Before HubSpot's AI Agents Go Live, Count the Runs

Picture the third week of a busy month. The customer agent on your website stops answering chats. Nothing is broken: the account used up its included HubSpot Credits, nobody bought an extra pack, and HubSpot's documented default for that case is that usage-based features pause until the next reset (Manage HubSpot Credits, HubSpot Knowledge Base). The agent did exactly what the settings said. The settings had no owner.

That week is avoidable, and HubSpot provides the tools to avoid it. It takes an hour with a spreadsheet before you switch anything on. This is the step I would add in front of our free HubSpot AI playbook, which covers which AI features are worth deploying and the setup each one needs. A disclosure before I go on: NetWebMedia sells its own CRM, so weigh my HubSpot opinions with that in mind. The billing facts below come from HubSpot's own pages.

What changed this summer

According to HubSpot's knowledge base, on July 23, 2026 all agents started consuming HubSpot Credits, and HubSpot strongly recommends simulating runs to estimate credit costs and setting monthly run limits to avoid unexpected charges (Test and control agent credit usage). Its credits article lists the customer agent, the prospecting agent, the data agent and Data Studio syncs among the features that require credits, and says credits are also used when a Breeze action runs inside a workflow (Manage HubSpot Credits).

HubSpot is open about all this, and I think usage-based pricing for agents is fair: an agent that does more work costs more to run. The risk sits on the customer's side, when nobody on the team is watching the meter.

Sort the jobs into drawers

Open the toolbox before you use it. List every AI job you plan to run in HubSpot and put each one in a drawer.

• Plain workflows. HubSpot's agent builder page says workflows come with Starter, Professional and Enterprise editions and do not consume HubSpot Credits (HubSpot agent builder). If a job is a fixed rule, like assigning a lead by region or sending a follow-up after a form, a workflow may be all it needs.
• AI steps and prebuilt agents. A Breeze action inside a workflow, the customer agent, the prospecting agent and the data agent all draw on credits, per the knowledge base articles above. Their cost follows how often they fire.
• Custom agents. Agents built in agent builder use credits on every run. HubSpot measures the work in action units and says the actions a custom agent takes vary from run to run (HubSpot agent builder), so two runs of the same agent can cost different amounts.

The first drawer is the one teams forget. Some jobs pitched as AI are rules in disguise, and a rule does not need a model to run it.

Estimate the runs, then test them

1. Count the triggers. For each job in the second and third drawers, estimate how often it will fire in a month: every new contact, every inbound chat, every deal that changes stage. Use last month's real volumes from your own portal.
2. Simulate before you switch on. HubSpot lets you run agent simulations that read CRM context without consuming credits or making changes, and shows the estimated credits for each simulated run (HubSpot Knowledge Base). Run a handful with sample data that looks like your real traffic.
3. Multiply and compare. Estimated credits per run times expected runs gives a monthly figure per job. Compare the total with the credits your subscription includes. HubSpot notes that estimates can differ from actual costs, so leave headroom.
4. Set a run limit per agent. Agent builder has a monthly run limit that shows an estimated credit range as you adjust it. Set one for every agent, even a generous one, so a surprise shows up as a limit rather than an invoice.

The sheet, row by row

Keep the whole exercise in one sheet with a row per job. The columns: the job in plain words, its drawer, what triggers it, how many times that trigger fired last month, estimated credits per run from your simulations, estimated credits per month, the run limit you set, what should happen when the limit is reached, and the owner. The last column is the one that gets skipped. Put a real name in it.

Sort the sheet by monthly credits, largest first, and read the top rows with fresh eyes. The expensive jobs are often the chattiest ones, like an agent that fires on every new contact, including the ones that arrive from a list import. Adding a filter to the trigger, such as only contacts from a web form, can change the cost of a job more than any setting inside the agent.

Decide what happens at the limit

The knowledge base describes two paths once included credits run out. Without extra packs, usage-based features pause until the next reset. With packs bought on or after September 16, 2026, the default is pay-as-you-go overage billed monthly in arrears, while accounts that bought packs before that date default to auto-upgrade, which moves the account to a larger pack for the rest of its commitment term (Manage HubSpot Credits). Credits reset every month, and unused ones do not roll over.

Neither default is wrong. They answer the same question differently: would you rather have an agent stop or a bill grow? A customer-facing agent going silent mid-month may hurt more than a modest overage. A prospecting agent pausing may be exactly what you want. Decide per job, write it down and name the person who checks usage each week.

Expect the names to keep moving

One practical note on reading any HubSpot AI guide, ours included. HubSpot's product names in this area have changed: its agent builder page says the tool was formerly known as Breeze studio, and HubSpot now presents its agents under Agent Hub (HubSpot Agent Hub). Match features by what they do rather than by what they were called when a guide was written, and check the current knowledge base for billing before you commit a budget.

Then make the features earn it

Counting runs tells you what the AI will cost. Whether it is worth that depends on setup, which is what the free HubSpot AI Playbook: Getting Real ROI from Breeze covers: which features deserve setup time, eight automations worth building first, predictive lead scoring configured on your own data, content settings that make output sound like your brand, the CRM data-quality fixes to make before any of it runs, and a 90-day upgrade roadmap. Bring your drawer list and your run estimates with you.

← Back to all articles