You Do Not Buy AI. You Consume It.

Software was something you bought. The company picked it, someone senior signed off on it, finance had a number, and that number sat still for a year. Whatever else was wrong with that arrangement, one thing about it was right: a person put their name on the decision, and you could find that person later.

AI is not sold that way. AI is sold by the action, in small amounts, continuously, to whoever happens to be at the keyboard. This piece is not about a prediction or a trend. It is about how you are being charged for AI. You can read all the details in the published rate cards right now.

Start with what is published, because the specifics matter more here than the theme.

Salesforce publishes a Flex Credits rate card, currently effective April 21, 2026. On it, a standard Agentforce action costs 20 credits. A voice action costs 30. Speech to text runs 150 credits per hour of transcription, and text to speech runs 6,000 credits per million characters. Salesforce’s pricing page lists Flex Credits at $500 for 100,000. So, a standard action costs ten cents, a voice action fifteen, and an hour of transcription seventy-five.

Microsoft publishes the same thing in different units. Copilot Studio bills in Copilot Credits, and its documentation lists the rate for each thing an agent can do: a classic answer is 1 credit, a generative answer is 2, an agent action is 5, a message grounded in your own tenant graph is 10, a page of content processing is 8, and a minute of premium generative voice is 75. Capacity is sold in packs of 25,000 credits at $200 per pack per month.

None of this is hidden. Both companies publish it plainly, in public, with the units defined. The rate cards are doing exactly what rate cards are supposed to do.

Now look at what is on them. Not the prices. The units.

A generative answer. An agent action. A message grounded in your own data. A minute of speech. Those are not products. They are the individual motions of work, itemized. What used to be a license has become your tab, and the tab is opened by anyone in the organization, without request, thousands of times a day.

Nothing on either list is expensive, and that is precisely why this is so easy to miss. A ten-cent charge does not trip a control. It never goes to procurement, does not need a business case, does not require a signature, and will never appear in front of the person whose job it is to approve spending for the firm. It is initiated by people who are trying to do their jobs, not making a purchasing decision, and many would be genuinely surprised to hear it described that way. Every prompt is a purchase order nobody signs.

At small volume it is a rounding error. At the volume these tools are being used to produce, it is a second procurement chain that materialized inside your company without anyone approving it.

How big are the numbers?

Gartner puts worldwide AI spending at $2.59 trillion this year, up 47 percent, and projects $3.49 trillion for 2027. Those are the numbers that get quoted. The one worth digging into is smaller and much more specific. In June, Gartner predicted that by 2028, AI coding costs will overtake the average developer’s salary, driven by rising token consumption and the shift to consumption-based licensing. Read that again, as a governance statement this time, not a budget one. A cost that was fixed, negotiated annually, and owned by a named person is being replaced by a cost that is variable, negotiated never, and owned by whoever presses send.

I want to be careful here, because the easy version of this piece is also the wrong one.

The easy version says vendors are gouging you with hidden meters. They are not. The rate cards are published, the units are defined, and consumption pricing is arguably the fairest way to sell something whose underlying cost genuinely does vary with use. Salesforce and Microsoft did not do anything to you. They changed how the thing is sold, announced it, and documented it. There is no villain in this story and looking for one will make you miss the real point.

What did not change is the structure on your side. No software vendor can do that.

Your procurement governance was built for a purchase: one decision, one accountable signer, a knowable price, and annual review. Your cost is now based on consumption, continuous decision making, by everyone, at a price that moves with how much the model decides to think about the question. Nobody chose to separate those two patterns. It is a default. In this case it is the difference between a pricing change the vendors announced and an accountability change nobody made.

I am not describing this from outside it. Fellowship Intelligence runs on metered AI. I can tell you what last month’s spend was, because I look at it. What I could not have told you six months ago is which part of those funds produced something a client actually saw, and which part was me thinking out loud. That is not a finance question. It is the entire question, and I could not answer it because I was reading the bill instead of matching the ledger.

This really is more about strategy than cost-control, and it should not be handled as just another line item.

A meter produces a record. Every one of those metered actions is logged. They come with a timestamp, an initiator, an input, an output, and a price. Taken together, that is a more honest account of how work actually gets done inside your organization than anything your last three process-improvement projects produced. And you are already paying for the record. It accrues whether you look at it or not.

Most organizations route it to finance, where it becomes a cost problem. It might get an owner in FinOps and stop. When read as a bill, its highest possible value is a smaller bill. Read as a ledger, it tells you which decisions in your company are now being made with machine assistance, by whom, how often, on what, and at what cost. Gartner’s own framing of that 2028 prediction names the gap directly: without clear visibility into token usage, organizations risk budget overruns and lose the ability to track cost against value. That visibility is itself the asset. The invoice is just the byproduct.

This is what having control as an advantage for your organization actually looks like. Not restraint, not a policy asking people to be thoughtful about usage, and not a moratorium. Control in the plain sense means knowing what is happening, being able to prove it, and having a named person who can answer for it. Adoption is not just encouraged but expected by the market now. Everyone will have the tools, at roughly the same prices, from roughly the same vendors. The difference between two companies running identical stacks will be that one of them can account for what it did with them.

Consumption pricing handed every company in the market a complete, time-stamped record of its own AI use, and most of them filed it under expenses.

The meter is already running. Decide whether it is a bill or a ledger.

SOURCES

Salesforce, Flex Credits Rate Card, effective April 21, 2026. https://www.salesforce.com/en-us/wp-content/uploads/sites/4/assets/pdf/agentforce/Flex-Credits-Rate-Card-04.21.2026.pdf

Salesforce, Agentforce pricing. https://www.salesforce.com/agentforce/pricing/

Microsoft Learn, Copilot Studio billing rates and management. https://learn.microsoft.com/en-us/microsoft-copilot-studio/requirements-messages-management

Microsoft, Copilot Studio pricing. https://www.microsoft.com/en-us/microsoft-365-copilot/pricing/copilot-studio

Gartner, worldwide AI spending forecast, May 19, 2026. https://www.gartner.com/en/newsroom/press-releases/2026-05-19-gartner-forecasts-worldwide-ai-spending-to-grow-47-percent-in-2026

Gartner, AI coding costs prediction, June 24, 2026. https://www.gartner.com/en/newsroom/press-releases/2026-06-24-gartner-predicts-ai-coding-costs-will-surpass-average-developer-salary-by-2028-as-token-consumption-surges