We’ve analyzed the protocol wars in agentic commerce. Three protocols competing for $385B in transaction volume. But protocol standardization won’t solve two deeper governance failures that sit underneath.
The first: who governs what agents buy. The second: who governs what agents cost.
These are distinct problems. They require different frameworks, different controls, and different organizational owners. Treating them as one problem is how companies lose control of both.
Problem one: discovery without oversight
A YC-backed founder described the moment his team chose a critical software vendor. An AI agent evaluated options, selected the tool, and the team moved. “The AI selected the tool and we moved, that was it,” he told Matt Williamson. No RFP. No procurement review. No vendor scoring matrix.
This is not an anecdote about a startup cutting corners. It is a preview of how enterprise purchasing will work when agents operate through Model Context Protocols, APIs, and developer documentation instead of UIs and sales decks.
Agents don’t browse landing pages. They parse changelogs, read API documentation, test integrations, and default to whatever works fastest. Williamson calls this “machine-readable growth”: the discovery channel where your product’s docs, not your demand-gen funnel, determine whether an agent selects you.
The governance problem is obvious once you see it. When humans select vendors, organizations wrap the process in controls. Approved vendor lists. Budget thresholds that trigger review. Compliance checks for data handling and security posture. Separation between the person requesting and the person approving.
Agents skip all of it. Not because they are malicious. Because those controls were designed for human workflows that agents never enter.
The concentration risk compounds quickly. Andrew Chen’s framework on tech-shift growth channels predicts that early winners in a new channel capture disproportionate share. When agents over-index on developer content (and they do), the vendors with the best docs, the cleanest APIs, and the fastest onboarding become default choices across thousands of agent-mediated decisions. One vendor’s documentation quality becomes the market’s moat.
This is how you get invisible lock-in. No contract signed it. No procurement team approved it. An agent chose the path of least friction, and a thousand other agents made the same choice the same week.
Problem two: the token margin crisis
The second problem is economic, and it is less visible but more dangerous.
Jamin Ball at Altimeter Capital published the arithmetic that should worry every CFO running AI workloads. An H100 GPU rents for $2 to $4 per hour. An NVIDIA GB300 NVL72 rack generates roughly one million tokens per second, or four billion tokens per hour. At current token pricing, that same hardware produces $600 to $800 per hour in revenue.
Read those numbers again. The same unit of compute costs single-digit dollars to rent and produces triple-digit dollars in token revenue. Median SaaS gross margins already exceed 80%. Token-based pricing pushes margins further.
This looks like a windfall. It is actually a governance crisis.
Ball’s warning is precise: “Price too low and you’re literally paying customers to use your product.” The opposite is also true. Price too high and you create an arbitrage that competitors or customers’ own infrastructure teams will exploit. The margin between those two failure modes is narrow, and it shifts with every hardware generation.
NVIDIA’s Vera Rubin architecture, expected next, delivers roughly 5x the throughput at 10x the cost reduction per generation. Every eighteen months, the floor drops. Pricing models built for one hardware generation become margin traps in the next.
For enterprises consuming tokens (not selling them), the crisis is different but equally urgent. When one agent call can trigger hundreds of downstream token charges, and when agents autonomously select tools that consume tokens on your behalf, cost governance becomes a function that didn’t exist twelve months ago.
The total addressable spend within each account is, as Ball puts it, “theoretically limitless.” An agent that finds value in a tool will use it without budget consciousness. It will not pause to check whether this month’s token spend exceeds the forecast. It will not compare the per-token cost of one provider against another unless someone built that comparison into its decision loop.
The structural intersection
These two problems intersect in a way that makes both worse.
When agents choose tools autonomously (problem one), they generate token consumption patterns that nobody forecasted (problem two). The agent that selected a vendor based on documentation quality has no awareness of the per-token pricing of that vendor’s API. The procurement team that might have caught the cost exposure never saw the decision.
Discovery governance without cost governance means agents select expensive tools freely. Cost governance without discovery governance means you control spend on tools you didn’t choose. You need both, and they need to talk to each other.
What governance looks like here
For discovery, the controls are structural. Approved tool registries that agents must query before selecting new vendors. Audit trails for every agent-initiated procurement decision, even when (especially when) the decision looks small. Threshold triggers that escalate to human review based on spend, data access, or vendor risk profile.
For cost, the controls are economic. Token budgets per agent, per workflow, per business unit. Real-time monitoring of consumption against forecast. Automated circuit breakers when spend exceeds thresholds. And critically, pricing model governance: a formal process for evaluating whether your token pricing assumptions survive the next hardware generation.
Neither control set is exotic. Both borrow from disciplines that already exist (procurement governance, cloud cost management). The difference is speed. Human procurement decisions happen on a timeline of weeks. Agent procurement decisions happen in seconds. The governance frameworks must operate at agent speed or they don’t operate at all.
The uncomfortable forecast
Agentic commerce will not wait for governance to catch up. Agents are already selecting tools, consuming tokens, and making purchasing decisions on behalf of organizations that have no framework for overseeing those decisions.
The organizations that build discovery and cost governance now will compound an advantage. Not because governance is exciting. Because ungoverned agent spend is the enterprise budget crisis that nobody sees until the quarterly review.
Protocol wars will sort themselves out. Hardware costs will keep falling. The question that persists is whether your organization governs what agents buy and what agents cost, or whether you find out after the fact.
This analysis synthesizes Selling to AI Agents by Matt Williamson (March 2026) and Per Token Pricing Is Coming for Every AI Company by Jamin Ball (March 2026).
Victorino Group helps organizations build governance frameworks for AI agent procurement and cost control. Let’s talk.
All articles on The Thinking Wire are written with the assistance of Anthropic's Opus LLM. Each piece goes through multi-agent research to verify facts and surface contradictions, followed by human review and approval before publication. If you find any inaccurate information or wish to contact our editorial team, please reach out at editorial@victorinollc.com . About The Thinking Wire →
If this resonates, let's talk
We help companies implement AI without losing control.
Schedule a Conversation