Craig Stanley
Home / Cost / Budgets / Budgets that follow teams and agents

Budgets that follow teams and agents

Why I set AI budgets per team and per agent, how Microsoft's spending policies and agent limits support that, and a worked example.

· 3 min read · Craig Stanley
In short, explained

Instead of one big pot of money for all the computer helpers, each team and each helper gets its own smaller pot. Then you can see who's using what.

A single company-wide AI budget hides where the money goes. Giving each team and each agent its own budget makes spending visible, lets the people closest to the work manage it, and stops one busy agent eating everyone else's share.

Structure AI budgets around accountable units: a team allowance for seat-plus-metered use and a separate cap per production agent. Map them to Microsoft 365 spending policies (per group and per user limits) and per-agent Copilot Credit limits in the Power Platform admin center, with alerts below the cap.

The problem with one pot

Most organisations I've seen start with one line in the budget: "AI". Licences go in, the Azure bill goes in, and someone in finance watches the total. That works until the total jumps and nobody can say why.

AI spend has two shapes. Licences are fixed per person. Metered use, paid in Copilot Credits or Azure consumption, moves with what people and agents actually do. A single pot mixes the two, so a runaway agent and a busy month-end look the same on the bill.

Two kinds of budget

I find it clearer to set budgets around the people who can change the spending.

A team budget covers the seats the team holds and the metered use its people drive, such as Cowork tasks. The team lead owns it, because they can see why use went up.

An agent budget covers one agent in production. The agent's owner holds it, because they can change its design, its knowledge sources or its triggers, which are what drive its cost.

What Microsoft gives you to enforce them

Two sets of controls line up with this split, and both are documented by Microsoft.

For Copilot experiences billed by usage, such as Cowork, the Microsoft 365 admin center has spending policies under Copilot and Cost Management. A policy names a group of users, the services it covers, a policy-level limit and per-user limits, with alerts as people approach them. Microsoft's documentation makes one point I'd underline: a spending policy is also the access control. Anyone in scope of a policy that selects Cowork can use it, however small their limit. So the group membership is part of the budget decision.

For agents built in Copilot Studio, the Power Platform admin center lets you set a monthly Copilot Credit limit for an individual agent. Microsoft's billing page also says that if a tenant runs on prepaid capacity and goes to 125% of it, custom agents are disabled until capacity is added. An agent cap set well below that point is a gentler way to find out.

Azure costs for models in Foundry sit in Microsoft Cost Management, where budgets can be set on a subscription or resource group. Those budgets send alerts. They don't stop the spending on their own.

A worked example

These figures are illustrative.

A 40-person operations team has 15 Copilot licences and access to Cowork for 10 people. It also runs two agents: a supplier query agent and an internal policy agent.

BudgetOwnerMonthly capAlert at
Ops team seatsTeam lead15 seats (fixed)Fewer than 10 active users
Ops team CoworkTeam lead60,000 credits, 6,000 per user75%
Supplier query agentAgent owner50,000 credits70%
Policy agentAgent owner10,000 credits70%

At $0.01 a credit, the metered caps add up to $1,200 a month. The seat budget alert is about use, because a seat costs the same whether it's used or not.

Setting the first numbers

For a new agent, I'd set the cap from a forecast of decision volume (see Forecasting Copilot Credits from decision volumes) plus a margin, then tighten it after a month of real data. For a team, I'd start from the last month's actual use and review it weekly, as in the Friday routine.

Where I got stuck

I'm still unsure how cleanly the two admin centres will join up. Microsoft's September 2026 announcement says cost management in Agent 365 is expanding to cover agents built in Copilot Studio, planned for October. Until that's live and I've seen it, I'm treating team budgets and agent budgets as two reports that someone has to put side by side.

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