Short answer
AI connects to HubSpot through Agent Hub’s built-in agents, which run on HubSpot Credits, through the CRM API, or through HubSpot’s remote MCP server, generally available since 13 April 2026. For New Zealand teams, account research enrichment works best when every claim written to a record carries a source URL and date, and contact-level enrichment is planned around IPP 3A of the Privacy Act.
- HubSpot’s remote MCP server reads and writes CRM records and respects existing user permissions.
- HubSpot Credits cost US$0.010 each; the Data agent uses 10 credits per prompt per record.
- Write AI research to separate properties, never over fields your reps maintain.
- Since 1 May 2026 IPP 3A requires notice when you collect personal information indirectly.
What HubSpot’s built-in AI does today
HubSpot now sells its AI agents as Agent Hub: Customer, Prospecting, Data, Content, Campaign, Nurture and Revenue agents plus an Agent Builder, included in Starter, Professional and Enterprise editions, with “many AI features and agents” running on HubSpot Credits (HubSpot AI, checked 19 October 2026). The Breeze name survives only for Breeze Assistant, which comes with every subscription. The Data agent is the one relevant here: it answers custom questions about a contact or company using calls, emails, documents and other signals already in your CRM.
The gap is outside evidence. The Data agent works from what HubSpot already holds. If you want research from company websites, the NZ Companies Register or news, with a link behind every claim, you are building an integration.
Integration routes and when to use each
| Route | What it gives you | Use it when |
|---|---|---|
| Agent Hub (Data and Prospecting agents) | Answers and outreach from CRM signals, billed in credits | Questions your CRM data can already answer |
| CRM API | Companies, contacts, properties and notes under your own code | A research pipeline with its own schema, tests and sources |
| Webhooks | Events when records are created or change | Starting research when a company is added |
| Remote MCP server (mcp.hubspot.com) | Read and write tools for any MCP client, OAuth 2.1 with PKCE | Reps working in Claude, ChatGPT, Copilot or Gemini |
| n8n, Make or Zapier | HubSpot triggers plus AI and web steps | Small volumes and simple enrichment |
The MCP server reads CRM records, marketing content and activity history, and can create or update contacts, companies, deals, tickets, line items, products and activities. HubSpot says “all actions respect your existing HubSpot user permissions” (changelog, 13 April 2026). See MCP for business systems for when to put your own server in front.
Automation 1: account research with a source for every claim
- Trigger
- A company is created or reaches your target lifecycle stage.
- Inputs
- Company name and domain; the company’s website, its NZ Companies Register entry and recent news.
- Model step
- Fill a fixed schema (what they do, locations, size signals, recent events, likely priorities). Every statement must carry a source URL, a short supporting quote and the retrieval date.
- Checks in code
- Drop any statement without a URL; confirm each URL loads and the quote appears on the page; confirm the domain matches the record.
- Write-back
- Separate custom properties and a timeline note. Never overwrite rep-maintained fields.
- Human review
- The account owner reads the brief before using it in outreach and can mark it wrong, which clears the properties.
- Measures
- Statements removed by checks, briefs marked wrong, time saved per account.
| Research output | HubSpot property | Rule |
|---|---|---|
| Summary, under 120 words | ai_research_summary (multi-line text) | Replaced on refresh; old version kept in the note |
| Sources | ai_research_sources (one URL and date per line) | Required; empty means nothing is written |
| Industry suggestion | ai_industry_suggestion (dropdown) | Values must match your internal options exactly |
| Last checked | ai_research_checked_at (date) | Drives a 90-day refresh |
Our lesson on researching a customer account covers what good research looks like, and our write-up on AI agents for go-to-market research covers where agents fail at it.
Automation 2: duplicate companies, merged only on approval
- Find. Weekly, list companies that share a domain or have near-identical names.
- Compare. The model proposes which record to keep and why, citing the fields that differ.
- Approve. A HubSpot admin approves or rejects each pair in a review queue.
- Merge and log. Only approved pairs are merged through the API, with both record IDs logged. Treat merges as irreversible.
Permissions, privacy and approval design
Run the pipeline through an app with only company read and write scopes, not a super admin’s login. For the MCP server, connect with a user whose HubSpot permissions match what the assistant should touch. If your account has HubSpot’s sensitive data setting turned on, the MCP server blocks calls, emails, meetings, notes and tasks; the standard APIs do not.
Since 1 May 2026, IPP 3A requires reasonable steps to tell people when you collect their personal information from another source, covering the purpose, recipients, who holds it and their access and correction rights (OPC guidance). Company facts are not personal information; a person’s role history is. Keep automated enrichment at company level unless you have planned the notice. Our agent permissions and approvals guide covers the action tiers.
Costs and limits
HubSpot Credits cost US$0.010 each pay-as-you-go, or US$10 per 1,000-credit monthly pack. Main-hub editions include 500 (Starter), 3,000 (Professional) or 5,000 (Enterprise) a month, which do not roll over. The Data agent uses 10 credits per prompt per record and the Prospecting agent 100 per lead (HubSpot rate sheet, checked 19 October 2026). Marketplace apps get 110 requests per 10 seconds per account; privately distributed apps on Professional get 190 per 10 seconds and 625,000 a day (limits).
Illustrative example: a custom pipeline reading about 30,000 tokens of web pages and writing 1,500 tokens per company costs around US$75 in Claude Sonnet 5.5 fees for 1,000 companies (US$2 and US$10 per million tokens, 19 October 2026), plus any search API.
Where it goes wrong
- Wrong company. “Harbour Builders” exists in three cities. Match on domain and company number, not name.
- Citations that do not support the claim. A real URL can sit beside an invented fact. Checking the quote on the page catches most of these.
- Stale research. Show the checked date on every brief and refresh on a schedule.
- Dropdown mismatches. The API rejects values that are not exact internal option names.
- Reauthorisation. After HubSpot adds MCP scopes, connections can show REQUIRES_REAUTHORIZATION until reconnected.
- Manual recovery. HubSpot keeps property history. Because AI writes only its own properties, a bad run is undone by clearing those properties for records with a given checked date.
Comparing CRMs? See AI and Salesforce and AI and Pipedrive. We build these pipelines as part of our AI integration services.
Sources
- Remote HubSpot MCP server is now generally available, HubSpot Developer Changelog, 13 April 2026, corrected 15 April 2026
- HubSpot MCP server, HubSpot Developers, last modified 19 September 2026
- HubSpot AI and Agent Hub, HubSpot, accessed 19 October 2026
- HubSpot credits rate sheet (product and services catalogue), HubSpot Legal, accessed 19 October 2026
- API usage guidelines and limits, HubSpot Developers, accessed 19 October 2026
- Enhancing the Privacy Act, Ministry of Justice, updated 1 May 2026
- IPP 3A guidance, Office of the Privacy Commissioner, accessed 19 October 2026
Written by the Levy Street team in Wellington. Last updated 19 October 2026.