Salesforce will soon charge every time a registered AI agent successfully calls into its platform, whether that call arrives through MCP or a direct API. It set out the model in a help article on 17 September 2026 and is targeting November for the billing and controls to go live.
For as long as most of us have bought CRM, the unit has been the seat. You pay for each person who logs in. That works when people are the ones doing the work inside the system, and it gets shaky when agents start doing it for them.
HubSpot is heading in a different direction. It charges for its own agents when they deliver a defined result, keeps its MCP connection free, and is piloting a simpler seats-plus-credits model with new customers in Europe. Both are reasonable answers to the same problem, and both will be harder to budget for than a seat.
MCP stands for Model Context Protocol. It's an open standard, introduced by Anthropic in late 2024, that gives AI tools one consistent way to connect to other software. Before it, every AI tool needed a custom integration for every system it wanted to read from. Now a software vendor publishes one MCP server, and any compatible AI client can use it.
Think of it as a USB-C port for software. The CRM provides the port, and Claude, ChatGPT, Copilot or a home-built agent plugs in.
In practice, you ask Claude which deals over £50,000 have had no activity in 30 days. Claude uses the CRM's MCP server to find the right objects, search the records, read them and, if you allow it, update them. It works with your permissions, so it only sees what you could see yourself.
Both vendors run one. HubSpot's remote MCP server became generally available on 13 April 2026 and powers its Claude and ChatGPT connectors. Salesforce's hosted MCP servers followed on 29 April.
One detail matters for everything that follows. A single question from a person rarely means a single call. The agent might look up the data structure, search, read three records and write a note, and each of those steps is a separate interaction with the CRM.
"Headless" means the system without the screen. The data and the logic stay in the CRM, but people stop clicking through it, because an agent does the updating, searching and logging in the background and they get answers in Slack, Claude or wherever they already work.
Salesforce has gone furthest here. At Dreamforce it launched AIforce, a layer that opens the whole platform by API, MCP and command line, with the browser interface treated as optional. That's a bold move for a company whose revenue is built on per-user licences. If a handful of seats can power a CRM that agents keep busy all day, the seat count stops reflecting how much the system is used, and the revenue attached to it starts to shrink.
HubSpot has framed the same shift differently. CEO Yamini Rangan's rule, set out at HubSpot's September Analyst Day, is that AI which helps a person work more effectively belongs in the seat, while AI that does work beyond the people using the software belongs in credits. Her shorter version is that "the quantity goes down only when the value goes up", as Steven Forth's analysis of HubSpot's pricing records it. HubSpot's paid seats have grown about 80% since its 2024 pricing change, to more than 2.5 million, so it's adding a second meter rather than defending the first.
Under the new model, each external AI agent has to be registered and given its own identity, credentials and permissions, rather than borrowing a human user's login. Once it's registered, every successful call it makes to a production org counts as a Headless Platform Interaction (HPI) and draws down Flex Credits. Salesforce's own description is that registration makes an agent's activity "visible, governable, and billable" (Salesforce Help).
The governance part is a genuine improvement. Knowing which agent did what, with which permissions, is something most CRM owners will want.
The pricing is harder to plan around. Flex Credits list at $0.005 each, but Salesforce hasn't yet published how many credits one HPI consumes, and has promised 30 days' notice before metering starts. Sandboxes and developer orgs aren't metered, and traditional non-agentic integrations keep their current pricing. Moving an agent from MCP to the REST API doesn't avoid the charge, because both routes are metered.
The main criticism, raised by Salesforce Ben and Aviando among others, is that customers using Claude or ChatGPT end up paying twice: once to the AI provider and again to Salesforce for access. Because cost scales with the number of calls, a chatty agent costs more than a disciplined one for the same result. That creates an incentive to copy data out of Salesforce to save money, which is the opposite of the clean, governed setup the registration step is meant to encourage.
HubSpot opened UNBOUND in September by declaring the Outcomes Era: the idea that AI should be judged on the business results it produces, given the right data and context, and paired with human judgement. Rangan summed up the shift in HubSpot's Q3 results as a move "from tools that help people do work to agents that do the work".
The pricing had already started moving. Since 14 April 2026, Breeze Customer Agent costs $0.50 per resolved conversation, down from $1 per conversation, and Breeze Prospecting Agent costs $1 per lead recommended for outreach. If the visitor ends up asking for a person instead, there's no charge. HubSpot's own line is that "AI should be measured in outcomes, not output" (HubSpot).
