Usage-Based Pricing for Copilot Is Good for Microsoft's Investors. Read That Sentence Again.

TheStreet ran a piece this week arguing that, of Microsoft's two Copilot announcements, the shift to usage-based pricing matters more to investors than the DeepSeek flirtation. That read is correct. It is also the tell.

Here is what Microsoft actually did. Copilot Cowork, the agent that reaches across Microsoft 365 to run multi-step work on your data, is coming off the flat per-seat add-on and moving onto consumption billing the company calls "Copilot Credits." Charles Lamanna, who runs Copilot, told Axios the product could not be offered on an unlimited-use basis. The users he pointed to are the ones doing hundreds of tasks a week. He called them "way productive." And then he said the part vendors normally keep off the slide: their costs go very high.

So the most productive users are the expensive ones. Hold that thought, because the whole argument lives there.

What "good for investors" is really saying

A pricing model earns the label "good for investors" when three things are true. Revenue starts to track cost-to-serve. Revenue scales with consumption instead of sitting flat per seat. And the vendor stops eating the margin on its heaviest users. All three are true here. None of them is a statement about whether a customer got value.

That is the gap I want to sit in for a minute.

Usage-based pricing meters an input. Tokens, compute, credits, whatever the unit. The customer does not buy tokens because they want tokens. They want a finished report, a resolved ticket, a reconciled spreadsheet. The token count is the cost of producing the outcome, not the outcome. And the relationship between the two is loose at best.

TSIA put it plainly in its May analysis of AI pricing: usage does not equal value, and consumption models often fail to reflect actual business value. That is not a critic talking. That is a research firm whose audience is the vendors building these models.

Agents make the coupling worse, not better

A chatbot answers and stops. An agent keeps going. It reads files, calls tools, checks its own work, hits a wall, tries again. Each of those steps burns compute, and the steps are what get metered. Every retry, every verbose detour, every loop the agent runs to second-guess itself adds to the bill. The customer pays for all of it.

Now ask yourself the hard question. Does a workflow that took the agent three retries and a long chain of self-checks deliver more value than the same workflow done cleanly in one pass? Of course not. It delivers the same outcome and costs more. Under seat pricing, that inefficiency was the vendor's problem. Under usage pricing, it is line-itemed onto the buyer's invoice.

The billing platform Flexprice, which sells the plumbing for this, says it out loud to its own customers: retries, loops, and background jobs are friction, not value, and usage-based pricing only works when customers gain something real as the meter climbs. Their warning to vendors is the buyer's whole case.

The people who get punished are the people who bought in

We do not have to guess how this lands, because GitHub Copilot already ran the experiment. On June 1 it moved to token billing. The median user barely noticed. The pain landed on the top five to ten percent, and it landed hard: community projections of bills jumping ten to fifty times, one developer modeling a move from roughly $29 a month to nearly $750, another claiming $50 to $3,000. TechCrunch called it the end of Copilot's golden age.

Look at who those heavy users are. They are not abusers. They are the people who took the vendor's three-year advice to use the tool for everything, built agentic workflows around it, and made it part of how they work. The pricing change penalizes exactly the depth of adoption every vendor claims to want. And the old safety net, where running out of premium budget dropped you to a cheaper model so you could keep working, is gone. What is gone, too, is the cost ceiling.

Lamanna's "way productive" power user and GitHub's top-decile developer are the same person. The model charges most to the customer who is succeeding most. Reward and penalty have swapped places. The reward now goes to the vendor.

The structural problem, which is bigger than the bill

Now the second half, and this part is more important than any individual invoice.

Think about where accountability for value sits in each pricing model. With outcome-based pricing, the vendor gets paid when a result lands and not before. Fin's (formerly known as Intercom) Fin charges 99 cents per resolution, billed only when the customer confirms the AI actually solved the problem. Under that model, every failed attempt costs the vendor. So the vendor has a direct, financial reason to make the agent efficient, accurate, and sparing with compute. Their margin depends on it.

