Skip to content

Running AI Agents in HubSpot: Governance, Monitoring, and KPIs

How to run AI agents in HubSpot after go-live: roles, governance rules, monitoring of quality and credits, and KPIs that show the value.

Once they are live, AI agents in HubSpot need one owner per agent, clear rules for access and approvals, and a regular check on quality, credits, and value. In the long run, your own team takes on these tasks. Until roles and routines are in place, a partner runs operations and enables your team in parallel. At Mark Lotse, we do this with Managed AI Services during the first three to six months.

Before go-live, the focus is on data quality and choosing the first agent, as described in our article Preparing HubSpot Data for AI Agents.

What changes after an AI agent goes live?

After go-live, the project becomes an ongoing task. Products, prices, and processes change, and every change has to reach the knowledge sources. For the customer agent, HubSpot automatically re-syncs knowledge base articles as soon as you update them. According to the help article, HubSpot re-syncs website pages, files, and imported URLs once a week. So if you change a price on your website, trigger the sync yourself with Refresh.

Then there are the costs: since July 23, 2026, all AI agents consume HubSpot Credits, according to HubSpot. Without clear ownership, sources go stale and nobody notices rising usage.

Which roles do you need to run AI agents?

In operation, the four roles from the implementation remain. Their tasks shift from building to maintaining and reviewing. In smaller teams, one person often covers several roles.

  • Executive sponsor: approves the budget for credits and new use cases and receives the results every quarter.
  • Business owner per agent: knows the process, reviews samples, and decides on changes to instructions and knowledge sources. For the customer agent, this is someone from service; for the prospecting agent, someone from sales.
  • CRM and RevOps ownership: maintains the agents' permissions, actions, and limits. According to HubSpot, anyone who creates and customizes agents in the agent builder needs Super Admin rights or the Agent Builder permission. When an agent runs through a workflow, it uses the permissions of the agent owner, according to HubSpot. If that person lacks access, the affected actions fail. So check this assignment whenever someone leaves the company or changes roles.
  • Enablement: trains new colleagues and collects feedback from the team. Over time, a small core team, often called a Center of Excellence, pools this knowledge.

How these roles take shape during implementation is described in our roadmap for AI in marketing and sales.

Which rules belong in AI agent governance?

Governance defines what an agent may do, who changes it, and who approves its results. For custom agents from the agent builder, HubSpot offers several settings for this (as of November 2026).

  • Access: For each agent, you define who runs it and who edits it. By default, all users can run an agent, and only Super Admins can edit it. Review this default for every agent that writes data or communicates externally.
  • Actions and knowledge: You choose the actions an agent may perform, such as reading or writing CRM records, and the knowledge sources it can access. Only enable what the use case needs.
  • Approvals: In the agent inbox, runs waiting for an action to be approved appear with the status Needs approval. Decide who reviews these runs and how quickly.
  • Run limits: You set a monthly run limit for each agent. Once it is reached, the agent stops.
  • Testing changes: Test new instructions or sources in a simulation first. According to HubSpot, simulations don't consume credits. Connected MCP tools can still run, so use sample data.
  • Documentation: For each agent, record what it is for, who is responsible, and when instructions or sources changed. That way you can trace why a result has changed.

How do you monitor AI agents in HubSpot?

You monitor three things: the quality of the results, handoffs to humans, and credit usage. For the customer agent, HubSpot provides dedicated reports; for custom agents, the agent inbox.

Customer agent: reading the resolution rate correctly

Under Analyze, HubSpot shows reports on the resolution rate, human handoffs, and the share of conversations without human involvement (deflection rate). Three details determine how you read the numbers:

  • Delay: HubSpot only evaluates whether a conversation is resolved 72 hours after the customer's last reply. The current week therefore shows a resolution rate that is too low.
  • Deflection shows relief: It counts every conversation handled without a human, even if the customer left without a suitable answer. Quality shows in the resolution rate.
  • Manual assignment: Only a handoff triggered by the customer prevents a resolution, for example a request for a human or a configured handoff phrase. If your team assigns a conversation manually or via workflow, it can still count as resolved. It does, however, appear in the human handoffs report.

Under Coaching opportunities, HubSpot collects conversations in which the agent lacked knowledge, sources contradicted each other, or an action failed. From there, you make improvements in the Knowledge, Actions, Handoff, and Experience tabs.