Everything settles in HubSpot Credits, priced at one cent each. Starter, Professional and Enterprise include 500, 5,000 and 10,000 credits a month, and admins can set limits by account, feature, action and agent. The overseas pilot sits on top of this: new customers in the Nordics and Benelux get a simplified model of seats plus credits, with spend that can shift between the two, and HubSpot is extending it to new customers across EMEA this quarter.
The contrast with Salesforce is sharpest on MCP. HubSpot's MCP server and its Claude and ChatGPT connectors carry no separate charge on any plan, so a team can point Claude at its HubSpot data today without a second meter running.
There are caveats worth knowing:
HubSpot has also been open about why it moved. Rangan told analysts in August that "predictability has become a defining theme in AI adoption" (Constellation Research), and the price cut has cost HubSpot some short-term revenue growth.
Salesforce charges for access, and HubSpot charges for results. That single difference shapes who each model suits.
|
Salesforce |
HubSpot |
|
|---|---|---|
|
What triggers a charge |
Each successful call by a registered agent, over MCP or API |
A defined result from a HubSpot agent; most other AI features are priced per action |
|
Unit price today |
Flex Credit at $0.005; credits per call not yet published |
Credit at $0.01; $0.50 per resolved conversation, $1 per recommended lead |
|
Claude or ChatGPT via MCP |
Metered once the agent is registered |
No separate charge |
|
What drives the bill |
How many calls each agent makes |
How many results the agents deliver |
|
Spend controls |
Digital Wallet usage alerts |
Limits by account, feature, action and agent; pay-as-you-go or fixed packs |
|
Best suited to |
Enterprises with engineering teams building their own agents, who want each one governed and can model call volumes |
Mid-market teams who want AI doing defined jobs like support and prospecting, with spend tied to finished work |
|
Main budgeting risk |
An unpublished rate, and call counts that climb as agents get busier |
Spend rises as agents improve, and monthly credits don't roll over |
For most mid-market revenue teams, HubSpot's model is easier to explain to a finance director, because each charge maps to something the business recognises. Salesforce's model gives more control to teams with the engineering capacity to optimise how their agents call the platform.
The shift that matters is planning AI spend as a cost per result rather than a licence. Six steps make that workable.
A service team handles 2,000 conversations a month and Customer Agent resolves 65% of them, the rate HubSpot published in April. That's 1,300 resolutions at $0.50, or $650 a month (65,000 credits). The 700 conversations that reach a person cost nothing on the HubSpot side. The number to compare it against is what those 1,300 conversations would cost your team to handle.
An agent completes 10,000 tasks a month, and each task takes six calls to Salesforce: 60,000 HPIs. Because the credits-per-call rate isn't published, the honest answer is a range. The rates below are planning scenarios modelled by Stacksync, not Salesforce prices.
|
Credits per HPI (scenario) |
Cost per call |
Monthly cost for 60,000 calls |
|---|---|---|
|
1 |
$0.005 |
$300 |
|
5 |
$0.025 |
$1,500 |
|
20 |
$0.10 |
$6,000 |
A twentyfold spread is hard to put in a budget. It also moves with agent design: cut the task from six calls to three and the bill halves, whatever the rate turns out to be. Your AI provider's charges sit on top of either figure.
Neither model is the finished article. Both are first attempts at pricing software that does work rather than just stores it, and both will be hard for teams to budget around in their current form.
My view is that buyers won't accept an open-ended meter for long. Where this settles is committed annual spend that can move between seats and credits, priced in units a finance director recognises: a resolved ticket, a booked meeting, a qualified opportunity. HubSpot's EMEA pilot is already close to that, because it lets customers shift spend between people and agents as their mix changes.
Salesforce's per-call meter charges for the plumbing rather than the result. That puts a toll on the open, multi-tool setups that make agents useful, and I'd expect the rate to be negotiated hard once it's published.
HubSpot's harder question is how it defines the outcome. A lead recommended for outreach is still a long way from a booked meeting, and the closer the billable unit gets to revenue, the easier it is to justify.
We're a HubSpot Elite Partner, so weigh my view accordingly. Whichever platform you're on, the practical step now is the same: know which of your AI costs are per seat, per action and per result, and set a ceiling on each before your agents get busy.