- Subject
- Microsoft 365 & SharePoint
- Filed
- Source
- [email protected]
- Loans
- 769
AI FinOps: Track Token, Agent and License Costs Across Your Business
AI costs can rise with usage, agent growth and overlapping licenses. Inventory agents, assign owners and review spend alongside access and business value.

AI spending can grow without a new server or a large change to an architecture diagram. More prompts, a different model or an agent calling another agent can raise costs as adoption spreads. That makes AI FinOps a practical management task: connect spending to the systems, owners and outcomes behind it before the bill becomes difficult to explain.

The scope is broader than cloud invoices. Teams need to account for model tokens, inference compute, GPU capacity, agent usage and licenses across platforms. They also need to know which agents exist, who owns them and what data they can access. A cost report without that context can show that spending has risen, but not whether the increase reflects valuable work, duplication or an abandoned experiment.
AI costs have become a mainstream FinOps responsibility
The FinOps Foundation’s State of FinOps 2026 report found that 98% of FinOps practitioners now manage AI costs, compared with 31% two years earlier. The shift reflects a change in what cost-management teams must track: AI services can generate variable usage charges alongside familiar infrastructure and subscription costs.
The FinOps Foundation’s remit has broadened from cloud value to technology value, as reflected in its mission update. That change matters for organizations buying AI through multiple channels. A per-seat license may be easy to forecast, while an agent’s usage charges can vary with ticket volume, model choice and the number of retrieval or agent calls needed to complete a task.
Consider an organization that licenses Microsoft 365 Copilot for 2,000 employees and separately runs a retrieval-augmented generation agent for customer support. The Copilot subscription is largely a per-seat cost. The support agent’s costs can change as workload and usage change. Looking at either product alone won’t reveal the total cost of AI across the organization, or whether several tools are doing similar work.
Why finance is taking a closer role
AI investment is competing with other technology needs, while finance teams are being asked to explain what that investment delivers. In Gartner’s 2026 CIO Agenda, 52% of surveyed CIOs said cost reduction would become more important over the next two years, even as many organizations continued to increase AI funding. BCG’s IT Spending Pulse found that cost management had become a top-three CIO priority, rivaling growth for the first time since 2023.
McKinsey’s research on technology budgets in the AI era found that AI was consuming up to a third of companies’ change budgets while also adding to the cost of running existing systems. When AI takes that much room in a technology budget, finance needs enough detail to compare spending with expected outcomes. That calls for collaboration between finance, IT and the teams responsible for data access and risk.
Gartner’s 2026 CIO Agenda provides the cost-reduction figure, while BCG’s IT Spending Pulse provides its finding on CIO priorities. Those findings point to a shared operational challenge: organizations want to expand AI use while improving their ability to identify, assign and review its costs.
Four sources of AI cost growth
Usage based charges
Token and inference costs follow activity. Prompts, retrieval calls and handoffs between agents consume resources, so spending can increase as more employees use a service or as a workflow becomes more involved. A small amount of usage by one person may look insignificant; repeated across many users and embedded features, it can become a substantial budget item.
This pattern differs from costs tied mainly to provisioned capacity. A workload’s consumption can change even when its architecture appears unchanged, because adoption rises, a model changes or an agent begins making additional calls. Teams need to see usage at a level that connects it to a workload or owner, rather than treating a combined invoice as an explanation.
Agent sprawl
Each agent can carry its own compute, token and license costs. When teams create agents independently, an organization may accumulate a large fleet without a reliable inventory, a named owner for every system or a clear record of what each agent does.
Gartner projects that the average Fortune 500 enterprise will run more than 150,000 AI agents by 2028, up from fewer than 15 in 2025. Gartner also found that only 13% of organizations believe they have the right governance in place for that growth. Those projections make inventory and ownership important parts of cost control: without them, teams may struggle to identify which agents are active, duplicated or still justified.

