People are not good at understanding consumption models. We want a price tag. Tell us what something costs and we can budget for it. Ask us to work out what it will cost if this happens, then that, then a few more things after that, and most of us give up and guess.
That’s the problem IT teams are walking into with AI. It looks new, but it isn’t. It’s the same problem the industry ran into when cloud first became available, and most of what we learned then applies almost unchanged.
Seats Were Easy. Work Isn’t.
For most of the history of enterprise software, budgeting meant counting: users, devices, servers, licenses. You knew how many people you had, so you knew roughly what you’d spend.
AI agents don’t fit that model. They’re metered by the task, not the seat. Activity grows with adoption rather than headcount, and the charges can land directly on an Azure subscription. In Microsoft’s environment the unit is the Copilot Credit, and what a task costs depends on the model it uses, how much organizational context it pulls in, how many tools it calls, and how long it runs. In practice, every round of back-and-forth between a person and the agent turns into credits.
It helps to think about AI spend in three layers.
- Licensing pays for access. These are per-user costs, and they’re the part that looks like traditional software.
- Consumption credits pay for work performed. This is the variable layer. It changes with the type of task and how often people ask for it.
- Azure resources are the infrastructure underneath. The applications, AI services, and networking are still there, and they still cost money.
Most budgets account for the first layer. The second is where the surprises come from, and the third is easy to forget because it was already on the bill.
Capacity Versus Consumption
There are two common ways to pay for cloud services, and AI uses the less familiar one. With capacity-based products like Microsoft Fabric, you buy a set amount of capacity and plan your work within it. The budget is decided up front.
Consumption works differently. You pay for what gets used, which is what makes it flexible: a team can start small and grow without buying capacity it doesn’t need yet. The trade-off is that there’s no natural ceiling. An organization that turns it on without a plan can use far more than it expected before anyone notices.
That’s why the controls matter, and Microsoft has built them in. Its cost management view for Copilot Credits shows administrators where credits are going. It shares a name with the Azure tool cloud teams already use to track spend, and it applies the same approach to a new meter.
The Decision that Belongs to the Business
The controls themselves are straightforward. Credits are managed across the organization, and administrators can carve them out by department or business line and set limits for groups and for individual users.
The important choice is what happens when a group hits its limit. One option is a hard stop: the credits run out and the AI capabilities stop working until the next period. The other is overflow, where work continues and the extra use is billed as it happens.
Neither is wrong. A hard stop protects the budget and frustrates people in the middle of a task. Overflow keeps work moving and puts the forecast at risk. That’s a business decision, not a technical one, and it should be made on purpose before launch rather than discovered on the first invoice. Very small companies sometimes set limits per user, but for most organizations the department is the right level.
This visibility is built for IT, not for end users. Its job is to let IT report on spend and manage compliance against it.
Pilot Small Enough to Do the Math
When an organization has no idea what a new Microsoft technology will cost, the best practice is the same every time: test it, then scale it.
Start with a small group of pilot users who are comfortable with IT and who represent the departments and lines of business that will actually use the capability. Let them work. Watch the costs. Then open it up gradually.
The point of starting small is that it gives you numbers to multiply. If one person in HR used 1,000 credits in a month, five people in HR will likely use around 5,000. Repeat that across the business and you can forecast how many credits you need before you commit to them.
Microsoft’s guidance helps with the first estimate. It sorts agent work into three tiers by how many sources a task draws on, how much reasoning it needs, and how many outputs it produces. A light task, like a recurring weekly status update built from someone’s calendar, typically uses 100 to 300 credits. A medium task, like a customer meeting briefing that pulls together emails, files, and a presentation, uses 300 to 700. A heavy one, like analyzing six months of usage data into a leadership-ready report, can use more than 700. Those are starting points, not quotes. The real number depends on the environment and on how much back-and-forth the task needs.
The alternative is what got organizations into trouble with cloud. People are asking for AI, so turn it on and see what happens. Some run through their budget in the first few weeks. Others set aside a sum of money, release it, and hope. Either way, the first real forecast is the bill.
Governance Is the Real Answer
The biggest mistake in rolling out AI is really two mistakes at once: cost and security. They share a root cause, which is uncontrolled adoption, and they share a fix, which is governance.
Cost governance is everything above: visibility, accountability, and controls set before launch. The other half is data governance.
AI works with the access each person already has. In a lot of organizations, especially smaller ones, sensitive files sit in places with loose permissions. Someone in accounting might never think to open the spreadsheet with everyone’s salary in it. Ask an AI assistant what the CEO makes, though, and it may go and find it, because technically that person can reach it.
The external version of the same risk is sensitive data leaving the company through public AI tools. Governance means deciding which AI solution is sanctioned and enforcing that decision, with data loss prevention policies in tools like Microsoft Purview that classify sensitive data and block it from being shared or uploaded where it shouldn’t go. That’s why the AI readiness workshops Microsoft and its partners deliver are built around enabling AI in a secure environment first.
Sensitivity labels and auto-labeling help content carry its own classification wherever it travels.
Assistants inherit the user’s existing permissions. Tighten access and site permissions before broad rollout, not after.
Purview DLP policies can restrict labeled content from being processed or summarized in Copilot and control what leaves the tenant.
Audit, activity reporting, and eDiscovery coverage extend across prompts and responses, applying the same expectations as any other data.
Extend existing insider risk, retention, and compliance policies to AI activity rather than standing up a parallel program.
None of this requires new thinking. It requires the discipline cloud already taught us: know what you’re consuming, decide who owns the number, and start small enough that you can do the math.
Connection’s cloud cost optimization team helps organizations see where their cloud spend is going across Azure, AWS, and Google Cloud, and build a plan to manage it. Learn more about Managed Cloud Cost Optimization.