Usage pricing inverts that incentive. The vendor is paid for activity regardless of whether the activity worked. An agent that burns more tokens, retries more often, and reasons more verbosely produces more revenue, not less. I am not claiming Microsoft will deliberately bloat Cowork to pump credits. I am saying the financial pressure that used to push toward lean, effective agents has been switched off; and switched-off incentives have a tendency of showing up in the product eventually.

The demand side pulls in the same direction. There is a name for it now: tokenmaxxing, the workplace habit of treating AI usage as a proxy for productivity, where people get judged on how many token they burn rather than on what they shipped. Built In's writeup is blunt about it: the habit rewards visible activity, not results. So stack the three forces. Buyers under pressure to run up consumption as a status signal, a vendor that meters by consumption, and an agent that inflates consumption on its own. Everything drives the meter up. Nothing points it at the outcome.

That is the real cost of the model. It moves the vendor one step further from owning the question of whether you got value, and it hands that entire question to you. The vendor essentially plays Pontius Pilate. The buyer now runs FinOps for AI. You set the budget caps. You write the spending policies. You read the consumption dashboard. You type /cost to see what a task burned. Microsoft, to its credit, is shipping all of those controls, and they are better than the ones GitHub fumbled out the door. But notice what they are. They are tools for the customer to govern value. They are not the vendor guaranteeing it.

The honest counterargument

I would be doing the same vendor-spin thing I just criticized if I left out the other side.

Flat pricing for agentic tools is inherently unsustainable. The economics are upside down: the model subsidizes the heaviest five percent and overcharges the lightest fifty. Metered billing is the rational fix for that, and for a low-volume or experimental buyer it is a better deal than paying a fat seat fee to barely use the thing. Aligning price with cost-to-serve is a good thing. It is just a vendor virtue, not a customer one, and the trick to watch is anyone presenting the first as if it were the second.

Cheaper models do not solve the coupling problem. A fine-tuned DeepSeek on Azure lowers the unit price of the metered thing. It does not make the metered thing track value. You are paying less per token for a number that still has a loose relationship to your outcome. And who knows how many additional tokens a potentially inferior model burns.

Where this lands

On my orchestration battleground, Cowork is the M365 layer that coordinates work across your apps and your Graph. Pricing that orchestration by consumption reframes it from a capability you own into a utility you rent by the drink. That is a substantial shift in who carries the risk when an orchestrated workflow goes long, and it is the buyer.

So, on the null hypothesis. Is usage-based pricing good for customers? Largely no, and for the reasons the question assumed. The metered unit is loosely coupled to value, agents widen that gap rather than closing it, the model bills the most engaged users the most, and it relocates the entire burden of value accountability from the vendor onto the buyer. Good for investors and good for customers are not in alignment here. On this one, they partly trade off.

Three things to do if you are buying

Model your power users, not your average. The average user will not break your budget. The fifteen people who actually adopted the thing will, and they may very well be the ones delivering your return.

Make the vendor define the unit before you sign. If a task can cost anywhere from a few credits to a few hundred depending on how many times the agent talks to itself, that is not a price, it is a range. And you'll end up at the upper end, trust me. Ask vendors to commit to a per-outcome cost and watch how fast the conversation gets vague.

Push for outcome terms on anything that has a definable outcome. Resolution, completion, ticket closed. If the vendor will only price the effort and not the result, they are telling you something about how confident they are in the result.

The interesting question is not whether usage pricing is here. It is. The question is whether buyers will accept a model where the vendor is paid the same whether the agent nails it on the first try or flails through ten, or whether the market pushes back toward paying for outcomes the way Fin does. Or Zendesk. Or Hubspot. Or others. I do not know which way that goes. But the vendor whose margin improves when its agent works harder is not, structurally, the vendor most motivated to make the agent work better.