Your Copilot Studio Chat Could Cost $7.84 on a Reasoning Model

Copilot Studio costs $0.11 to $0.41 per internal HR conversation and about $0.15 per support chat on a standard model. A reasoning model can push one conversation past $7.

By Rajesh Beri·November 16, 2026·17 min read
Share:
A finance analyst's desk with a printed monthly cloud invoice covered in highlighter marks, a calculator, and a laptop showing a bar chart of daily usage climbing sharply at the end of the month.

Illustration generated using AI

On a standard model, one Copilot Studio conversation costs between $0.11 and $0.41 for an internal HR agent and about $0.15 to $0.18 for a customer support agent. Switch either agent to a reasoning model and Microsoft's own token example puts a three-turn conversation between $7.61 and $7.84. The bill depends on the agent's design: which features fire on each turn, which model runs the orchestration, and whether the person asking holds a Microsoft 365 Copilot licence.

The verdict: build internal agents in Copilot Studio when most of the people using them already hold Microsoft 365 Copilot seats, because their usage is not billed. For a customer-facing support agent, Copilot Studio on a standard model is the cheapest of the three options here per conversation, but only if you set a per-agent monthly limit before launch and keep reasoning models off the hot path. Agentforce is the expensive choice for a support agent unless your cases already live in Service Cloud. Intercom's Fin costs more per conversation than Copilot Studio and charges only when Fin produces an outcome, which suits teams that can't staff a Power Platform admin.

Prices below were checked on each vendor's live page on November 16, 2026.

Price unit (checked Nov 16, 2026) Internal HR agent, 8,000 conversations/month Support agent, 30,000 conversations/month Who should not buy it
Copilot Studio, standard model $200 per 25,000-credit pack per month; $0.01 per credit pay-as-you-go $0 for Microsoft 365 Copilot licensed users; about $2,630 for unlicensed staff about $4,400 (22 packs) Tenants with nobody assigned to watch the Power Platform admin center
Copilot Studio, reasoning model Same credits, plus 10 credits per 1,000 billed tokens about $62,700 for unlicensed staff about $228,000 Anyone who hasn't measured token counts on real conversations
Salesforce Agentforce $500 per 100,000 Flex Credits; 20 credits per action about $3,200 about $15,000 Companies whose support cases don't live in Salesforce
Intercom Fin $0.99 per outcome Not built for internal HR use about $17,800 if 60% of conversations end in a billed outcome Teams that need an internal employee agent

What Copilot Studio Costs per Credit in November 2026

A Copilot Credit is the unit Microsoft uses to meter agent usage in Copilot Studio; every billable step an agent takes deducts a fixed number of them. There are three ways to buy credits.

Microsoft's Copilot Studio pricing page lists a Copilot Credit pack at $200 per pack per month for 25,000 credits, which works out to $0.008 per credit. The Power Platform pay-as-you-go meters page prices the Copilot Studio meter at $0.01 per credit, billed to an Azure subscription. The third option is the Copilot Credit prepurchase plan, a one-year pool of Copilot Credit Commit Units that you buy in the Azure portal, per Microsoft's standard harness licensing page. The pricing page says it saves "up to 20%" and sends you to the Copilot Studio Licensing Guide for the tiers. Microsoft doesn't publish the tier breakpoints on the pricing page, so get the quote in writing.

The pack versus pay-as-you-go decision is plain arithmetic. One pack costs the same as 20,000 credits on the meter. If an environment reliably uses more than 80% of a pack each month, the pack is cheaper. Below that, pay-as-you-go wins, because unused pack credits don't carry over to the next month.

Two other prices sit next to the credits. The Microsoft 365 Copilot licence is $30 per user per month, paid yearly, and it covers internal agent use by the person holding it. And since November 1, 2026, new and renewed Power Apps, Power Automate and Dynamics 365 premium licences no longer include seeded AI Builder credits, per Microsoft's AI Builder end-of-credits page. AI Builder prompts in your existing flows fall back to Copilot Credits when the old credits run out, and if no Copilot Credits are available, "usage will be blocked." If your Power Platform estate runs document processing or prompts, that usage is now on the same meter as your agents.

