Copilot Studio Pricing: What a Small Business Customer FAQ Agent Costs

A customer FAQ agent in Copilot Studio typically costs around 7 to 10 Copilot Credits per conversation, which is $0.07 to $0.10 on pay-as-you-go at $0.01 a credit. For 900 conversations a month that is roughly $63 to $90. A $200 pack of 25,000 credits only beats pay-as-you-go above about 20,000 credits a month.

Those figures rest on how Microsoft meters agent work: a scripted answer, a generated answer and an action each cost a different number of credits, and one conversation mixes them. So the useful question is not "what does Copilot Studio cost" but "how many credits does my conversation burn, and how many conversations do I get". Answer those two, and the choice between the three ways to pay mostly makes itself.

Follow me on Instagram@sagnikteaches

The price list in credits and dollars

Microsoft's billing rates page, updated on 3 November 2026, gives the unit costs. The dollar column below uses the pay-as-you-go rate of $0.01 a credit that Microsoft's Copilot Studio licensing guide (November 2026) and Azure's published price list both show.

Connect on LinkedInSagnik Bhattacharya

Microsoft's page says one interaction can use several of these at once. Its example is an agent grounded in a tenant graph using 12 credits for one complex prompt: 10 for grounding and 2 for the generative answer. For a customer FAQ, the two rows that matter are the first two. Everything else is a way for the bill to creep up.

Subscribe on YouTube@sagnikteaches

What staff using it do not pay, and why customers do

The rate table has a column for people licensed with Microsoft 365 Copilot, and for most rows it says no charge. That applies to employee-facing use in which the user is signed in with the licence. A customer on your website is not signed in with your licence, so none of that exemption applies. The Microsoft 365 Copilot licence is listed at $30 a user a month on the Copilot Studio pricing page, and it covers staff building and using agents inside Microsoft 365. The same page reserves publishing to websites and use by people without a licence for the standalone Copilot Studio route. Buying more Copilot seats does nothing for a public FAQ agent.

That is worth saying plainly because owners sometimes believe that "we already pay for Copilot" covers a website agent. It covers your staff asking an internal agent. Your customers are metered.

Main worked month: a plumbing firm's FAQ agent

Illustration: a plumbing firm with a website agent that answers questions about call-out areas, emergency hours, what a boiler service includes and how to send photos of a leak. The owner wants a number for one month.

First, the conversation shape. The firm writes eight fixed replies as classic answers (hours, prices, photo instructions) and lets the AI answer the rest from a short document. An average conversation turns out to use:

• 4 classic answers, 4 credits.
• 2 generative answers, 4 credits.
• Total: 8 credits, which is $0.08.

Microsoft's own customer support example uses the same pattern, four classic answers and two generative answers per run, which is a sign the mix is typical, not exotic. Now the volume across three kinds of month:

The pattern is the point. In quiet and typical months the pack costs roughly three to eight times as much. Only in the busy month does it win, by $40. And you cannot buy a pack just for winter: Microsoft's licensing guide lists packs at $200 a month billed annually, so one pack is a $2,400 commitment for the year. A firm with a seasonal business should lean to pay-as-you-go, and check the worst month.

Where the break-even sits

The sum is short. A pack costs $200 and holds 25,000 credits. Pay-as-you-go costs $0.01 a credit, so $200 buys 20,000 credits on that meter. The break-even is 20,000 credits a month. At 8 credits per conversation that is 2,500 conversations; at 10 credits it is 2,000; at 6 credits it is 3,333. Below the figure for your own conversation shape, use pay-as-you-go.

A pack is a tenant-wide licence, so the credits are pooled across the whole tenant, and you assign them to environments. If you buy one and then use less than a third of it, you have paid for the rest, and you cannot bank it: the licensing guide says purchased capacity is enforced monthly and unused Copilot Credits do not carry over from month to month. Microsoft's capacity page makes the same point from the other side, listing "no more unused capacity expires at the end of the month" as a benefit of pay-as-you-go.

The third route: pre-purchase

Microsoft's pricing page describes a pre-purchase plan and says it saves up to 20 per cent, with automatic pay-as-you-go cover when prepaid credits run out. The licensing guide gives the detail: the Copilot Credit Pre-Purchase Plan is a one-year, pay-up-front pool of credits, unused credits expire at the end of the annual term, and the discount rises with the size of the pool. The smallest tier is 300,000 credits at 5 per cent off; the 20 per cent figure belongs to the largest tiers, measured in hundreds of millions of credits. A separate Agent Pre-Purchase Plan, shared with Microsoft Foundry, starts at 20,000 commit units, each worth $1 or 100 credits.

Run the small-firm sum. The plumbing firm's typical 7,200 credits a month is 86,400 a year. The smallest pre-purchase tier is 300,000 credits, which at $0.01 less 5 per cent is $2,850 paid up front, about three and a half years of the firm's usage, and anything unused at the end of the year is lost. So pre-purchase is a tool for organisations already spending thousands a month on agents. For a small firm with one FAQ agent, it is not a sensible route.

