admin must turn them on in Power Platform and allow each provider in the Microsoft 365 admin centre. It is a decision for your DPO, not just for whoever builds the agent. Your own models deployed in Microsoft Foundry are also supported, billed separately.
For governance, Copilot Studio inherits Power Platform controls: separate environments, data policies that block connectors or channels, agent inventory and consumption limits.
Agent 365: agent inventory at scale
Microsoft Agent 365 is the control plane for observing, securing and governing AI agents across the organisation. It is a separate product: $15 per user/month with an annual commitment, or included in Microsoft 365 E7. In its fiscal year 2026 fourth-quarter earnings (29 July 2026), Microsoft said Agent 365 had "nearly 40 million agents registered across tens of thousands of companies". An SME with two or three agents does not need it; a company with agents built by several departments outside IT's control probably does.
How do you roll out Copilot Studio in 5 steps?
- Pick a process with volume and documentation: a help desk with hundreds of queries a month or an order flow, not "one agent for everything".
- Choose the tool using the table above: Agent Builder if internal knowledge is enough; Copilot Studio if there are actions, an external channel or a formal lifecycle. Also check language limits if the agent will be autonomous or voice-based.
- Estimate consumption before building and measure it afterwards: use the official agent usage estimator, start the pilot on pay-as-you-go with a monthly limit per agent, and fill in the measurement template above. With two or three weeks of data, apply the 77% rule to choose between packs and pay-as-you-go.
- Prepare permissions and sources: clean up the SharePoint the agent will read and define which channels and models are allowed in the admin centre.
- Train both builders and users: Article 4 of the EU AI Act, Regulation (EU) 2024/1689, applicable since 2 February 2025, requires deployers of AI systems to take AI literacy measures for their staff. For those who build agents, our AI agents training for technical teams covers design, testing and governance; in Spain, that training can be subsidised through FUNDAE, as we explain in our guide to FUNDAE-subsidised AI training.
The context: 20.0% of EU companies with 10 or more employees used AI in 2025 according to Eurostat, and in Spain the figure reached 21.1% in the first quarter of 2025 (+8.7 points), according to Spain's INE ICT survey. For a full SME plan, see our guide to implementing AI agents in an SME.
If you would rather not go it alone, at Technova we design and deploy agents on Copilot Studio and other platforms as part of our data and AI services, specifically our AI agents for business service, and we train the teams who maintain them through our Copilot training for business, part of our AI training programme.
When is Copilot Studio not the right choice?
Copilot Studio is not the right choice when your company does not work in Microsoft 365, when an external channel has high volume with reasoning, when you need vendor independence or when the use case runs into its language limits:
- You do not work in Microsoft 365: if your email, documents and CRM live in Google Workspace or other platforms, you lose the main advantage (Graph, Entra ID, Teams).
- Very high-volume customer service with reasoning: given the per-answer cost of reasoning models, an external agent handling tens of thousands of conversations a month can end up costing more than a custom solution with a more efficient model.
- You need vendor independence or your own model: Copilot Studio supports your own models through Microsoft Foundry (formerly Azure AI Foundry), but they are billed separately and remain within the Microsoft ecosystem.
- Voice in Catalan, Basque or Galician, or autonomous agents in languages other than English: voice does not support those three languages, and the language page only lists English (United States) for the event triggers that fire autonomous agents. Validate it in the pilot before designing the process around them.
- Highly specific logic or critical integrations: custom development gives you more control there; our guide to AI agent development costs compares both routes.
Alternatives to Copilot Studio and when they fit better
| Alternative |
Pricing model |
Fits better when… |
Source |
| Agent Builder |
Comes with Microsoft 365 Copilot; its generative answers without Graph grounding consume no credits |
A knowledge-only internal agent is enough |
Microsoft Learn (October 2026) |
| Microsoft Foundry |
Azure consumption (models and resources) |
You have a development team and need your own code, model choice (OpenAI, Anthropic, Meta, Mistral) or network isolation |
Microsoft Cloud Adoption Framework |
| n8n |
Cloud from €20 per month (2,500 executions, annual billing); free self-hosted Community edition |
AI automations across many tools, inside or outside Microsoft |
n8n pricing |
| Gemini Enterprise or ChatGPT Enterprise |
Per-user licence |
The company works in Google Workspace or has already standardised on ChatGPT |
Google and OpenAI websites |
| Custom development |
Project + model API consumption |
High external volume, critical logic or vendor independence |
Technova's view |
n8n prices verified on 2 November 2026. We do not include Google or OpenAI prices because they depend on edition and contract: check them on their websites.
If n8n is your alternative, we compare it with Zapier in our Zapier vs n8n comparison; for other platforms, see our AI agent platforms comparison and our ranking of the best AI agents for business.
Evaluating Copilot Studio, or already running agents with no cost control? Tell us about your case and we will help you choose the tool, estimate consumption and deploy your first agent.
Decision checklist: packs, pay-as-you-go or Microsoft 365 Copilot
Decide based on who uses the agent, how much they use it and which harness it runs on:
- Only employees with a Microsoft 365 Copilot licence, in Teams, Copilot Chat or SharePoint, with the agent on the standard harness → do not buy credits for that agent; just watch agent flows with other triggers, computer-using agents and fair usage limits. If the agent runs on the GitHub Copilot harness, budget credits even if everyone is licensed: the exemption does not apply.
- Unlicensed employees or autonomous agents → calculate credits with the estimator. If volume is stable and above 77% of a pack, buy packs; otherwise, pay-as-you-go.
- External customers (website, WhatsApp) → always credits. Start on pay-as-you-go with a monthly limit; switch to packs once you have three months of stable usage.