The Source Usage tab sorts your knowledge sources into four groups: frequently used with a high resolution rate, frequently used with a low resolution rate, rarely used with a high resolution rate, and rarely used with a low resolution rate. The second group deserves your attention first, because the agent draws on these sources often and rarely resolves anything with them.

For these analytics, HubSpot requires a Professional or Enterprise edition, such as Service Hub, Smart CRM, or Revenue Hub, plus HubSpot Credits (as of November 2026).

Custom agents: reviewing runs in the agent inbox

You review runs of custom agents in the agent inbox. For each run, HubSpot shows the status, trigger, user, and estimated credits, along with the output, thinking steps, and sources used. The status filter shows failed runs and runs waiting for approval. For the inbox, a Super Admin enables access to generative AI, CRM data, and Breeze Assistant, among other settings, in the AI settings.

Credits: what agents consume and when they stop

How an agent consumes credits depends on its type. According to HubSpot, the customer agent consumes credits only for resolved conversations, regardless of the number of messages. Custom agents from the agent builder consume credits per run, depending on how much they think, act, and generate. HubSpot shows the estimated cost per run in the agent inbox and in the simulation.

You see usage under Account & Billing in the HubSpot Credits tab. There you set limits for the whole account, for individual features, and, for buyer intent and the customer agent, per action. At 75, 85, and 90 percent usage, HubSpot sends alerts to the billing contacts. If none are defined, they go to Super Admins and Billing Admins (HubSpot, as of November 2026). In the new pricing model, emails and almost all workflow actions also consume credits (HubSpot).

If you only have the included credits and they run out, all features that use credits pause until the next monthly reset or until you buy more. In the new pricing model, this also stops emails and workflows. With purchased capacity packs, usage continues until the account's spend limit is reached. By default, HubSpot sets it 50 percent above the monthly allotment and either bills the overage or upgrades your pack, depending on your billing setting. Unused credits expire at the end of each month.

So plan run limits and feature limits so that a single agent leaves enough credits for the rest of your processes.

Review cadence for operations

This split works as a starting point, which you adapt to volume and the number of agents:

WhenWhat you review
WeeklyCoaching opportunities for the customer agent, runs with Needs approval, and failed runs in the agent inbox
MonthlyCredit usage per feature and agent, run limits, Source Usage
With every change to products, prices, or processesUpdate knowledge sources, trigger a sync with Refresh, test instructions in a simulation
QuarterlyKPIs with the sponsor, roles and access, new HubSpot features

Which KPIs measure the success of an AI agent?

You measure success with a before-and-after comparison. Each agent gets baseline values before go-live, which you regularly compare with current values. Three groups of KPIs belong here.

  • Quality: resolution rate, share of handoffs to humans, and the results of your sample reviews. They show whether the agent works reliably.
  • Impact: time saved, response times, qualified meetings, or revenue effects per use case. They show the value for sales and service.
  • Cost: credits per resolved conversation for the customer agent, credits per run for custom agents. They show whether the agent pays off.

Review the values once a quarter together with the sponsor and adjust tasks, sources, and limits. For reliable statements on pipeline and revenue, you need at least one quarter. Which metric fits individual AI workflows is shown in our article AI Workflows in HubSpot.

Who looks after AI agents in ongoing operation?

At Mark Lotse, we look after AI agents in HubSpot with Managed AI Services, as a HubSpot Platinum Partner with offices in Landau in der Pfalz and Bremen. We enable your team to steer operations on its own. During the first three to six months, we jointly maintain agents, prompts, rules, and knowledge bases, monitor quality with defined thresholds, and deliver a quarterly outcome report to management. The support runs in four steps: Takeover, Guidance, Handover, and Sparring. After that, your Center of Excellence steers operations.

In the LOTSE model, with the stages Landscape, Orientation, Trial, Scale, and Earnings, operation begins after the trial. We support setting up an agent in Stage T, described on our page Breeze & Agents. Our Managed AI Services cover Scale and Earnings.

Scope, process, and how to get started are on our page Managed AI Services for HubSpot.

Questions about running AI agents in HubSpot

Let's set your course.

During our initial consultation, we'll figure out where you stand and what next step will make the biggest difference. 30 minutes, to the point, with no obligation.