We configure and govern Breeze AI agents inside your CRM so they research, route, and resolve real work, with a human always able to check what they did and why.
Book a discovery callEvery agent is scoped to a real bottleneck your team already feels. Filter by where you want to move first.
Resolves support inquiries across channels using your knowledge base.
Finds buying signals, sources contacts, drafts outreach for review.
Enriches records and answers custom questions from CRM and web data.
Turns one piece of content into emails, social posts, and landing pages.
Agent actions built into your own workflows and deal stages.
HubSpot's AI agents are moving fast, and since 23 July 2026 every agent run draws HubSpot Credits. We test new agents and betas in our own HubSpot portal before we recommend them to a client, so you know what they cost and where they break before your team relies on them.
Talk to our teamIV-Lead, a HubSpot Gold Solutions Partner in Tel Aviv, implements and governs AI inside HubSpot for B2B companies in Israel. There are five routes: Breeze agents, HubSpot's custom-LLM workflow action, the HubSpot connector for Claude, HubSpot's Agent CLI, and a custom integration. Which one fits depends on who drives the work, what it is allowed to write, and which HubSpot tier you hold. Four of the five are HubSpot's own and need no code from us.
Use the custom-LLM workflow action, which HubSpot ships on Enterprise hubs. Connect your own provider key — Anthropic for Claude, OpenAI for ChatGPT; Cohere, xAI and Google are supported too — write the prompt against the record data you choose, and the action returns an output that later steps can write to a property and branch on. That is the automation route, and it runs unattended, so the approval step is something you design rather than something you get. Asking Claude questions through the connector, and letting an AI coding agent work through the Agent CLI, are different jobs with different permissions — the table below separates them.
| Route | Best for | Requires | Reads / writes | Approval and audit | What it costs | Limits and what we test first |
|---|---|---|---|---|---|---|
| 1 · Breeze agents — HubSpot's own agents, running HubSpot's models (not your Claude or ChatGPT account) | Support resolution on chat, email and WhatsApp; prospect research; record enrichment; a custom agent for a bottleneck of your own | Professional or Enterprise hub plus HubSpot Credits | Reads and writes CRM records inside the portal; the customer agent answers from your synced content | Hands off to a human on low confidence or your escalation rules; testing a custom agent consumes no credits; monthly run limits are configurable | Credits — $0.50 per conversation resolved for the customer agent; custom agents estimate their credit cost before launch | Hebrew behaviour of the customer agent, and resolution quality against your actual knowledge base |
| 2 · Custom-LLM workflow action — native, your own model provider | Summarise a record, classify what arrives, draft a personalised text, then branch the workflow on the result | An Enterprise hub; your own API key from Anthropic, OpenAI, Cohere, xAI or Google | Sends your prompt with the record context you choose; the output can be written to a property and used by later steps | Runs unattended by default. A review task does not pause a workflow — you need an explicit approval wait and an approved branch before anything customer-facing | Your provider usage on your own key; HubSpot states no extra fee for the action | HubSpot publishes no token or rate limits for it; the action carries a BETA label on the workflow-outputs page; prompt injection from text inside records |
| 3 · HubSpot connector for Claude — a person drives, in Claude | Account briefs, pipeline questions, drafting notes and tasks, ad-hoc analysis, tidying up to 10 records at a time | Any HubSpot plan; a paid Claude plan (Pro, Max, Team or Enterprise); a Super Admin or App Marketplace user connects it, then grants access — a grant that cannot be undone | Creates and updates contacts, leads, companies, deals, tickets, custom objects, pipelines, engagements, landing pages, website pages, blog posts and marketing emails; creates but does not update campaigns and marketing events; read-only on invoices, orders, payments and conversations; never deletes | Each user sees only what their HubSpot permissions allow; with Sensitive Data switched on the connector gets no engagement data at all; conditional property rules and pipeline stage validations apply to its writes, association label validations do not | $0 in the marketplace; the Claude plan is the cost | The 10-record cap; Anthropic's terms for data shared through it; it is built around a person in the loop rather than scheduled runs |
| 4 · HubSpot Agent CLI (beta) — an AI coding agent works on the portal | Bulk clean-ups, property and pipeline surgery, building and fixing workflows, portal administration with an agent doing the typing | All HubSpot products and plans; Claude Code, the Claude app, the ChatGPT app or another compatible agent; a signed-in HubSpot user | Acts with the permissions of the user who signed in: creating, updating and deleting records, and reaching workflows, pipelines, properties and associations | The permission scope is the guardrail, so we scope a dedicated user rather than running it as an admin, and we work against a sandbox first | No HubSpot fee stated; your Claude or ChatGPT plan | It is beta, and it can delete — the connector's caps do not apply here. We agree a blast radius and a restore path before it touches a production portal |
| 5 · Custom integration — private app plus a model API, or an integration layer | Multi-step processes, ERP in the loop (Priority, SAP Business One), Hebrew-specific logic, anything the built-in routes cannot express | A private app with scoped access; hosting; our build | Whatever the scopes allow and nothing more; every write goes through our queue | Approval step, retries with persistent operation state, an audit log, and a restore path agreed before go-live | Build and support from IV-Lead; hosting; your provider usage | HubSpot API limits: 190 requests per 10 seconds per private app on Professional and Enterprise, against a daily ceiling shared across the whole account (625,000 or 1,000,000) |
Tiers, limits and prices as published by HubSpot and checked on 6 September 2026. HubSpot changes them without notice; we re-check monthly, and the date above is the last time we did.