How Each Agent Feature Burns Credits

The billing rate table is where the cost actually lives. Each turn of a conversation can trigger several features, and each feature bills separately. These are the rates from Microsoft's billing rates page, checked November 16, 2026:

Agent feature Copilot Credits At $0.01 pay-as-you-go
Classic answer (a scripted, author-written reply) 1 $0.01
Generative answer 2 $0.02
Agent action (triggers, deep reasoning, topic transitions, computer use) 5 $0.05
Tenant graph grounding (retrieval over Microsoft Graph) 10 $0.10
Agent flow actions 13 per 100 actions $0.13
Text and generative AI tools, premium (reasoning models) 10 per 1,000 tokens $0.10 per 1,000 tokens
Content processing tools 8 per page $0.08
GenAI voice 35 per minute $0.35

Microsoft gives its own example: an agent grounded in the tenant graph can use 12 credits on a single complex prompt, 10 for the grounding and 2 for the generative answer. Tenant graph grounding is the single most expensive per-turn feature, and it is optional. You can turn it on or off for each agent.

The reasoning-model meter changes the math

Reasoning models are billed on two meters. You pay the normal feature rate, plus 10 credits per 1,000 tokens the model processes, counting input and output and excluding tokens read from its prompt cache. Microsoft's worked example on the same page uses Claude Opus 4.7 on a support agent: one follow-up question took two model calls, billed 14,100 tokens and cost 141 credits before feature rates. If the same question opened a new conversation, with most of the prompt not yet in the cache, the message cost 461 credits.

At pay-as-you-go, that's $1.41 for a follow-up and $4.61 for an opening message, against $0.02 for a generative answer on a standard model. Microsoft labels these token counts illustrative, and your agent's instructions, knowledge sources and conversation length will move them. A cost analysis from CloudZero, which sells cost-management software, puts the spread for the same agent at $8 a month or $800 depending on how it's built, without showing its arithmetic. Microsoft's own example is enough to treat the model choice as a budget decision.


When a Microsoft 365 Copilot Licence Covers Agent Use

A Microsoft 365 Copilot licence covers an agent's usage when three things are true: the agent is employee-facing, the person using it holds the licence, and the agent runs with that person's authenticated identity. Every feature in the rate table, including reasoning-model tokens, shows "No charge" for that user, subject to fair-use limits Microsoft can change, according to footnote 1 of the billing rates table.

The coverage stops in several places that catch budget owners:

  • Staff on Copilot Chat without the paid seat are billed at the standard rate. Microsoft's own sales-agent example on the billing page charges for the 100 unlicensed users and not for the 50 licensed ones.
  • Customers, partners and anonymous website visitors are never covered by an employee's licence, so every customer-facing agent is fully metered.
  • Agent flow runs called by an agent that an event trigger or a Power Automate flow started consume credits even if everyone in your organization holds a Copilot licence. The same applies to agent flows that use any trigger other than "When an agent calls the flow."
  • Computer-using agent (CUA) usage is excluded from the licence inclusion and billed at the agent action rate.

The licence math rarely justifies buying seats just to zero out an agent bill. Covering 2,000 HR agent users with Microsoft 365 Copilot costs $60,000 a month at list price. The same population on pay-as-you-go costs about $2,630 to $3,280 a month in the HR example below. Buy seats for the broader Copilot features and treat the agent inclusion as a side benefit.

One support thread on Microsoft Q&A from March 2026 shows a less obvious gap: a Teams agent in a pay-as-you-go environment answered greetings at 1 credit each but returned "It has reached its usage limit" for every SharePoint-grounded question. The moderator's explanation was that grounded generative answers needed a further AI entitlement, and the thread records no confirmed fix. Test grounded questions with an unlicensed account before launch.