What happens when prepaid credits run out

This is the difference most owners miss between the routes. Microsoft's billing page says that for prepaid capacity, enforcement starts when a tenant reaches 125 per cent of its prepaid capacity. At that point the environment's agents are blocked: an agent stops responding, even in a conversation that is already running, and users see a message that the agent has reached its usage limit. For a website agent that means customers see an error in a cold snap.

If the environment bills overage to a pay-as-you-go plan, running out of prepaid credits does not block the agent. The usage the pack does not cover goes to the plan. So a common set-up is a pack for the base load and pay-as-you-go for the overflow. For a small firm the simpler version is pay-as-you-go alone, with an agent limit as the safety.

Fixed and one-off costs outside the meter

The credit meter is not the whole cost. Add these before you decide whether an agent is worth it.

• Build time. Illustratively, six hours to write the replies, load the FAQ document and test. At an owner's time valued at $50 an hour that is $300, once.
• Upkeep. Thirty minutes a week to read the transcripts and fix wrong answers. Over a month that is two hours, $100 of time. This is the cost that most often goes unbudgeted, and it is the one that keeps an agent accurate.
• An Azure subscription. Pay-as-you-go requires one linked through a billing plan. The setup is a one-off, covered in the Azure budget setup tutorial.
• A trial before spending. Microsoft's licensing page says a trial lets you build and test but not publish, so you can price your conversation shape in the test chat before paying for live traffic.

Cost per conversation against cost per enquiry handled

A more useful comparison than the credit price is what an enquiry costs you to handle without the agent. Illustration, with assumptions you should replace: a typical customer question by phone or email costs about five minutes of staff time, and staff time is worth $24 an hour, which is $2 an enquiry. If the agent resolves half of its 900 monthly conversations without a human, it saves about 450 × $2 = $900 of time, against $72 of credits and $100 of upkeep time. Even if the real resolution rate is a fifth, 180 × $2 = $360 still clears the cost.

The weakness is the assumption. If your enquiries are mostly "can you come on Tuesday" bookings that need a person, the agent resolves few of them, and the sum collapses. Read two weeks of real enquiries before you use my arithmetic, and count how many were questions with a fixed answer.

The reasoning-model bill nobody saw in testing

A pattern worth knowing. A firm tests its agent with a reasoning model because the replies read better. Nothing looks wrong in the test chat. Microsoft's billing page works through a similar case: a customer follow-up in an ongoing conversation, with the agent calling the model twice and 14,100 billed tokens, comes to 141 credits for the model charge alone, plus the normal feature rates on top. At $0.01 a credit that is over $1.40 for one message. If the firm's 900 conversations average four such messages, the month is $5,000, not $72. The numbers in Microsoft's example are illustrative and token counts depend on instructions, knowledge and tools, but the order of magnitude is the lesson: ask for a standard model unless the job truly needs reasoning, and check the Product or AI feature column in the usage report for the premium text and generative AI tools line.

Cutting the credits per conversation

1. Move the top ten questions to classic answers. A classic answer costs 1 credit and a generative one costs 2, but the saving is also in consistency: no wrong prices.
2. Leave tenant graph grounding off for a public FAQ agent. It costs 10 credits a message and your FAQ lives in one document, not across the firm's mailboxes.
3. Keep instructions short and knowledge focused. Microsoft's own tips for reducing consumption include concise instructions, focused knowledge sources and new conversations for new topics.
4. Cap the conversation. End after six exchanges with a hand-off to a phone number or form.
5. Set an agent limit. Microsoft lets admins set monthly credit limits per agent, with notifications and a hard stop.

Before and after for one reply, illustrative: the generative answer "What areas do you cover?" costs 2 credits each time and may vary. A classic answer with the area list and a booking form costs 1 credit and says the same thing every time. Across 900 conversations where 30 per cent ask about areas, that is 270 questions, and the saving is 270 credits, or $2.70. Small money, but the consistency is the real gain.

Voice is a different price

If you were hoping to put the agent on a phone line, Microsoft's page prices voice per minute: classic voice at 10 credits, generative voice at 35 and premium generative voice at 75 per minute, with core agent activity included. At $0.01 a credit that is $0.10, $0.35 and $0.75 a minute. A 3-minute call on the generative tier is about $1.05, which makes the maths against a person very different from a text chat. Price voice separately and do not assume the FAQ figures carry over.

Cheaper ways to answer the same questions

An agent is not the cheapest answer to "customers keep asking the same thing". Try these first, in order of cost.

1. A good FAQ page. Free, indexed by search, and where the agent's knowledge would come from anyway. Start with building the FAQ your chatbot needs; a bot is only as good as that document.
2. Saved replies in the shared inbox. Ten templates, a person sends them. A minute of staff time per reply, with a human check.
3. A classic-only agent. Every answer is written by you at 1 credit. For 900 conversations of 6 classic answers that is 5,400 credits, $54, and nothing can be wrong that you did not write.
4. A different tool. If you need automations beyond chat, Copilot Studio against n8n compares the costs, and training a chatbot on your policies and prices covers non-Microsoft options in general.

Whichever route you take, read the general guide to AI pricing by seat, credit and usage once, because the same meter logic shows up in most tools.