What HubSpot does on its own: the customer agent hands off to a human on low confidence or your escalation rules; the connector cannot delete and caps bulk edits at 10 records; Sensitive Data blocks engagement access entirely. What we build into an engagement: least-privilege scopes for every private app and connector user; an explicit approval wait with an approved branch before anything customer-facing, because a review task by itself does not stop a workflow; allow-listed outputs with everything unrecognised sent to a review queue; persistent operation state so a retried run cannot write twice; and keys kept in HubSpot's provider connection or secrets store, never in a prompt or a property. What is proven: only what has been tested in your portal. We separate what is reversible (property history) from what needs a compensating action and what cannot be undone at all — workflow re-enrolment restarts processing, it does not undo it — and we agree that restore path before an agent gets live scope.
Use HubSpot's built-in custom-LLM workflow action, available on Enterprise hubs. You connect your own provider key — Anthropic for Claude, OpenAI for ChatGPT, and Cohere, xAI or Google are also supported — write the prompt, and the action's output becomes a value later workflow steps can write to a property and branch on. That is the automation route. A person asking Claude questions through the HubSpot connector, and an AI coding agent working through the Agent CLI, are different jobs with different permissions.
Usually not. The custom-LLM workflow action, the HubSpot connector for Claude and the Agent CLI are all HubSpot's own, and none of them requires code from us. Breeze agents are HubSpot's own agents running HubSpot's models, which is a different thing from calling Claude or ChatGPT yourself. Code or an integration layer is the fifth route, for multi-step processes, ERP in the loop, or logic the built-in routes cannot express.
Through the HubSpot connector, Claude can read, create and update contacts, leads, companies, deals, tickets, custom objects, line items, products, pipelines, landing pages, website pages, blog posts, marketing emails and engagements; it can create but not update campaigns, marketing events and properties; quotes are in beta; invoices, orders, payments and conversations are read-only. It cannot delete anything, and bulk changes are capped at 10 records. If Sensitive Data is switched on in your account, the connector gets no engagement data at all. The Agent CLI is different: it acts with the permissions of the person who signed in, and it can delete.
The connector for Claude and the Agent CLI work on all HubSpot products and plans. The custom-LLM workflow action needs an Enterprise hub. Breeze agents need Professional or Enterprise plus HubSpot Credits. A custom integration runs on any tier, inside that tier's API limits.
Anthropic does not publish a Hebrew benchmark and tells customers to test their own languages, so we test each Hebrew process end to end before it goes live rather than assume. HubSpot says its customer agent is available in all HubSpot-supported languages but publishes no agent-specific list, and Hebrew is a HubSpot content language rather than a software-interface language — so for that agent specifically we verify behaviour in your portal before recommending it for Hebrew-facing support.
For the connector, a Super Admin or a user with App Marketplace permissions connects it once; granting other users access is a separate step that cannot be undone, and each person then sees only what their HubSpot permissions already allow. Sensitive Data properties are never exposed. The Agent CLI inherits the permissions of whoever signed in, which is why we scope a dedicated user for it rather than running it as an admin.
Three different things, and it is worth keeping them apart. HubSpot's own behaviour: the customer agent hands off to a human on low confidence or your escalation rules; the connector cannot delete and caps bulk edits at 10 records. What we build into an engagement: an explicit approval wait and an approved branch before anything customer-facing (a review task on its own does not pause a workflow), allow-listed outputs, a review queue for anything unrecognised, and keys held in HubSpot's provider connection or secrets store rather than in a prompt. What is proven: only what we have tested in your portal, which we run and report before an agent is given live scope.
The connector is listed at $0 in HubSpot's marketplace; the paid Claude plan (Pro, Max, Team or Enterprise) is separate. The custom-LLM action bills your own provider usage. Breeze agents consume HubSpot Credits — the customer agent is priced at $0.50 per conversation resolved, and custom agents show an estimated credit cost before you launch them. A custom integration adds hosting plus our build and support. We size all five against your real volumes on a diagnostic call.
Your team owns the outcomes and the approvals. IV-Lead owns the configuration, monitoring and changes under a support agreement, or hands over with documentation and training if you would rather run it yourselves.
Technical review: Chen Yehoshua, IV-Lead. Facts checked 6 September 2026 against HubSpot's documentation for the custom-LLM workflow action, workflow action outputs, the connector for Claude, the Agent CLI, the customer agent, the agent builder and HubSpot Credits, HubSpot's marketplace and product pages, HubSpot's API usage guidelines, and Anthropic's multilingual-support documentation. The connector's object access was re-checked on 2 October 2026 against HubSpot's article updated 24 September 2026. Each route links to its primary source above.
Book a 30-minute diagnostic call