Worked Cost: An Internal HR Agent at 8,000 Conversations a Month

An HR policy agent for 2,000 employees, each starting four conversations a month, handles 8,000 conversations. Assume each conversation has three turns, and each turn produces a generative answer grounded in the tenant graph, plus one agent action per conversation (a leave-balance lookup or a ticket). On the rate table that's 3 × 2 + 3 × 10 + 5 = 41 credits per conversation.

  • If no user holds a Copilot licence, the agent uses 328,000 credits a month. Pay-as-you-go costs $3,280. Thirteen packs plus 3,000 credits on the meter costs $2,630, or $0.33 per conversation.
  • With tenant graph grounding off (answers grounded on a SharePoint knowledge source through generative answers only), it uses 11 credits per conversation and 88,000 a month: $880 on pay-as-you-go, or $800 in four packs, which is $0.11 per conversation at the meter rate. Grounding was 73% of the bill.
  • If every user holds a Microsoft 365 Copilot licence, the credit cost is $0, within fair use.
  • For unlicensed users with a reasoning model orchestrating, using Microsoft's illustrative token counts (461 credits for the first message, 141 for each follow-up), a three-turn conversation costs 743 premium credits plus 41 in feature rates, 784 credits or $7.84. At 8,000 conversations that's about $62,700 a month.

The Agentforce equivalent, using Salesforce's Agentforce pricing page at $500 per 100,000 Flex Credits and 20 credits per action, costs $0.10 per action. If the three answers and one lookup each count as an action, that's $0.40 per conversation, or $3,200 a month, before the Salesforce user licences an HR team would need. Salesforce lists no separate token meter on that page. That is close to Copilot Studio's grounded cost. In a Microsoft 365 shop, Copilot Studio needs no new user licences, and with grounding off it costs about a quarter as much.

Worked Cost: A Customer Support Agent at 30,000 Conversations a Month

A website support agent handling 30,000 conversations a month is the workload where the three vendors compete directly. Assume each conversation includes two scripted answers (returns policy, opening hours), three generative answers from product manuals, and two agent actions (order lookup and case creation).

On Copilot Studio with a standard model, 2 × 1 + 3 × 2 + 2 × 5 = 18 credits per conversation, all billable because the users are external. That's 540,000 credits a month: $5,400 on pay-as-you-go, or $4,400 for 22 packs, about $0.15 per conversation. The prepurchase plan's discount of up to 20% would lower it further.

On a reasoning model, with the same three model-answered turns at Microsoft's illustrative token counts, 743 premium credits plus 18 in feature rates is 761 credits, $7.61 per conversation, or about $228,000 a month. Copilot Studio is cheapest on this workload only on a standard model.

On Agentforce, if each of the three generative answers and two actions counts as an action, that's 5 × 20 = 100 Flex Credits, $0.50 per conversation and $15,000 a month. Salesforce also lists a $2 per conversation option on the same pricing page. Agentforce's per-action unit is twice the price of a Copilot Studio agent action at pay-as-you-go ($0.10 versus $0.05), and that gap is why it loses this comparison on price. Our Agentforce pricing teardown shows how the action count per case drives the bill.

For Intercom Fin, Fin's pricing page charges $0.99 per outcome, with one outcome billed per conversation however many actions Fin takes, a 50-outcome monthly minimum when you keep your current help desk, and no charge for a conversation handed to your team without an outcome. Resolutions count as outcomes, and so do procedure handoffs to a human, so the outcome rate runs at or above your resolution rate. If 60% of conversations end in an outcome, that's 18,000 outcomes and $17,820 a month. At 40% it's $11,880. Fin costs more per conversation than Copilot Studio on a standard model, but you only pay when it does something, and you don't need a Power Platform admin to watch a meter. Our comparison of per-resolution pricing across Fin, Zendesk and Agentforce covers what counts as an outcome.

Who should not pick each one? Skip Copilot Studio if nobody in your organization will own the Power Platform admin center's capacity pages; the consumption model punishes an unwatched agent. Skip Agentforce for support if your cases live in Zendesk, Intercom or ServiceNow, because you would pay Salesforce's per-action rate without the data advantage that justifies it. Skip Fin if you need one platform for internal and external agents, since Fin is a support product.


