Understanding Copilot Credits, Consumption-Based Billing, and the Real Business Case for AI
One of the most common questions I hear after discussing Business Central’s Sales Order Agent, Payables Agent, Expense Agent, or other AI-powered capabilities is surprisingly simple: “This sounds interesting, but what does it cost?” Fair question.
For years, software pricing was relatively predictable. You purchased licenses, implemented the application, trained your users, and paid a recurring subscription. Organizations generally knew what their software expense would look like next month and next year.
AI introduces a different model. Instead of paying only for access to software, organizations may also pay based on how much certain AI capabilities are used. That is where terms such as Copilot Credits, consumption-based billing, and occasionally tokens enter the conversation.
For finance leaders, this creates understandable concern. Nobody wants a surprise bill. Nobody wants to approve an investment without understanding how the charges are calculated. CFOs and Controllers want to see the assumptions, the math, and the business case.
The purpose of this article is not to convince anyone that AI automatically pays for itself. The purpose is to explain how Microsoft measures AI consumption, put realistic numbers around actual Business Central scenarios, and provide a framework for determining whether the expected benefit justifies the expected investment.
Tokens, Copilot Credits, and What Customers Actually Pay For
Behind the scenes, large language models process information using units called tokens.
A token is a small piece of text the AI model processes. When someone asks Copilot a question, uploads a receipt, submits a vendor invoice, or asks an AI agent to complete a task, tokens are used behind the scenes as the model reads information and generates a response.
Fortunately, Business Central customers do not need to track tokens.
Instead, Microsoft uses Copilot Credits as the customer-facing unit of consumption. Business Central’s billable AI agents consume Copilot Credits as they perform work. The number of credits consumed depends on what the agent does and how often it is used. Microsoft currently supports both Prepaid Copilot Credit capacity and Pay-as-you-go billing. When both exist, prepaid capacity is consumed first.
For illustration purposes, Microsoft currently publishes a pay-as-you-go Copilot Credit price of $0.01 per credit, although actual effective pricing can vary depending on agreements, region, purchasing arrangements, and future pricing changes. Keep in mind the distinction between tokens (how AI models operate) and Copilot Credits (how Microsoft measures and bills consumption).
How Much Does the Business Central Expense Agent Cost?
Microsoft currently assigns 50 Copilot Credits per receipt processed by Expense Agent, making estimated consumption easy to calculate: Receipts multiplied by 50 Copilot Credits.
Expense Agent is the easiest place to start because the rule is straightforward. Microsoft currently assigns 50 Copilot Credits per receipt processed, regardless of how many expense lines it creates from that receipt. That gives us a simple planning formula:
Receipts × 50 Copilot Credits = Estimated Expense Agent Consumption
The agent does much more than read a receipt. It extracts information, categorizes expenses, supports itemization, groups expenses into reports, validates policies, and prepares the resulting expense documentation for review and approval. Microsoft abstracts that complexity into a predictable, receipt-level consumption rule.
A Practical Expense Management Example
Imagine an organization with:
- 25 employees who regularly submit expenses
- 10 receipts per employee per month
That produces:
- 250 receipts per month
- 50 credits per receipt
Monthly consumption:
250 × 50 = 12,500 Copilot Credits
At $0.01 per credit:
12,500 × $0.01 = $125 per month
Annualized:
$125 × 12 = $1,500 per year
Now we know the estimated cost.
The next question is whether the organization receives more value than it spends.
Suppose employees spend five minutes preparing each receipt and finance spends another three minutes reviewing and processing it.
That equals eight total labor minutes per receipt.
250 receipts × 8 minutes = 2,000 minutes
2,000 ÷ 60 = 33.3 labor hours per month
If Expense Agent reduces that effort by a conservative 50%, the organization recovers:
33.3 × 50% = 16.7 hours per month
At a fully burdened labor rate of $40 per hour:
16.7 × $40 = approximately $668 per month
Comparing the two sides:
- Monthly labor value: $668
- Monthly AI cost: $125
- Net monthly benefit: $543
- Net annual benefit: $6,516
This does not prove every organization will achieve the same result. It simply demonstrates the math. Since the assumptions are visible, a finance leader can substitute actual receipt volume, labor effort, labor cost, and expected efficiency gains to create their own business case.
How Much Does Payables Agent Cost?
Payables Agent currently uses 50 Copilot Credits to process an invoice plus 5 Copilot Credits for each invoice line, allowing organizations to estimate consumption based on invoice volume and average invoice lines.
