AI won’t fix productivity – it will price it

We used to have a simple story for Microsoft productivity tools: buy a license, assign it to a user, enable the feature, move on. That model is changing from fixed licensing to consumption-based pricing, and looking more like cloud consumption.

Copilot Cowork, GitHub Copilot, and the new Copilot Studio experience based on the GitHub Copilot harness all point in the same direction: AI work is becoming metered work. The license gives access, while usage creates the bill, and in the process changes a lot for customers, partners, makers, developers, and finance teams.

The license is only the entry point

Copilot Cowork is the northstar example. You still need Microsoft 365 Copilot, but the Cowork work itself is charged through Copilot Credits. If a task uses more context, calls more tools, runs longer, or uses a more expensive model, the cost goes up. That is technically reasonable, because a small summary and a multi-step agentic task are not the same workload. But the sentence “we already bought Copilot” no longer closes the discussion.

💡 Did you know? You can use the built-in skill “/cost” shows you the consumption in the current conversation.

GitHub Copilot is already there. It started as a coding assistant in the editor, and is now much closer to an agentic developer platform: chat, code review, command-line help, repository context, cloud agents, and workflows that can run for longer than a quick suggestion. Charging this through AI Credits makes sense, but it also means engineering leaders need to understand which patterns burn money and which patterns actually save it.

Copilot Studio adds the next challenge. With the new announced GitHub Copilot harness, building and testing agents also becomes part of the consumption picture. That means cost can start before production (during design, testing, evaluation and iteration), not when the agent is rolled out.

Invoice visibility is a challenge

Of course AI costs money. Models need compute, retrieval needs services, tool calls and automation need platform resources – nobody should be surprised by that. The challenge is that the bill does not always show up in one clean place.

Cowork and Work IQ costs are tracked in Microsoft 365 Admin Center, Copilot Studio sits in Power Platform. And GitHub? A world on its own. Dataverse Storage, APIs, Tools, custom models, fine-tuned models, MCPs, Trainings, Change Management, human reviews – yes, you are still needed – and technical dept, all belong to the real AI cost picture.

In a perfect fresh setup, you definitely want to point all these consumption services to Azure Subscription(s) to collect a global overview.

This matters for partners as well. If we use AI to deliver faster, do we pass that efficiency to the client? Do we charge differently because our own AI tooling creates direct consumption cost? What happens when the project margin improves on hours but shrinks on credits, storage, APIs, and platform usage?

Three places where this becomes concrete

Copilot Cowork: This is delegated work with a meter attached. Tenant, group, and user limits are not admin details anymore. They are part of financial governance. If a task can run longer and call more tools, someone needs to decide who is allowed to run it and under which budget.

GitHub Copilot: The question is no longer only “does it make developers faster?” The better question is: which Copilot usage creates net value? Model choice, cloud agents, code review, repository context, and long-running workflows can all change the commercial equation.

Copilot Studio: Agent building needs cost discipline earlier than many teams expect. Experiments are fine. Endless experiments with no budget owner are not. If building, testing, and evaluating agents consumes credits, the maker experience needs guardrails before the first “successful pilot” becomes the next expensive habit.

How to clear the fog

• The first decision is ownership. AI consumption cannot sit everywhere and nowhere at the same time. If the business wants agentic work, the business needs to own part of the cost discussion.
• The second decision is control. Who can use Cowork? Who can build agents? Who can select expensive models? Who can connect data sources? Who gets alerts before a budget is gone? These questions are boring until they are suddenly urgent.
• The third decision is measurement. Saved hours are not enough. Add adoption effort, security review, data readiness, support, process redesign, licenses, credits, storage, and Azure consumption. Then talk about ROI. Not before.
• The fourth decision is client communication. We need to be transparent about where AI reduces effort and where it creates new cost. Clients do not need buzzwords, but a clean commercial story.

Is Microsoft Licensing the bad guy, again?

My view: consumption-based AI billing is not wrong. It is probably the only realistic model for agentic systems. But it forces a level of discipline many organizations do not have yet. The weak point is not the technology. It is how we govern, explain, and price the work around it.

AI will create productivity gains. No doubt. But productivity without cost visibility is just optimism with an invoice attached. The winners will not be the companies that enable every new AI feature first. The winners will be the ones that know where AI creates value, where it creates waste, and who pays for both.

Sources: 1; 2; 3; 4; 5; 6; 7