The single biggest blocker we see when teams consider HubSpot's AI agents isn't scepticism about the technology. It's the fear of a surprise invoice. HubSpot credits are the unit that meters agentic work across the platform, and because some agents charge per outcome while others charge per run, the cost model looks opaque until someone maps it out properly. It isn't. Once you understand what consumes credits, how each agent is metered, and where the spend caps live, credits become one of the more predictable line items in your stack.
This is our working guide as a HubSpot Solutions Partner: what consumes credits, what each key agent costs, how to forecast the variable ones, and the controls we configure before we switch anything on.
Credits are consumed primarily by AI agents, plus a small number of non-agent features such as buyer intent. In practice, the vast majority of consumption we see today comes from deploying agents across go-to-market motions.
There are two ways to fund credit consumption:
Every plan — Starter, Professional and Enterprise — includes a number of credits by edition. This is the question we field most often, so it is worth stating plainly: using your included credits does not tip you into paid overage. If you have not purchased additional credits, you cannot accidentally incur extra charges. Consumption simply stops when the included allowance is exhausted, and the product tells the user the monthly total has been reached.
One caveat that catches people out: unused credits do not carry over to the next month.
Source: HubSpot
Agent Hub is the unified place to build, manage and run your agents. It lists every available agent and the outcomes each one is driving for the business, and it is where you create a new custom agent. You can reach it by several routes in the portal; the destination is the same.
Each agent is metered against the outcome it produces, not against a generic per-message charge. That distinction matters enormously when you model cost.
Source: HubSpot
The content agent generates content with full HubSpot context: answer engine optimisation (AEO) recommendations, brand, voice, identity and your contact base. A typical pattern we like: AEO flags that your site has no coverage of your loyalty programme, you prompt the content agent to action that recommendation, and it produces the post. You pay when the content is generated.
Source: HubSpot
The nurture agent sits inside workflows and automated nurture motions, personalising an email immediately before it sends. You control the prompt. The agent can review CRM data, past activity and business context, research the web, and then personalise subject line, body and calls to action. You pay per email personalised and sent — roughly €10 for 100 contacts.
Source: HubSpot
The flagship and longest-standing agent handles inbound questions across text-based channels such as live chat and email, and now across voice and calling channels too. Crucially, text-based conversations are charged per resolution, not per back-and-forth reply. A long conversation with eight exchanges that ends in one resolution is one charge.
A resolution means one of two things:
Source: HubSpot
The prospecting agent is paid per outcome: you are charged at the point of enrolment, when the agent drafts outreach for a recommended lead. It does not act autonomously — it must be prompted and configured in the portal — but enrolment itself can be automated. It reviews your addressable market, target audience and target accounts, sources warm-fit leads from buying signals, and recommends them for a play. Manually enrolled leads trigger research plus drafted outreach and are charged the same way.
Source: HubSpot
The data agent is pervasive across the platform and keeps the CRM current by updating properties. You pay when it updates an individual property — for example, reviewing past CRM activity or researching a contact on the web and writing the finding back into a HubSpot property.
Source: HubSpot
Custom agents are the most complex piece of the model and the part worth understanding properly. You define prompts, instructions, knowledge sources, guardrails, the things to care about and the things to ignore — much like working with an LLM directly. You pay per run, but cost per run is not fixed.
Source: HubSpot
Every action an agent takes consumes a number of action units — the metering unit for how much work the agent has done. At the end of a run, the total action units equate to the number of credits that run costs.
Take an agent triggered whenever a demo is booked. It must summarise the company record, research recent activity, personalise a script, create and save a note, and create a task to notify the company owner. The mechanical work — creating the note and the task — is consistent across every company. The agentic work is not. Summarising a brand-new company record with almost no CRM history costs materially less than summarising a decade-old record with years of activity behind it.
This leads to the most useful rule we apply: broad instructions are expensive instructions. Telling an agent to "search my entire CRM" creates enormous work per run. Guardrails and a defined order of operations make the agent run far more efficiently. Good agent hygiene is a cost control.
When we inspect the credits breakdown on a working custom agent, the pattern is consistent: most of the cost per run comes from LLM and context work — analysing, researching, sourcing from the web, interpreting portal data inside the agent's context. Very little comes from actions such as creating, saving or updating records. Those are bounded, predictable and cheap.
That is a surprising result for many admins, who assume writing to the CRM is the expensive part. It isn't. Reasoning is.