Microsoft recently published detailed Payables Agent consumption examples, making cost forecasting much easier. The published model currently assigns 50 Copilot Credits to process an invoice and 5 Copilot Credits per invoice line. The planning formula becomes:
Invoices × [50 + (5 × Average Invoice Lines)]
For example:
- 100 invoices per month
- 3 invoice lines per invoice
Consumption:
100 × [50 + (5 × 3)]
100 × 65
= 6,500 Copilot Credits per month
At $0.01 per credit:
6,500 × $0.01 = $65 per month
Microsoft’s published example shows something many organizations find surprising: AI consumption costs are often considerably lower than the labor cost of manually processing invoices.
How Much Does Sales Order Agent Cost?
Sales Order Agent consumption varies by the activities performed during an interaction, with Microsoft’s published examples assigning different Copilot Credit amounts to activities such as analyzing emails, checking item availability, creating orders, and generating responses.
Sales Order Agent follows a different model because several different activities can occur during an interaction. Microsoft publishes the following examples:
| Activity | Credits |
| Analyze incoming email | 2 |
| Check item availability | 5 |
| Create or update quote/order | 5 |
| Generate response | 2 |
A typical quote or order workflow therefore consumes 14 Copilot Credits, not including attachment processing. Microsoft’s example assumes:
- 100 customer requests per month
- 50% include a usable attachment such as a purchase order
The resulting calculation is:
[14 + (5 × 50%)] × 100 = 1,650 Copilot Credits per month
At $0.01 per credit: 1,650 × $0.01 = $16.50 per month
Again, the key point isn’t the credit count itself. The key point is that the cost of AI is often far smaller than the manual effort it replaces.
Does AI Cost More Than the Labor It Replaces?
Not necessarily. In many Business Central scenarios, AI consumption costs can be considerably lower than the value of employee time recovered through automation.
The labor savings examples above intentionally focus on measurable capacity. They do not assign a dollar value to faster reimbursements, faster invoice processing, fewer data-entry errors, improved compliance, better vendor relationships, improved customer responsiveness, and better employee experience. Those benefits may be meaningful, but you should not include them in a ROI model without supporting measurements. It is also important to recognize what recovered capacity means. Reducing effort by ten or twenty hours per month does not automatically reduce payroll expense. More commonly, organizations redirect that time into higher-value work.
Finance teams spend less time typing and correcting transactions and more time analyzing results, strengthening controls, and supporting decision-making. Customer service teams spend less time managing routine order administration and more time serving customers. The objective is not headcount reduction. The objective is to increase productive capacity.
How Should You Calculate the ROI of Business Central AI?
The best AI business cases start with the organization’s existing process, measuring transaction volume, labor effort, bottlenecks, rework, delays, and current costs before comparing those numbers with projected AI consumption.
The best business cases begin with the current process. Ask questions such as:
- How many transactions occur each month?
- How much time does each transaction require today?
- Where are the bottlenecks?
- How often does rework occur?
- What delays create downstream business impact?
- How much labor is involved?
Once a baseline exists, the math becomes straightforward:
Annual Benefit = Annual Hours Recovered × Fully Burdened Labor Rate
Annual AI Cost = Projected Copilot Credits × Applicable Credit Rate
Net Annual Benefit = Annual Benefit − Annual AI Cost
ROI = Net Annual Benefit ÷ Annual AI Cost
This is the same financial discipline organizations apply when evaluating software, equipment, outsourced services, automation initiatives, or additional headcount. AI should be no different.
What Is the Right Financial Question to Ask About AI?
The right question is not simply how much AI costs, but what the organization is spending and what business value it receives in return.
“How much does it cost?” is not the wrong question. It is an essential question, but it’s incomplete without the second half of the equation. The better question is “What are we spending, and what are we getting back?” The organizations deriving the greatest value from Business Central AI will not necessarily be the organizations consuming the fewest Copilot Credits. They will be the organizations that can demonstrate the greatest business value from the credits they consume.
Not Sure Where to Start?
At . Understanding Copilot Credits, consumption-based billing, and Microsoft’s AI architecture is important. Understanding business processes, current effort, baseline performance, governance, and measurable outcomes matters just as much.
The objective is not to deploy AI because it is new. The objective is to identify opportunities where AI can improve productivity, reduce operational friction, strengthen customer experiences, and support meaningful business outcomes.
Once the assumptions and the math are visible, the AI conversation becomes much less mysterious. It becomes a business decision.