Between October 15 and October 19, HubSpot, SaaStr and diginomica each put one question in front of people who buy AI software: who holds the ceiling on the bill. They gave three different answers. Read together, they show the ceiling moving out of the contract and into settings the buyer controls.
What each account says
CMSWire on HubSpot’s spend controls
HubSpot announced new AI spend controls on October 15. The company turned spend controls on by default and sets an account-level credit limit of up to 50% above a customer’s monthly credits. Customers can also set limits at the feature and action level, swap seats for credits without waiting for renewal, and choose between flat-rate credit packs and Pay-as-You-Go billing for usage beyond their monthly credits. HubSpot says it will not upgrade a business into a bigger plan only because its usage is growing.
HubSpot says the controls build on a change earlier this year, when its Customer Agent and Prospecting Agent moved to outcome-based pricing so customers pay only when the work gets done. The new controls apply to credit use across the platform.
Advertisement
CMSWire’s report on the changes framed them as “easing budget anxiety around AI agents.” It placed the move in a wider pattern of pricing changes in AI software and pointed to Treasure AI, which the outlet says moved last week to charge for engagement instead of email sends. CMSWire also listed the places a buyer on consumption pricing should pressure-test a contract: seasonal spikes, the gap between pilot and production, minimum commitments, renewal pricing and usage visibility.
HubSpot’s own explanation is that a seat has a fixed price, while the credits AI uses vary with how much work the AI does. Jon Dick, Chief Customer Officer at HubSpot, said: “Nobody should have to miss out on getting the most out of AI because they don’t know what the bill will be.”
diginomica on what a contract can carry
diginomica’s Monday roundup led with a report on AI contract negotiation and came to a sober view of what a contract can do. It cited a survey of the diginomica network in which 65% of CIOs said AI was getting the largest share of their new technology investment this year, while 3% were very confident their governance was fit for the AI tools they were deploying and 12% admitted they were deploying faster than they could govern.
The report’s conclusion is that responsibility sits with the buyer. In diginomica’s words, “It comes down to you as a buyer to understand the risk you’re willing to accept and think about how you can internally manage against it.” It adds that the insurance market does not yet offer many options for protection, and that the EU AI Act calls the organization using a system a deployer, which for an enterprise buying AI software will often mean the buyer.
SaaStr on falling unit costs
SaaStr’s weekly takes, published October 18, come from the cost side. Jason Lemkin, founder of SaaStr, wrote that Haiku 5.5 is 90% cheaper than Haiku 4.5 for prompts under 100K tokens and that Sonnet 5.5 is about 10% cheaper than before. He added: “This is the first time in two years my AI costs have actually gone down.” He also wrote that token prices had kept falling but usage grew faster, so none of the portfolio companies he mentions saw their cost of goods drop until this month.
Where the accounts disagree
The three accounts do not share a remedy. HubSpot builds a ceiling into the product, so a customer’s exposure is bounded by a setting. diginomica argues that a ceiling in the contract is the wrong place to look, because the buyer keeps the risk whatever the paperwork says. Lemkin points the other way from both: his advice is to re-run model costs this month and “use the savings to run AI on every record instead of a sample.” One account caps usage, one questions the paper, and one treats cheaper units as a reason to use more.
They agree on the mechanism. In all three, the bill follows how much work the AI does, and the price of one unit of work is the smaller driver. HubSpot says as much when it describes credits as variable because they depend on how much work AI does. Lemkin reports that falling token prices did not lower costs until the price drop outpaced usage growth. diginomica’s survey shows organizations spending ahead of their ability to govern that usage.
What the three add up to
In our reading, the common finding is that AI spend risk has become an operating setting. A seat license put the risk in the contract, where finance and procurement handled it once a year. A credit model puts it in a dial that someone has to set, watch and adjust every month, and the person closest to that dial is usually in RevOps or sales operations.
The default HubSpot chose shows how that works. A limit of up to 50% above monthly credits bounds the damage. For a hypothetical account with 10,000 monthly credits, the ceiling would sit at up to 15,000. That protects against a runaway agent, and it still leaves a team free to burn through an extra half of its allowance in a bad month. A cap is a guardrail. Someone still has to decide how high it sits and what happens at the limit.
Our earlier coverage found the same pressure from the vendor side. Microsoft’s sales development agent starts consuming Copilot credits on November 16, and revenue software vendors now pitch outcomes as the product. Each of those moves shifts some variability onto the customer. HubSpot’s controls give the customer a default way to limit it.
What it means for the sales leader
Three things follow for anyone running a sales or revenue team that buys AI agents on credits.
First, treat the cap as a budget line with an owner. Whoever signs the contract is rarely the person who watches usage, so name one person in RevOps who reviews credit consumption by feature each month and has the authority to change the limit.
Second, set limits per action instead of one limit per account. HubSpot now allows limits at the feature and action level, which lets a team protect the work that books meetings from the work that only drafts notes. An account-wide ceiling will stop both at the same moment.
Third, track the cost of each outcome and leave the unit price of a credit out of the headline number. If a prospecting agent costs a set amount per meeting booked, a falling unit price helps only when the meeting count stays steady. Lemkin’s note that usage outran price for two years is the case to plan for. Our opinion piece on token-metered sales tech contracts argued for the same discipline, and the new controls make it easier to apply.
Before the next renewal, ask each vendor three questions in writing. What is the default limit, and who can change it. What happens to running agents when the limit is reached. Whether a rise in usage can move the account to a larger plan without a signature. HubSpot has published its answer to the third in a company post, and a written answer from every other vendor lets you compare them line by line.
Source: HubSpot
