A team reviews an AI usage dashboard tracking credits, spending, budgets, guardrails, and workflows in a modern office.
Remember the manager who congratulated you for topping the AI leaderboard in 2024? Microsoft Digital, the company's internal IT organization, says that manager may now be stopping by your desk to ask you to hold off on using AI. A Microsoft Inside Track post published November 19, 2026, argues that the reason is simple. AI is increasingly billed by consumption instead of a flat license fee. The "how do we drive adoption?" question has turned into "what are we getting for what we spend?"

This piece covers what Microsoft Digital says it is doing, what Microsoft Learn documents for admins, and where the framing needs caution.

A team reviews an AI usage dashboard tracking credits, spending, budgets, guardrails, and workflows in a modern office. From flat fees to metered work​

Microsoft Digital describes the shift this way. Two years ago most AI tools had a fixed subscription price, so leaders encouraged heavy use. Today a growing share of agentic and reasoning work is billed by consumption, often as credits. Employees often don't know which experiences are included in their license and which can run up a bill.

The post comes from Microsoft's own IT team, which calls itself "Customer Zero" for Copilot, Copilot Cowork, Microsoft Foundry and Agent 365. These are Microsoft's own accounts of its internal practices and customer conversations. They are not independent measurements.

The core idea: return on token spend​

Microsoft Digital frames "human capital" and "token capital" as two finite resources. It proposes a leadership metric called return on token spend, defined as business value delivered divided by tokens consumed.

This is a conceptual measure, not an accounting standard. The post doesn't say how to put a number on business value or how to compare unlike outcomes. It also doesn't say how to fold in costs beyond tokens, such as licensing, development and human review. Treat the metric as a discussion tool until your organization defines its own method.

The post describes three phases of enterprise adoption:

  1. Tokenmaxxing: drive usage at almost any cost to build habits.
  2. Cost control: budgets, alerts and hard caps. Microsoft says most organizations are here now.
  3. Valuemaxxing: compare tokens spent against outcomes delivered.

Microsoft's advice is to design for phase three even while you operate in phase two. It also borrows four lessons from cloud FinOps:

  • Meter before you cap.
  • Tag usage at the source.
  • Budget for your worst month rather than your average one.
  • Track cost per outcome rather than cost per token.

Cheaper tokens, bigger bills​

The post leans on a Stanford figure. The 2025 AI Index says the inference cost for a system performing at the level of GPT-3.5 dropped over 280-fold between November 2022 and October 2024. Coverage of the report describes the comparison as a fixed GPT-3.5 performance level on the MMLU test (64.8 points), going from US$20 to US$0.07 per million tokens. So it measures the price of one fixed capability level. It is not a price history for all models or workloads.

Microsoft's point is that falling unit prices don't settle total spend. The post notes that CEO Satya Nadella said on the April 30, 2025 earnings call that Microsoft's cost per token had "more than halved." Microsoft Digital stresses that this describes the company's own cost efficiencies, not a matching drop in customers' bills. More capable models running multistep agent workflows use more tokens.

Microsoft Digital also says usage can differ by 10 to 100 times between a light chat user and a heavy agent user. It presents this as an observation, with no published methodology. If it holds for your tenant, a few enthusiasts may drive most of your consumption. Microsoft says the aim is to educate those users, not block them.

What admins can actually control​

The post lists a layered set of tools:

  • Microsoft 365 Cost Management for Copilot Credits.
  • Power Platform and Copilot Studio controls.
  • The Viva Insights Consumption Dashboard.
  • Azure Cost Management and Foundry observability.
  • AI gateway in Azure API Management, for rate limits, quotas, routing and caching.

It adds that Agent 365 is generally available as the control plane for agents, with deeper financial telemetry on the roadmap. Microsoft Digital also says it began showing its own employees and vendors their AI usage and token spend in November 2026.

Microsoft Learn, last updated November 17, 2026 according to the page, gives the operational detail.

Billing models vary by service​

Microsoft's usage-based billing model charges customers based on actual usage, measured in Copilot Credits. Depending on the service, organizations can use fixed licensing, the Copilot Credit Pre-purchase Plan (P3), pay-as-you-go, prepaid capacity packs, or supported combinations. So not every Copilot experience is metered per token, and not every license is unmetered. What applies to you depends on the service and your configuration.

The Learn documentation lists the services currently managed this way. They include Copilot Cowork, advanced work in SharePoint and OneDrive, Work IQ API, Copilot Managed Runtime, Teams Phone Agent, Copilot Code, Autopilot, Copilot Studio and Windows 365 Agent. Microsoft says it will add more over time, so don't treat the list as exhaustive.

