Procurement agent for AI tokens #
There is a paradox developing for the AI foundation model companies. After playing around a bit with Fable, the latest offering from Anthropic, I felt that the length and complexity of tasks that the model can accomplish has greatly expanded. Let’s assume this continues as the models become more and more advanced on the frontier.
The less you actually need to interact with the model via an interface (in my case Claude Code) the less you care about the underlying model. If some of the work was accomplished during my Fable session by a lower cost Haiku model, I wouldn’t have cared. I probably wouldn’t have noticed at all, given that I was working on other things 1 while the Claude Code was working on a long-running task in the background.
The paradox is that the better the model, the more susceptible to the introduction of the procurement agent and the less profit available for model companies.
The procurement agent, since this is 2026, would be an AI agent, of course. The procurement agent would be tasked with accomplishing the same goal, at the most cost effective token costs, usage, and timing. The procurement agent gets great at understanding tradeoffs for the user and their business: how they value their time, the level of specificity in certain requests, when they want to work in real-time vs. asynchronous, and the shape of their budget tolerances.2
The frontier might keep much of the planning and orchestration, but the tasks get parceled out to eager, cheaper models. Knowing which areas of planning and orchestration truly need bleeding edge capabilities is not something that is easy for most users to have intuition over.
The procurement agent has an existing baseline to improve on. Token spend for many software orgs is moving to become the second largest cost item after salaries. Ramp data released this week shows that the top 1% of AI spending organizations spend an average of $7,450 per month per employee on AI tokens. The cost difference between low-end and high-end tokens can be 100x.3
Procurement work can be framed in different ways in an organization: it makes us more profitable, it takes work off the plate of others, it unlocks being able to do things we couldn’t do before. The AI procurement agent is aimed right at the last bucket.
One of the change management issues you would encounter in a legacy business bringing in procurement for the first time pre-AI would be that it usually isn’t viewed as unlocking doing things that couldn’t be done before. Cost cutting makes people nervous because they see that they too are a cost that can be trimmed. Sure, the gratuitous waste can be axed, but beyond that it becomes harder sledding.
Having just lived through the brief moment of tokenmaxxing 4 that happened at a very small number of firms, we have all collectively recalibrated away from the simple
more is moreto the more rational measuring of both inputs and outputs. Set against that backdrop, the procurement agent’s entry into the AI token arena becomes a stabilizing force wrangling the live wire of spend.
As I saw firsthand with procurement in the telecom 3G and 4G buildouts, procurement enters the scene when there is some interchangeability in the tech and between brands. Telecom was built on costs (network capex) creating revenue and smarter procurement orgs at telcos understood that. Now with AI, token spend can create revenue and reduce costs if applied smartly.
It may be that the profit maximizing point for many businesses is more AI spend and not less. AI procurement agents will be at the vanguard of this—helping overcome natural human biases and hesitations.
As we’ve seen, SaaS contract lengths have been dropping. Many businesses are using the AI tokens without contracts at all. The procurement agent fights the peanut-butter lock-in that smoothly covers everything—it stratifies the spend and is surgical. It doesn’t consider its career and can have many of the principal-agent problems addressed in its marching orders.
Who could be at the forefront of this? It’s easy to rule out most of the usual suspects—none of the model companies themselves would be trusted to be objective in their analysis and decision making for a procurement agent product.5
There is a great opportunity here for a laser focused startup to build this. Latch on to some of the most profligate spenders and ride the cost curves with them. There are CFOs out there that want to tell investors about how they are using AI to manage spend and maximize output. We might all be surprised where this takes us.