This is The Agentic Commercial Model, a newsletter about the operating model crisis hiding inside the AI gold rush
Here’s the thing nobody in SaaS wants to say out loud right now.
Usage-based pricing is not the answer to agentic AI monetization. It’s a component of the answer, confused with the whole answer, being deployed by companies who haven’t thought clearly enough about what’s actually changed.
The industry has convinced itself that the transition from seat-based to consumption-based models is the transformation. It isn’t. It’s a repricing exercise dressed up as a commercial reinvention. And the evidence that it isn’t working is already in the earnings reports, the customer forums, and the pricing pages that have been changed three times in eighteen months.
Let me explain what I mean, using the companies who’ve already run this experiment, and the bill they’re currently paying for it.
What “usage-based pricing” actually taught us
Before we get into AI, it’s worth revisiting where consumption models came from and what we learned from the first wave of companies that built their businesses on them.
Twilio made usage-based pricing famous. You paid per API call, per message, per minute of voice. The model was elegant: value delivered, value captured. It scaled beautifully in the growth era and was held up as the template every infrastructure company should follow.
Then the macro turned. When budgets tightened in 2022 and 2023, Twilio’s customers didn’t churn, they optimised. They cut message volumes, consolidated API calls, re-engineered workflows to consume less. The platform that had won on flexibility was now suffering because of it. Dollar-based net expansion rates collapsed. Not because customers stopped valuing the product. Because the commercial model had no floor, no friction, no stickiness beneath the variable consumption layer. Usage-based pricing in a downturn is an open invitation to cost-optimise your vendor.
Snowflake built a similar model around compute credits, you consumed what you needed and the bill followed the workload. This created genuinely impressive land-and-expand dynamics when customers were growing and adding use cases. But it also introduced something Snowflake disclosed repeatedly in its own filings: limited visibility into revenue timing. When customers have full flexibility over consumption, the vendor has full exposure to their discretion. Snowflake’s filings from 2023 to 2024 cite, again and again, the risk that customers consume more slowly than expected, particularly in response to adverse macroeconomic conditions. The model that creates ceiling-less upside also creates floor-less downside.
The lesson from the Twilio and Snowflake era isn’t that usage-based pricing is wrong. It’s that usage-based pricing, on its own, is an incomplete commercial architecture. It answers the question how do we capture more value when customers use more? It doesn’t answer how do we retain commercial value when customers use less, when they optimise, or when the unit of value itself changes?
That incomplete lesson is exactly what the industry is now applying, at speed, to agentic AI.
Salesforce: a pricing model changed three times in eighteen months
Agentforce launched in September 2024 with a price of $2 per conversation. The model was clean, intuitive, and immediately controversial.
The problem was definitional. What is a conversation? When a single customer query triggers eight backend processes, a multi-step data lookup, a policy check, and an escalation decision is that one conversation or five? The $2 per conversation model was elegant on a whiteboard and ungovernable in production. The backlash was significant and fast. Enterprise procurement teams couldn’t build a budget model around a unit they couldn’t define. CFOs couldn’t approve spend they couldn’t forecast. And SMBs did the maths and walked away.
By early 2025, Salesforce had moved to Flex Credits $0.10 per action. More granular, more aligned to actual work performed. But “action” turned out to carry the same definitional problem as “conversation.” One action triggered sub-actions. Budgeting became guesswork at scale.
By late 2025, Salesforce had introduced a third model: per-user licensing bundled into Salesforce Foundations. Seats again. The model they’d implicitly been moving away from, reinstated because enterprise buyers needed a number they could put in a spreadsheet. As of now, all three models run simultaneously.
Three pricing revisions in eighteen months is not iteration. It’s a company discovering, in public, that they didn’t have a clear answer to what their agent actually delivers and how that delivery should be valued. Salesforce has the brand, the relationships, and the enterprise contracts to survive this learning curve. Most vendors don’t.
HubSpot: the credit layer problem
HubSpot’s journey is instructive from a different angle. In early 2024, they shifted to a cleaner seat-based model, Core Seats, View-Only Seats, a simplified structure designed to scale with customer growth. A sensible move for a company whose complexity had become a sales problem.
Then came the AI layer. Credits on top of seats. Specific consumption costs per agent action, 100 credits per contact for the Prospecting Agent to monitor outreach opportunities, 10 credits per Workflow execution, 10 credits per company for buyer intent monitoring. The credits model is transactionally transparent. You can see exactly what you’re consuming and what it costs.
But here’s what transparency doesn’t solve: predictability. Customers who’d signed up for a known monthly seat cost now had a variable layer they couldn’t forecast. As one customer put it, the combination of fixed licence cost plus fixed AI add-on cost plus variable credits was simply “too much.” Not too expensive in absolute terms. Too complex to manage, budget for, or defend internally.