What Happens When Capacity Runs Out Mid-Month

Prepaid credits are pooled across the tenant, and admins allocate them to environments. An environment with no allocation draws from the tenant pool by default, per Microsoft's capacity management guide. When an environment's allocation is spent, it can draw from unallocated tenant capacity, bill to a pay-as-you-go plan, or stop.

Microsoft allows some overage as a grace period. Enforcement starts when the tenant reaches 125% of its prepaid capacity, according to the overage enforcement section of the billing rates page. At that point the environment's agents are blocked, "including in a conversation that's already in progress," and most users see: "This agent is currently unavailable. It has reached its usage limit. Please try again later." Microsoft 365 Copilot licensed users "might not be blocked." The tenant's designated administrator gets an email, and the Power Platform admin center shows a notice.

Three details matter for planning:

  1. One agent can take down the others. Every environment drawing on the tenant pool shares the 125% threshold. A runaway test agent can block a production support agent in another environment. Environments with their own allocation keep running until that allocation is gone.
  2. Agent flows are enforced separately. When an environment has no credits, new agent flow runs are blocked while the agent keeps answering questions that don't need a flow. Flows already running finish normally.
  3. Pay-as-you-go environments aren't blocked for running out of prepaid credits. Usage beyond prepaid capacity bills to the Azure plan instead. That is the right setting for a customer-facing agent, as long as a monthly limit sits on top of it.

Running over is also a licensing matter. Microsoft's licensing page says that if you exceed purchased capacity you should "adjust your quantity per standard Microsoft terms to stay in compliance."

The Controls That Stop a Runaway Bill

Copilot Studio has per-agent monthly limits, and they are the most important control on the platform. In the Power Platform admin center (Licensing, Copilot Studio, Manage Agents), an admin can set a monthly credit limit on any agent, get alerts as it nears the limit, and choose a hard stop that turns the agent off when it hits it, per Microsoft's capacity guide. The same page lists every billable agent in the tenant with month-to-date billed credits and a status of within, nearing or over limit.

The limits have two gaps you need to plan around. They are not applied instantly, so an agent can keep responding for some time after it reaches its limit; Microsoft advises setting limits and notification thresholds "with some margin." And an agent's limit does not stop the agent flows it calls. Those are controlled only by the environment's allocation and billing settings.

The reports are useful but coarse. You can download consumption by environment, agent or user, and the detailed report adds columns for the AI feature, channel, model and knowledge source. The reports do not show usage for a single conversation, token counts, or why a row wasn't billed. To trace a reasoning-model spike, Microsoft's guide says to filter the feature column to "Text and generative AI tools (premium)" and multiply credits by 100 to estimate billed tokens. Pay-as-you-go charges show up in Azure Cost Management.

The Microsoft 365 admin center now has a second control plane. Its Cost management dashboard sets spending policies, user and agent limits, and alerts for Copilot Credits across services like Cowork and the Work IQ API. Copilot Studio policies there are opt-in, and Power Platform admin center controls keep applying until a Microsoft 365 spending policy covers that scope. Spending policies have "Auto-apply new services" on by default, which means new metered services join a policy automatically. Decide whether you want that before you rely on it.

Before you build, run the agent through Microsoft's Copilot Studio agent usage estimator, which models credits from agent type, traffic, orchestration, knowledge and tools. A consultancy writeup from Stoneridge Software in August 2026 makes the practical case: budget the flat licence and the metered services as separate line items, and turn on cost controls before you scale.

Which Choice Predicts Regret

The decisions that cause regret with Copilot Studio are about design, and most are made before anyone looks at the price page.