Shadow AI
Tools used outside IT’s view can create both unbudgeted spending and access risks. IBM’s 2025 Cost of a Data Breach Report found that one in five breached organizations traced the incident to shadow AI. Those organizations paid an average of $670,000 more than a standard breach, and 97% lacked basic AI access controls, according to IBM.
The cost of an unsanctioned tool may not appear under a clear AI budget category. It can surface instead through an unexpected subscription, duplicated service or incident response expense. That gives IT and finance a reason to include discovery of unapproved tools in cost reviews, alongside access and security checks.
Overlapping licenses and services
Different platforms can offer similar capabilities under separate subscription and usage models. A Microsoft 365 Copilot seat, a Google Gemini license and Salesforce Agentforce credits may each be purchased by different teams, leaving no shared view of overlap. Projects can also duplicate effort: one team builds an agent in Copilot Studio, another uses Salesforce Agentforce for a similar workflow, and a third starts a Vertex AI agent while waiting for the other work to progress.
Each decision may look reasonable when viewed on its own. Across Microsoft 365, Google Workspace and Salesforce, the same business need may be funded several times through licenses, compute and development effort. Reviewing one platform’s admin console at a time can miss that pattern. A cross-platform review helps determine whether the tools serve distinct needs or whether a capability is being bought and built more than once.
AI FinOps and cloud FinOps use different cost signals
Traditional cloud FinOps often examines compute hours, storage, data transfer and whether infrastructure is idle or oversized. AI FinOps adds usage units such as tokens, inference calls and agent-hours. Because activity and model choice can change costs, teams need to look beyond provisioned capacity.
| Area | Cloud FinOps | AI FinOps |
|---|---|---|
| Cost units | Compute hours, storage and data transfer | Tokens, inference calls and agent-hours |
| Cost changes | Often linked to provisioned capacity | Can change with usage and model choice |
| Common sources of waste | Idle or oversized infrastructure | Untracked agents, shadow AI and overlapping licenses |
| Governance scope | Cloud accounts and resource tags | Agents, prompts, ownership and accessible data |
| Optimization options | Rightsizing and reserved capacity | Model selection, agent consolidation and clear ownership |
The distinction has practical consequences. Cloud cost tags can help show which account or project uses infrastructure, but they may not reveal which business process an agent supports, who approved it or what information it can access. For AI services, those details help teams decide whether an increase is justified and whether an agent should remain in service.
What a mature practice looks like
A useful AI FinOps program progresses from discovering spend after the fact to managing cost, ownership and access together. The difference is visible in how teams respond when a bill changes.
| Stage | Typical practice | Likely result |
|---|---|---|
| Reactive | Teams find AI charges on a monthly invoice; there is no complete agent inventory or assigned owner. | Surprises are difficult to trace to a workload or team. |
| Managed | Spending is tagged by team or project, agents are tracked manually and costs are reviewed monthly. | Visibility improves, but ownership and access questions may remain unresolved. |
| Governed and optimized | Each agent has an owner; budget limits are connected to intended outcomes; cost and access are reviewed together. | Teams can relate spending to value and identify redundant or unmanaged agents earlier. |
A dashboard is useful, but it cannot answer every management question. It can show that agent spending rose 40% in a quarter. It may not show who approved the agent, which data it can reach, whether it is still serving its intended purpose or what a shutdown would affect. Those answers depend on an inventory and clear ownership as well as billing data.
AvePoint describes this combined approach through its AgentPulse Command Center. The company says the service inventories agents across Microsoft 365 and Google Cloud, including Copilot Studio, Microsoft Foundry, SharePoint agents and Vertex AI, and presents ownership and cost information alongside access and policy data. This is a vendor’s description of its product; the broader operational point is that spending information becomes more useful when it can be tied to an agent’s owner and access context.
Steps to start controlling AI spend
- Inventory agents in production. Include systems built by teams outside IT. An incomplete inventory makes it difficult to identify duplicate work or understand total usage.
- Name an owner for each agent. The owner should be able to explain its purpose, usage and continued need, and take part in cost reviews.
- Connect spending to outcomes. Compare cost with a useful measure for the workload, such as the task it completes or service outcome it supports. Team-level totals alone can hide differences between a widely used service and an unused pilot.
- Review licenses across platforms. Examine Microsoft 365, Google Workspace and Salesforce together, and check whether separate teams are paying for overlapping capabilities.
- Set budget limits before rollout. Estimate expected usage and establish a review process before scaling. A limit set only after a spike helps explain the past but offers little warning during expansion.
- Include shadow AI in cost discovery. Look for unsanctioned tools as both possible access concerns and expenses outside planned budgets.
Organizations can start with provider cost dashboards to examine token and compute charges, then add agent tracking as the inventory grows beyond what spreadsheets and manual tagging can reliably cover. The key gap to close is connecting a cost record to an identifiable service, its owner and its purpose.
Measure efficiency against useful work
Raw token consumption says how much a service used, but not whether it performed well. A more informative primary measure is cost per completed task or business outcome. Supporting measures can include the number of agents compared with the number in active use, the share with named owners and month-to-month cost variance.
These measures help separate productive growth from waste. An increase in usage may be appropriate if an agent is completing more valuable work. A duplicate agent, an abandoned pilot or a service with no active owner may call for a different decision. Cost optimization therefore doesn’t mean simply reducing AI use; it means reviewing whether spending continues to support a worthwhile outcome.
For new deployments, set a spending ceiling based on expected usage before launch and arrange for an alert as costs approach it. The source guidance suggests an alert at 70 to 80% of the ceiling, giving the named owner time to review the trend before it reaches the limit. The right review interval depends on how quickly deployment and usage are changing; fast-growing services need closer attention while their costs are still settling.
Make cost and access reviews part of the same routine
AI FinOps works across teams. Finance can own budget accountability, IT can provide technical cost data, and security can assess agent access and associated risks. Keeping those responsibilities connected helps prevent two common blind spots: costs without an accountable owner, and technical context without anyone responsible for the budget.
For Microsoft 365 Copilot, the per-seat license is only one part of the picture. Copilot Studio agents and other AI services may introduce usage-based charges, depending on the service and pricing model. Organizations should therefore assess Copilot licensing alongside the wider AI environment, including agents, model usage and overlapping services.
The FinOps Foundation’s State of FinOps 2026 report offers a benchmark for the growing role of AI spend in technology cost management. For individual organizations, the next practical check is more immediate: can each agent be matched to an owner, a purpose, its costs and the data it can access? That record gives finance and IT a basis for deciding which AI spending to expand, adjust or retire.