Because cost per run varies, the only honest way to forecast is to build the agent and test it. Testing is the part most teams skip, and it is the part that de-risks everything:
Two levers give you consistency. The first is the prompt: instruct the agent to review only the past month of activity rather than all historical data, and your outliers collapse. The second is representative testing across record types. HubSpot documents best practices for agent prompts, and they are worth following before you start engineering your own.
Source: HubSpot
Spend controls apply specifically to customers who have purchased additional credits. If you are living inside your edition's included allowance, they are not relevant to you — you cannot overspend.
The controls nest hierarchically, and that structure is the key mental model:
The sum of your action-level limits cannot exceed the feature-level limit, and the sum of your feature-level limits cannot exceed the portal-level limit. The purpose is deliberate: concentrate credit spend on the tools genuinely driving value, and put guardrails on everything else.
One change we welcome — HubSpot now sets a portal-level spend cap automatically the first time you buy additional credits. Runaway costs and shock invoices are prevented by default, and you can adjust the cap up or down at any time.
Source: HubSpot
Everything lives under Account & Billing in HubSpot. That is where limits are set and adjusted and where usage reporting sits: which features are on, consuming, not consuming or paused; individual actions; and historical usage trends. Access depends on permissions — admins and those granted access see the full reporting picture, including historical trends.
Source: HubSpot
You can pause a feature entirely if priorities shift and you never want it consuming credits. Pausing stops consumption until it is switched back on, which is excellent protection against accidental spend.
The misconception worth correcting: pausing does not let you keep using the feature for free. It pauses the credit-consuming feature itself, which may affect users and teams across the portal. Communicate before you pause.
Source: HubSpot
That second control is underrated. It moves cost awareness to the person taking the action, not just the person reading the invoice.
Source: HubSpot
We get asked constantly whether to build custom or use what ships in the box. Our default position:
On scale: there is effectively no hard ceiling on custom agents, in volume or complexity. Where an agent must reference external tooling consistently and frequently on every run, it is worth mapping whether to build it outside HubSpot on top of the MCP and connect it back in via connectors to the core LLMs. That can produce a more centralised stack. It is a solutioning decision, not a limitation.
A fair challenge: why spend credits on the content agent when Copilot, Claude or GPT will draft copy at effectively no marginal cost?
Our honest answer is that they solve different problems. General LLMs produce raw copy efficiently — you get a drafted document. What you do not get is a publish-ready, optimised page. The content agent's advantage is proximity to your business context:
The practical drawback of external LLMs is repetition: restating brand rules on every prompt — "our brand hates the Oxford comma" — is a tax you pay forever. A properly configured content agent already knows. For pure copy editing and drafting, an external LLM is perfectly fine, and we use both in different capacities. Agentic tools are only ever as powerful as their understanding of your business.
If credit anxiety is what is holding you back, sequence it like this:
Yes. Included credits can be spent across any credit-consuming feature, up to the included threshold. As a rough illustration, if you have 3,000 included credits and a lead costs 100 credits to enrol, that is around 30 leads a month — the 31st cannot be enrolled, and the product tells you the monthly credit total has been reached. Treat the arithmetic as indicative only and check the current rate sheet for the credit values that apply to your portal.
No. If you have not purchased additional credits, there is no automatic overage. Overage only applies where additional credits have been bought — and even then, HubSpot now sets a portal-level spend cap by default.
No. Unused credits do not carry over to the next month.
Build it, test it across a representative spread of old and new records, and use the average cost per run from those tests. Tighten the prompt — for example, limiting the agent to the last month of activity — to reduce outliers.
Yes, on the LLM and context side rather than the tool and action side. More context to process and more reasoning to perform is what drives cost, not the record updates at the end.
Not because of scale. The trigger is dependency on external tooling. If an agent must reference external tools consistently on every run, map whether building on top of HubSpot's MCP and connecting back in gives you a cleaner, more centralised architecture.
Credits stop being frightening the moment you treat them as an operational design problem rather than a billing mystery. Build one custom agent, test it before you publish, read the credits breakdown to see where the reasoning cost sits, then set your portal and feature caps to match the budget you actually have.
If you would like a second pair of eyes on your agent architecture, your prompt hygiene or your spend cap structure before you scale, our team works on exactly this kind of HubSpot problem every week. Start a conversation with us — bring your use case and we will map the cost model with you.