Craig Stanley
Home / Cost / Budgets / Budgeting for an agent before it goes live

Budgeting for an agent before it goes live

How I estimate an agent's monthly cost from its design before launch, using Microsoft's Copilot Credit rates and a range instead of one figure.

· 3 min read · Craig Stanley
In short, explained

Before a new helper starts work, guess how many jobs it will do and how much each job costs. Then you know roughly what to save up.

You can estimate an agent's cost before it launches. Count how many conversations or tasks it will handle, work out what each one uses from the price list, and give a low, likely and high figure. Then set a cap and check the real numbers after a month.

Estimate pre-launch agent cost bottom-up: events per run by billing type (classic, generative, grounding, actions, flows, AI tools) times Microsoft's Copilot Credit rates times run volume, split by licensed versus unlicensed users. Publish a low, likely and high case and set the cap at the high case.

Why do it before launch

An agent's cost is decided mostly at design time. Whether it grounds every answer in the tenant graph, how many actions it takes per run, and whether it runs on a trigger or waits to be asked all set the bill before anyone uses it. Once it's live and popular, changing the design is harder.

So before an agent goes live I write down what one run looks like, then multiply.

Step 1: describe one run

List what happens in a typical run in the terms Microsoft bills. For Copilot Studio agents, Microsoft's billing page sets these rates in Copilot Credits:

EventCopilot Credits
Classic answer (authored response)1
Generative answer2
Agent action5
Tenant graph grounding10
Agent flow actions, per 100 actions13

There are separate rates for AI tools and voice, which I'd look up if the agent uses them. Pay-as-you-go credits cost $0.01 each.

Step 2: split by who uses it

This step is easy to miss. Microsoft's billing page says employee-facing use is included at no charge when the person using the agent has a Microsoft 365 Copilot licence and the agent runs under their identity. Unlicensed users with Copilot Chat consume credits for the same agent. So a team that's half licensed pays roughly half what the rate card suggests.

Flows started by other triggers are billed at the standard rate whoever is involved, so an agent that runs on a schedule costs credits regardless of licences.

Step 3: give a range

These numbers are illustrative.

An HR policy agent answers questions grounded in the tenant graph. A typical conversation has three questions, each using tenant graph grounding and a generative answer: 3 × (10 + 2) = 36 credits.

CaseConversations a monthShare unlicensedCreditsCost at $0.01
Low1,00050%18,000$180
Likely2,50060%54,000$540
High5,00070%126,000$1,260

I'd set the agent's cap at the high case and an alert at the likely case. If the first month lands near the low case, I'd lower the cap.

Step 4: decide how to pay

If the likely case is steady, a prepaid pack may suit. Microsoft lists a Copilot Studio pack at $200 a month for 25,000 credits, billed annually, and unused credits don't roll over. The likely case above is just over two packs. The high case would spill into pay-as-you-go, which Microsoft recommends setting up alongside packs to avoid interruption. For more on that choice, see Seat licence or Copilot Credits: the break-even.

Things that blow the estimate

The biggest surprises I'd expect come from triggers. An agent that runs on every new email or every new record can run thousands of times without anyone asking it a question. Reasoning models add a premium rate per 1,000 tokens on top of the feature rate, according to Microsoft's billing page, so switching an agent to a reasoning model changes the sum. Microsoft also publishes an agent usage estimator, which I'd use to check my own arithmetic.

What I'm still checking

I don't yet know how well a design-time estimate holds up for agents with open-ended conversations, where the number of questions per conversation varies a lot. I'd like a month of real logs from a pilot before trusting any range for those.

Sources

Read next

A question to take awayIf you moved last week's unused AI allowance to the people who ran out, who would get it?

About me

Craig Stanley

Microsoft AI consultant and technical architect, based in Whitley Bay. Over the last few years I've delivered Microsoft 365 Copilot, Copilot Studio agents, Microsoft Foundry (formerly Azure AI Foundry) work and governance for UK public sector and financial services organisations.

What interests me is the decision underneath the tool: what it costs, what it risks, and whether a small, transparent model can make it better. I write the methods up here and on Substack so anyone can use them.

I write this site to learn in public: explaining each idea simply is how I check I understand it. Why I write this site.

Find me