- Phone support → calculate by voice minutes (10, 35 or 75 credits) and consider Contact Center Intelligence licensing.
- Any agent with a reasoning model → measure actual consumption per message on the Monitor page during the pilot before opening it to more users.
- Always → link pay-as-you-go as cover so the 125% cut-off does not stop a production agent, and check that monthly per-agent limits are not set to stop usage if that agent is critical.
Frequently asked questions about Copilot Studio
What is Copilot Studio and what is it used for?
Copilot Studio is Microsoft's low-code tool for building AI agents and automations connected to company data. It is used for internal assistants (IT or HR help desk, onboarding), customer service agents on the web, WhatsApp or phone, and autonomous processes such as order or invoice handling, published in Teams, Microsoft 365 Copilot or your website.
How much does Copilot Studio cost?
€173.30 per month for a pack of 25,000 credits, or $0.01 per credit on pay-as-you-go, excluding VAT; the one-year pre-purchase plan saves up to 20%. Internal agents used by employees with a Microsoft 365 Copilot licence (€26.00 per user/month) consume no credits for their core functions.
What licence do I need to use Copilot Studio?
It depends on the use. To try it, the individual trial licence is enough: you can build and test but not publish. Some Microsoft 365 subscriptions include the Copilot Studio for Teams plan, without generative orchestration or premium connectors. For production you need credits (packs or pay-as-you-go, with an Azure subscription) or Microsoft 365 Copilot licences, which let you build agents and cover internal use by licensed employees on the standard harness.
Is Copilot Studio free?
Not for production. The individual trial licence lets you build and test agents, but not publish them. To publish you need credits (packs or pay-as-you-go) or Microsoft 365 Copilot licences, which cover internal agents used by licensed employees in Copilot Chat, Teams or SharePoint.
How long is the Copilot Studio free trial?
Microsoft Learn does not state a duration in days: it sends notifications and emails before it expires, and on expiry you can extend it by 30 days. Trial agents keep working for up to 90 days after expiry. During the trial you can build and use the test chat, but not publish.
Is Copilot Studio included in Microsoft 365?
Not in standard Microsoft 365 plans. Some subscriptions include a Copilot Studio for Teams plan with limited capabilities. The Microsoft 365 Copilot licence (€26.00 per user/month on Enterprise) does include building agents and using them internally without consuming credits; external or autonomous agents need separate credits.
Which languages does Copilot Studio support?
English, Spanish, French, German and Italian are generally available for the interface, generative answers and voice, according to Microsoft Learn's language page. Catalan, Basque and Galician are supported in the interface and in chat answers, but not in voice. The agent's primary language is chosen at creation and cannot be changed afterwards, and Microsoft only lists English (United States) for the event triggers of autonomous agents.
Which AI models does Copilot Studio use?
OpenAI's GPT Chat Latest by default, processed in-region in Europe. You can also choose Anthropic models (Claude Sonnet 4.6 and Claude Opus 4.6, 4.7 and 4.8), which in Europe process data outside the region, and add external models from xAI and Mistral if the admin enables them. GPT-4.1 and GPT-5 Chat retire on 12 November 2026. Your own Microsoft Foundry models are billed separately.
Copilot Studio is low-code SaaS for business teams to deploy agents quickly; Microsoft Foundry is the PaaS platform for developers, with your own code, model choice and network control. Microsoft's Cloud Adoption Framework recommends prototyping for one or two weeks on each option and sticking with low-code if it meets the functional and security requirements. They can be combined: a Copilot Studio agent can use models deployed in Foundry.
What is the difference between Copilot Studio and Microsoft 365 Copilot?
Microsoft 365 Copilot is the assistant built into Word, Excel, Outlook and Teams, paid per user. Copilot Studio is the tool for building your own agents, with actions on your systems and publishing to channels such as Teams, your website or WhatsApp, paid by credit consumption. They complement each other: Copilot Studio agents can be used inside Microsoft 365 Copilot.
What is the difference between Copilot Studio and Power Automate?
Power Automate automates rule-based flows and is licensed through Power Automate; its cloud flows do not consume Copilot Studio credits. Copilot Studio builds agents that converse and reason, and its own agent flows (very similar to Power Automate's) are billed in credits. A Copilot Studio agent can call flows as tools.
Is Copilot Studio the same as Power Virtual Agents?
It is its successor. In November 2023 Microsoft merged Power Virtual Agents' capabilities into Copilot Studio and retired that name. Existing bots keep working, and Copilot Studio adds generative AI, harnesses, MCP servers, workflows and publishing inside Microsoft 365 Copilot.
How many credits does a Copilot Studio agent consume?
It depends on the design. On the standard harness, a classic answer costs 1 credit, a generative one 2, an agent action 5 and Microsoft Graph grounding 10. In Microsoft's examples, a customer service conversation uses 8 credits, an answer with Graph grounding 12, and a single message with a reasoning model can exceed 400. Use the official estimator before building.
Is there an official Copilot Studio certification?
Yes. The Microsoft Certified: AI Agent Builder Associate certification (exam AB-620) assesses planning, integrating and managing agents in Copilot Studio. For architects there is Agentic AI Business Solutions Architect Expert (exam AB-100), which also requires an Associate-level certification such as PL-200 or AB-620 itself.
Can you use Copilot Studio without coding?
Yes. It is a low-code tool: you describe the agent in natural language, add knowledge sources and connectors without writing code, and test it before publishing. For integrations with your own APIs, authentication on external channels or environment governance, it is worth having technical profiles or a partner, and training whoever will maintain it.
