Salesforce, HubSpot and Atlassian are now pricing AI agent access separately from seats, and the rates don't agree

Salesforce, HubSpot and Atlassian have each rolled out a separate price meter for AI agent traffic this year, splitting agent usage off the human-seat price they've charged for years — and the going rate for that meter differs by as much as 1,200 times depending on which vendor, and which caller, is asking.

The clearest account of the pattern came September 28 from SaaStr founder Jason Lemkin, who tracks the vendor stack his own company runs on. Lemkin's piece names HubSpot and Salesforce as the two most recent examples and says a niche CRM and another unnamed tool have quietly done the same. Each vendor drew the line differently: what counts as a billable agent action, whether existing API capacity is exempt, and how far ahead customers get to see the rate before it applies.

What changed

HubSpot moved first, and transparently. Since April 14, 2026, its Customer Agent no longer charges a flat $1 per conversation regardless of outcome — it now charges $0.50 per resolved conversation, billed as 50 HubSpot credits, only when the agent actually closes the case. The Prospecting Agent flipped from a recurring monthly fee per enrolled contact to $1 per lead it recommends, billed as 100 credits. HubSpot points to two numbers behind the switch: across the more than 8,000 customers already running Customer Agent, it closes about two in three conversations on its own and shaves over a third off how long a resolution takes — and Prospecting Agent adoption climbed 57% quarter over quarter once it launched.

"Outcome-based pricing removes that risk. You pay when it works, full stop."

That's Jon Dick, HubSpot's chief customer officer, on why the company moved off flat per-conversation billing. Salesforce's Agentforce runs on a different mechanic, Flex Credits: $500 buys 100,000 credits, a standard action costs 20 credits (about ten cents), and a voice action costs 30. That part is published on Salesforce's own pricing page and stable. What isn't yet public there, according to Lemkin's account of what Salesforce has told its customers, is a second meter coming for agent traffic that arrives through MCP or the API rather than the Agentforce interface — every successful call would draw Flex Credits, agents would need to be registered, and existing customers would move to the new billing at their next renewal. Lemkin estimates the proposed rate at $5,000 to $100,000 per million calls, against roughly $83 per million for the API capacity Salesforce already sells for human-built integrations today.

Atlassian took a third route, and Lemkin holds it up as the better-run version. Its Rovo credit system already counts calls made through the Teamwork Graph CLI and the Rovo MCP server, and overage billing begins December 3, 2026 at $0.01 per credit — a basic action running about ten cents, close to Salesforce's rate. Paid plans carry a pooled monthly allowance (25 credits per user on Standard, 70 on Premium, 150 on Enterprise), and pulling data through those channels isn't metered yet. What makes it land differently, in Lemkin's telling, isn't the price:

"Atlassian published the rate, published the date, and gave admins a switch. That's the difference between a meter you can plan around and one you can't."

Who it affects

This is a live pricing-page problem, not just a vendor-bill problem. Every one of these platforms is also a reference point for how a software company prices its own product's agent and API access — and there's no shared convention yet for a packaging team to lean on. A pricing lead deciding whether to meter agent calls, fold them into the seat, or price the outcome instead is watching three large vendors solve the same problem three different ways, with the gap between a published rate card and a quiet renewal email doing as much reputational work as the number itself.

It also changes vendor evaluation for marketing teams buying tools. Lemkin describes the instinctive response to an unpublished or steep agent meter: sync the vendor's data into your own database and have agents read from the copy instead of the live API, so less usage ever reaches the meter. He argues that pattern erodes the thing that made a system of record sticky in the first place — usage, not the software itself.

What to do differently

• Decide now whether your product's agent and API access sits inside the seat price or on top of it, and put that decision on the pricing page instead of leaving support to explain it at renewal
• If you do add a per-call or per-credit meter, publish the rate and the effective date in advance, the way Atlassian did, rather than send it only in an account email
• Price the outcome an agent produces where your data model allows it, not the number of times it queries your API — HubSpot's move from per-conversation to per-resolution is the template