  • The orchestration model changed the cost per conversation by a factor of 19 to 42 in the examples above. Make it a budget approval.
  • Tenant graph grounding, at 10 credits a turn, was 73% of the HR agent's bill. Turn it on only for agents whose answers measurably improve with it.
  • Who the users are decides coverage. An internal agent for Copilot-licensed staff is close to free. The same agent opened to contractors on Copilot Chat is fully metered.
  • Choose packs or pay-as-you-go per environment: packs for steady internal load above 20,000 credits a month, pay-as-you-go with a monthly limit for anything customer-facing that must not be blocked at 125%.

What changes the answer: if your support conversations need a reasoning model to resolve, Copilot Studio becomes the most expensive of the three: Agentforce's per-action price (Salesforce lists no token meter on its pricing page) and Fin's flat $0.99 per outcome both come in far below $7.61 a conversation. If your cases, customer records and entitlements are already in Service Cloud, Agentforce's higher per-action price buys you data access that would take engineering time to replicate in Copilot Studio.

What to Do Next

This Week:

  1. Open the Power Platform admin center, go to Licensing, Copilot Studio, Manage Agents, and set a monthly credit limit with notifications on every agent already in production.
  2. Download the agent-level consumption report and filter for "Text and generative AI tools (premium)." Any agent showing up there is on a reasoning model; check that someone approved it.

This Month:

  1. Check which environments draw from the tenant pool. Give each customer-facing environment its own allocation and link it to a pay-as-you-go plan, so a test agent can't block it at 125%.
  2. Find the Power Automate flows that use AI Builder prompts or document processing. Their seeded credits stop renewing with your licences, and they will draw on Copilot Credits next.

Before Your Next Microsoft Renewal:

  1. Get the prepurchase plan's tier breakpoints in writing and compare them with twelve months of downloaded consumption.
  2. For any support agent you plan to build, run the usage estimator on both a standard and a reasoning model, and price the same volume on Fin and Agentforce using the numbers above.

The Bottom Line

For customer support, Copilot Studio is the cheapest of the three platforms per conversation by a wide margin, as long as the agent runs on a standard model and its users are either Copilot-licensed employees or customers on an agent with a hard monthly limit. The pricing resembles early cloud compute: unit prices are low and published, and the bill comes from what nobody turned off. The reasoning-model meter is the line most budgets miss, and Microsoft's own example prices a single new-conversation message at 461 credits.

Set the per-agent limit before the agent goes live.

Continue Reading

Share:

Frequently Asked Questions

How much does a Copilot Credit cost in Copilot Studio?

As of November 16, 2026, Microsoft lists a Copilot Credit pack at $200 per month for 25,000 credits ($0.008 each) and the pay-as-you-go meter at $0.01 per credit billed to Azure. A one-year prepurchase plan saves up to 20%, with tiers in the licensing guide.

How many Copilot Credits does one conversation use?

It depends on the features each turn triggers: a classic answer is 1 credit, a generative answer 2, an agent action 5 and tenant graph grounding 10. A three-turn grounded HR conversation with one action uses about 41 credits. Reasoning models add 10 credits per 1,000 billed tokens, which can reach hundreds of credits per message.

Does a Microsoft 365 Copilot licence cover Copilot Studio agents?

Yes, for employee-facing agents used by a licensed person under their own authenticated identity, within fair-use limits. It does not cover unlicensed employees, external users, autonomous agents started by triggers, agent flows with other triggers, or computer-using agents.

What happens when Copilot Studio capacity runs out?

Microsoft allows some overage, then blocks agents when the tenant reaches 125% of prepaid capacity, including mid-conversation. Environments linked to a pay-as-you-go plan keep running and bill the overage to Azure. Unused prepaid credits do not roll over to the next month.

How do I cap Copilot Studio spending per agent?

In the Power Platform admin center, go to Licensing, Copilot Studio, Manage Agents and set a monthly credit limit with notifications and an optional hard stop. Limits are not applied instantly and do not stop the agent flows an agent calls, so leave margin.

Newsletter

More on AI ROI & Cost, every Sunday

The week's enterprise AI news, AI ROI & Cost included, in one Sunday email.

Weekly, free. Unsubscribe any time. Privacy

Latest Articles

View All →