Two other agents, priced the same way

The FAQ shape above is the cheapest kind. Two other illustrative agents show how fast the meter moves when the job changes.

An insurance broker's policy-wording agent

The broker loads policy wordings and wants customers to ask what is covered. Most replies must be generative, because nobody can pre-write every question, and the owner is tempted to turn on tenant graph grounding so the agent can search the whole firm's content. Per conversation: five generative answers (10 credits) plus grounding on each of five messages (50 credits) is 60 credits, or $0.60. At 900 conversations that is $540 a month, compared with $90 for a document-only agent with the same five generative answers and no grounding at 10 credits per conversation. The grounding is a tenfold difference for a task a single uploaded document handles. A wording-heavy agent also has a risk to weigh against price: a wrong answer about cover is worse than no answer, so keep it to quoting the wording and handing off for advice.

An electrical contractor's quote-request agent

This agent asks six questions, books a survey slot and emails the office. Microsoft counts steps such as triggers and topic changes as agent actions at 5 credits each. Illustratively: six classic answers (6 credits), two generative answers (4 credits), and four agent actions (20 credits) for the booking and email is 30 credits, $0.30 a conversation. Actions dominate the bill, not the chat. If the booking also calls an agent flow, Microsoft prices flow actions at 13 credits per 100 actions, which is small per run but is billed and capped separately from the agent. At 250 enquiries a month the agent costs $75, and the real question is whether it replaces a half hour of phone tag each time.

Questions to put to a reseller before you buy anything

1. What is the term on a credit pack? The licensing guide says packs are billed annually and unused credits do not roll over; ask whether you can reduce the number of packs before the term ends.
2. What happens to the agent at 125 per cent of prepaid capacity, and can overage go to pay-as-you-go?
3. What is the pay-as-you-go rate per credit today, and does it include any partner margin?
4. Is the smallest pre-purchase tier still 300,000 credits a year, and what discount applies to it?
5. Who receives the usage notifications, and can I add a second person?

Take the answers to the costing sheet below and the figures come out of guesswork and onto a page you can defend to a partner or a bookkeeper.

A one-page costing sheet you can fill in

Copilot Studio FAQ agent costing sheet
Conversations a month (quiet / typical / busy): ____ / ____ / ____
Classic answers per conversation: ____   x 1 credit  = ____
Generative answers per conversation: ____ x 2 credits = ____
Agent actions per conversation: ____      x 5 credits = ____
Credits per conversation: ____
Monthly credits (typical): ____          x $0.01 = $____
Monthly credits (busy):    ____          x $0.01 = $____
Break-even vs $200 pack: 20,000 / credits per conversation = ____ conversations
Build hours ____ x hourly value $____ = $____ one-off
Weekly upkeep hours ____ x 4 x $____ = $____ a month
Enquiries resolved without a person (honest guess): ____ %
Staff minutes saved per resolved enquiry: ____
Agent limit to set (credits): ____   Azure budget: $____

Fill it in with two weeks of real enquiries beside you. If the typical month is well below the break-even and the saved staff time clearly beats credits plus upkeep, build it on pay-as-you-go with an agent limit. If not, the FAQ page and saved replies are the better first step, and you can revisit the agent when the volume justifies it. If you would like someone to run these numbers on your own enquiries, that is the sort of job my AI implementation consultation covers. Prices and rates change, so confirm the current rate table on Microsoft's pages before you commit.

Copilot Studio cost questions for small firms

Does a Microsoft 365 Copilot licence make customer chats free?

No. Microsoft's rate table shows no charge only when the person using the agent is signed in with a Microsoft 365 Copilot licence, which describes staff, not customers. A public website agent talks to unlicensed visitors, so every answer consumes credits.

Is a $200 credit pack always cheaper than pay-as-you-go?

Only if you use most of it every month. A pack gives 25,000 credits for $200 a month, about $0.008 each, while pay-as-you-go is $0.01 each. Below about 20,000 credits a month the pack costs more, and Microsoft's licensing guide says unused pack credits do not roll over to the next month and packs are billed annually.

What stops a runaway bill?

Set a monthly limit on the agent in the Power Platform admin center and an Azure budget for alerts. Microsoft notes budgets only warn, while an agent limit can turn the agent off. Limits are not instant, so leave some margin below your real ceiling.

How do I forecast credits before I build?

Count expected conversations a month, estimate classic and generative answers per conversation, and multiply by the credit rates. Microsoft also publishes an agent usage estimator you can use to cross-check your sums before you commit to a pack.

Further reads

• What Is Copilot Studio and Does a Small Business Need It? — Whether a small business needs Copilot Studio at all.
• AI Chatbot for Accountants: Website FAQs Without Giving Tax Advice — A website chatbot that avoids giving advice.
• Does Microsoft 365 Copilot Keep Your Business Data Private? — How Microsoft treats data used by Copilot.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Microsoft Copilot Studio pricing page, Microsoft Copilot Studio Licensing Guide (November 2026), Microsoft Learn Copilot Studio billing rates page (updated 3 November 2026), Microsoft Learn Copilot Studio licensing and capacity management pages, Azure retail prices list for the pay-as-you-go meter.