The problem isn’t that HubSpot got the price wrong. It’s that they layered a consumption model onto a subscription model without resolving the fundamental tension between them: subscriptions create budget certainty; consumption models undermine it.
The actual problem nobody is naming
Here’s what I think is happening, and why usage-based pricing alone can’t fix it.
When your product was a seat, the unit of value was obvious. A person used the software. The seat represented that person. The relationship between price and value was legible to every buyer.
When your product is an AI agent, the unit of value is a task completed, an outcome delivered, a decision made. None of those map cleanly to a conversation, a credit, an action, or an API call. The consumption metrics being used aren’t measures of value, they’re measures of activity. And buyers know the difference.
This is the consumption problem: the industry has correctly identified that seat pricing doesn’t work for agentic AI, and has replaced it with consumption metrics that also don’t work for agentic AI, because they’re proxies for value rather than measures of it.
A $2 conversation that deflects a customer service case worth £150 is priced correctly. A $2 conversation that runs eight backend processes, fails to resolve the issue, and escalates to a human agent is priced absurdly. The commercial model can’t tell the difference. That’s not a pricing problem. It’s a value architecture problem.
Zendesk understood this earlier than most. Their outcome-based model, charging per automated resolution, not per conversation or action, is the closest anyone in the market has come to pricing what actually happened rather than what was attempted. The model is still imperfect, still carries definitional challenges, and it only works because Zendesk invested heavily in the measurement infrastructure needed to verify what a “resolution” actually means. But the direction is right.
The progression isn’t seat → consumption. It’s seat → consumption → outcome. And most companies are stopping at step two and calling it done. The simple reason why they stop here is obvious, what is a consistent outcome per customer, per industry?
What the transition actually requires
If usage-based pricing is a component rather than the answer, what does the complete answer look like?
Three things have to be true simultaneously, and most companies are only managing one or two of them.
First: the unit of consumption has to track to a unit of value. Not a proxy. An actual unit of value, something the customer recognises as meaningful output. Resolutions, not conversations. Decisions automated, not API calls made. Time saved on a task the customer can quantify, not credits consumed on a process they can’t see.
Second: the model has to be forecastable for the buyer. Consumption pricing with no floor and no ceiling is not a commercial model. It’s a liability. Enterprise procurement teams need a number. They need to be able to defend a line item to a CFO. Hybrid structures, a committed base with consumption flex are not a compromise. They’re a commercial necessity for enterprise motion.
Third: the model has to survive customer efficiency gains. This is the test that almost nothing passes right now. If your customer uses your AI agent to automate a process and then optimises the agent to use fewer credits, your revenue goes down while their value goes up. That is a broken commercial model. It means you are economically punished for making your product better. The companies that will win in the agentic era are the ones whose commercial models get stronger as customers extract more value, not weaker. Own the value exchange, not as a one time activity, but a forever activity. Measure and track how customers gain value from your product, how this evolves, and how this is impacted by your roadmap. The companies that win the AI race will embed the monetization model into the product. Link every product decision to cashflow impact, this is what world class looks like!
None of these three conditions are met by simply switching from seats to credits.
What most companies will get wrong
Most companies will run the Salesforce playbook. They’ll launch with a consumption metric that sounds logical, per conversation, per action, per agent-hour, and they’ll discover in the first two quarters that their enterprise customers can’t budget for it. Your pricing model is a direct consequence of deals stalling. They’ll add a hybrid option to quiet the procurement teams. They’ll watch their SMB base optimise consumption when growth slows. And they’ll end up with three pricing models running in parallel, none of which is the answer, all of which require explanation.
The companies that avoid this aren’t the ones that find the right consumption metric faster. They’re the ones that start with a different question: what does value actually look like for the customer, and can we build a commercial model that tracks it directly?
That’s a harder question than “how do we price per use?” It requires investment in measurement infrastructure before the pricing model can be credible. It requires a product that delivers outcomes that are legible and verifiable, not just activities that are countable. And it requires a commercial team that can sell value delivered rather than usage committed. Today we barely operate on the minimal KYC principles, let alone the customer value journey, and delivering forever value to customers.
Most companies aren’t structured to do any of those three things. Which means they’re going to spend the next two years iterating through consumption metrics, changing their pricing page, and wondering why customer satisfaction and revenue momentum aren’t moving in the same direction. This lazy approach is directly contributing to a reduction in your Enterprise Value.
Usage-based pricing is necessary. It is not sufficient. And the gap between those two things is where a significant amount of SaaS value is about to disappear.
Next issue: The Broken Comp Plan, how legacy incentive structures are destroying AI-era commercial performance. If your sales team is still on a quota model designed for seat-based ARR, they are not equipped to sell agentic AI. Not because they lack capability. Because the model they’re working inside makes it economically irrational to try.
First published in the Agentic Commercial Model newsletter on LinkedIn, June 14, 2026. Read the original on LinkedIn.