Here is the claim, up front. The procurement gravity that shaped how enterprises buy cloud is now forming around how they buy intelligence. If you sell agentic software, or you buy it, this is going to shape your next few deals.
I have spent over a decade on the enterprise side watching procurement shape architecture, at Amazon, AWS, and Cisco. The pattern holds every time. The commercial structure of a deal quietly dictates the technical structure of what gets built. Now I buy frontier intelligence for a company north of 3,000 people, and the same forces are lining up around tokens.
The mechanic that already runs cloud buying
Start with what enterprises already do, because the new behavior is a copy of it.
A big company signs a multi-year committed-spend agreement with its cloud provider. It promises to spend a certain amount over the term, and in return it gets better rates and a closer relationship. Then something happens to every other software purchase in the building. Whenever the company can route a third-party software buy through the cloud marketplace, it does, because that spend counts against the commitment it already made.
That is the whole trick. The money was going out the door anyway. Routing it through the marketplace makes it do double duty, it pays for the software and it burns down the commit at the same time.
This is a big part of why marketplace private offers exist. A vendor negotiates a custom deal, the buyer transacts it through the marketplace, and the buyer's finance team is happy because the dollars land where they were already promised. The vendor who makes that easy wins. The vendor who forces the buyer to pay off the commit creates friction that has nothing to do with the product itself.
A procurement scenario you will recognize
Picture a real evaluation. A team wants to buy an agentic tool, say one that reads inbound support tickets and drafts the responses. The demo lands well and the workflow fits. Everyone in the room likes it.
Then the deal reaches procurement and finance, and the questions change. How is this priced? The vendor charges per seat, plus a usage fee that bundles in the model calls the agent makes under the hood. Finance looks at that usage line and asks the question that is going to define this whole category. We already committed to a large spend with a model provider this year, so why are we paying your markup on inference when we could run this on capacity we already bought?
If the vendor has no answer, the deal slows down, sometimes it stalls. The product was never the problem. The inference bill was stapled to the product, and the buyer already had somewhere cheaper and pre-committed to put it.
The early shape is already visible
This is not a far-off prediction, the plumbing is being poured right now.
Look at provisioned capacity on Bedrock and dedicated model deployments on Azure AI Foundry. Look at the model commitments getting written straight into enterprise agreements with the labs. Companies are starting to hold intelligence the way they hold cloud, as a pooled and committed resource with a run rate and a budget owner.
Once a company owns a pool like that, it wants everything to draw from it. The same gravity that pulled software into the cloud marketplace is going to pull agentic workloads onto committed inference.
The prediction: bring-your-own-inference becomes a checkbox
So here is where it goes. "Does your product run on our inference" becomes a line in the RFP, the same way "runs in our VPC" became one a decade ago.
Remember that fight. There was a time a SaaS vendor could say "trust us, it is safe in our cloud" and win the deal. Then "must run inside our VPC" showed up in the security questionnaire, and every serious vendor suddenly had to have an answer. Bring-your-own-inference is the same move, only driven by finance instead of security. For agentic B2B software it goes from a nice differentiator to table stakes.
What vendors should do now
Decouple your value from the inference bill. That is the whole strategy in one sentence.
In practice that means a few things. 1/ Price the workflow, not the tokens, charge for the outcome your agent produces, the resolved ticket or the shipped feature. 2/ Accept the buyer's model where you can, let a customer point your product at their own model commitment on their own account. 3/ Make your margin defensible without the inference markup, because that markup is the first thing a committed buyer is going to squeeze.
If your business only works because you are reselling tokens at a spread, you are building on ground that is about to move.
What buyers should do now
Treat token spend as one pooled commitment, not a dozen line items scattered across tools.
Two moves matter. 1/ Get a single owner for intelligence spend, the way you already have one for cloud, and give that person the run rate and the commit to manage. 2/ Ask the bring-your-own question in every agentic evaluation, can this product run on the model we already pay for, on our account, under our commitment? A "no" is not an automatic dealbreaker, but it is a real cost and it belongs in the comparison.
Common sense, mostly. The companies that treated cloud spend as one managed pool got years of pricing power out of it. The same prize is sitting there in intelligence, for whoever sets it up first.
This is post 2 of 9 in my Thoughtful interfaces series. The rest lives at hagestedt.com/writing.