Roles​

  • Global administrator or Billing administrator: can add, select and change billing methods.
  • AI administrator or License administrator: can edit spending policies, limits and alerts, but cannot create policies.
  • Reader roles: can view dashboards and reports with read-only access. This lets finance and governance staff see consumption without changing settings.

Microsoft recommends using the roles with the fewest permissions.

Setup path​

The path is Copilot > Cost management in the Microsoft 365 admin center. An administrator activates the default spending policy and chooses a billing method. They can then limit or not limit monthly spending, optionally set a per-user monthly limit, and define alerts as a credit amount or percentage. Alert emails start when the threshold is reached and repeat daily until the monthly period resets or the policy is changed. Microsoft recommends setting the per-user limit to prevent runaway spending by one individual.

Gotchas worth checking​

  • Limits don't inherit. The tenant-level limit is set on the default policy. Each additional policy has its own independent limit.
  • Auto-apply is on by default. The "Auto-apply new services" setting adds future supported Copilot services and agents to a policy. Turn it off if you want to review new services first.
  • Copilot Studio is opt-in. Environment and agent policies in the Microsoft 365 admin center are opt-in. Existing Power Platform admin center controls keep applying until a Microsoft 365 spending policy covers that scope.
  • Credits are not dollars. The Copilot Cost Management API reports AI credits. Microsoft says it does not include pricing, invoices, currency costs or purchasing forecasts. Finance teams still need billing data to reconcile actual spend.
  • Billing order. If you select both prepaid capacity packs and an Azure subscription with P3, billing applies capacity packs first, then pay-as-you-go.

Teaching people to spend tokens well​

Microsoft Digital says education is the single most important area. Its decision path for knowledge workers runs down a list, stopping at the first yes:

  1. Can Copilot Chat handle it in one conversation? Use it with the model set to Auto. Microsoft describes Auto as included in the license, but actual entitlements depend on your license and tenant policy. Choose a specific model only for business-critical work, then switch back.
  2. Does the task need multistep work and a finished deliverable? Use Copilot Cowork and scope it carefully.
  3. Should Copilot run in the background across your apps? Use an agent such as Autopilot, only for work that warrants it.
  4. Will a whole team repeat the process? Build a custom Copilot Studio agent with an owner and a budget.

For prompts, it offers the CRAFT checklist:

  • Context: include only what the AI needs.
  • Requirements: scope, length, tone, exclusions and success criteria.
  • Audience: who the output is for.
  • Format: the exact shape, such as an email, table or executive brief.
  • Task: exactly what you want done.

Its example contrasts "Prepare me for my customer briefing" with a request limited to the last 60 days. That request asks for five priorities, three talking points and three open risks in a one-page brief. Microsoft says the second version gets a sharper answer in one pass with far fewer tokens. It gives no measured savings, but the logic is plausible. Clear scope tends to cut retries.

A practical checklist for IT teams​

This is my reading of the guidance, not Microsoft's wording.

  1. Meter first. Review the Cost management dashboard and the Viva Insights Consumption Dashboard before setting hard limits.
  2. Set a default policy with alerts and a per-user cap. Then add targeted policies for power users and agents, remembering each needs its own limit.
  3. Decide on auto-apply deliberately. It affects how future services inherit your rules.
  4. Pair credit reports with finance data. Credits alone won't give you a dollar ledger.
  5. Name an owner and a budget for every custom agent. Microsoft's own decision path asks for this.
  6. Define "value" before you measure it. Without that, return on token spend is only a slogan.

The skeptical view​

This is a vendor describing a problem partly created by its own pricing models, and offering its own tools to manage it. That doesn't make the advice wrong. Metering before capping is sound, and so is tagging usage at the source. But the post is not independent evidence that the tools fully solve the problem. Microsoft itself says the discipline is young and the tooling is still maturing, and that deeper financial telemetry in Agent 365 is still on the roadmap.

The takeaway for admins is to treat AI consumption like cloud spend. Instrument it, assign owners, set guardrails, and ask what each token bought. Do that before the invoice does it for you.

 

References

  1. Inside Track - Maximizing return on token spend: How we're approaching FinOps for AI Microsoft 2026-11-19T17:00:00+00:00
  2. Usage-Based Billing and Cost Management for Copilot Credits | Microsoft Learn learn.microsoft.com
  3. Set up usage-based billing for Copilot Credits | Microsoft Learn learn.microsoft.com