HubSpot AI integration: Breeze, the MCP connector or a custom build

What are the four ways to do HubSpot AI integration?

The right route depends on two things: who starts the work, and how much custom logic sits between the AI and your records. If a rep asks a question and reads the answer, a chat connector is enough. If a form fill, a new deal or an inbound email should trigger research, scoring and updates without anyone typing a prompt, you are into workflows or a custom build.

Most HubSpot accounts end up using two or three routes at once. Support might run the Breeze customer agent on live chat, managers might query pipeline data through Claude or ChatGPT, and RevOps might run a custom service that enriches and routes every new lead. The table below sets the four side by side.

What do HubSpot Breeze agents cost after the 2026 credit change?

Breeze agents are billed in HubSpot Credits, and since 14 April 2026 the main agents charge per outcome rather than per attempt. HubSpot's Credits Rate Sheet (product and services catalog, 2026) lists 50 credits for each conversation the customer agent resolves, 100 credits for each lead the prospecting agent recommends for outreach, and 10 credits each for a data agent response on one record and for a workflow AI action. Additional credits cost $0.01 each pay-as-you-go, or $10 per capacity pack of 1,000.

HubSpot's knowledge base defines a resolved customer agent conversation as one where the agent answered from a content source or took an action and the visitor did not ask for a human within 72 hours, or where the agent set a lead's qualification status. The customer agent needs a Professional or Enterprise edition, and if credits run out it stops being assigned new conversations. Reply recommendations that a rep reviews and sends in the help desk do not need a live channel.

Because credits scale with volume, run the numbers before switching an agent on. At the rate sheet figures, 2,000 resolved conversations a month is 100,000 credits. Professional includes 3,000 credits a month, so the remaining 97,000 cost about $970 at the pay-as-you-go rate.

What can the HubSpot MCP server and AI connectors do?

HubSpot's developer documentation describes its remote MCP server as a hosted endpoint that lets an agent read, write and update HubSpot records, including contacts, companies, deals and tickets, as a user would. HubSpot's own connectors for Claude, ChatGPT, Microsoft Copilot and Gemini run on this server, and access is granted through OAuth, so the AI works within the permissions of the HubSpot user who connects.

In practice this is the fastest way to give managers and reps an AI that knows the CRM. A sales lead can ask Claude which deals over a set amount have had no activity in three weeks, or ask ChatGPT to summarize every open ticket for one company before a renewal call, and get an answer from live records instead of an export.

The same documentation positions the server for builders too: if you are creating your own agent, it can call the MCP server rather than wrapping each CRM endpoint yourself. That is useful for an internal assistant, but the connector still assumes a user session. It does not watch for new records, apply your scoring rules on a schedule or hold a write until someone approves it.

When do you need a custom HubSpot AI integration?

You need a custom build once the work has to happen without a person starting it, or once the logic is specific to your business. Breeze and the connectors cover common jobs well; they stop fitting when the rules are yours, the volume is high or the data lives partly outside HubSpot.

Workflow AI actions and custom code can bridge part of the gap, and they are worth trying first for short steps. HubSpot's developer documentation sets hard limits on custom code actions: they need Data Hub Professional or Enterprise, must finish within 20 seconds, can use up to 128 MB of memory and can hold no more than 1,000 characters of secrets in total. A model call that reads a long email thread and a company website can exceed that time, which is why heavier steps usually move to an external service that the workflow calls.

• Work is triggered by a record change, form, email or schedule, dozens or hundreds of times a day.
• Decisions follow your rules: territory, ICP fit, product line, contract status or SLA tier.
• The job reads data from outside HubSpot, such as an ERP, a billing system, a support inbox or the web.
• Output writes to fields that drive routing, forecasting or customer messages, so it needs validation and approval.
• You want to choose the model provider (OpenAI, Claude, Gemini or Bedrock) and log every prompt and output.
• You need audit history, cost caps per task and an eval set that reruns after each change.

How is a custom HubSpot AI integration built?

A custom build has five parts: a trigger, a context step, a model step, a rules step and a write-back step. Each one has a HubSpot-specific detail that decides whether the build holds up in production.

Rate limits are the constraint most demos ignore. HubSpot's API usage guidelines (2026) set the limits below for privately distributed apps. The burst limit applies per app, but the daily limit is shared by every private app in the account, so an AI service competes with your existing syncs. Going over returns a 429 error on subsequent calls, and HubSpot recommends batch APIs and caching to cut call volume. The same guidelines note that webhook calls made from workflows do not count toward the API rate limit.

• Trigger: a workflow webhook or a webhook subscription on object create, update or association events, so the service reacts instead of polling.
• Context: batch reads of the contact, company, recent emails, notes and associated deals, plus any outside data, with caching for records that have not changed.
• Model step: a fixed output schema (for example fit score, reason, suggested owner, next step) so results can be validated field by field.
• Rules step: duplicate checks, allowed values for each property, a confidence threshold and a cost cap per task.
• Write-back: batch updates to AI-specific properties, a note with the reasoning, and a task for the owner; anything below the threshold lands in a review view instead.
• Failure handling: queue requests, retry 429 and 5XX responses with backoff, and alert someone when a record fails twice. Inside workflows, HubSpot retries custom code actions that fail on rate limits or 5XX errors for up to three days, so make every write safe to repeat.

Worked example: AI lead triage on HubSpot

Here is a worked example with round numbers for illustration, not a client result. A B2B company gets 1,500 inbound leads a month through HubSpot forms. A sales ops person spends about 5 minutes on each one: checking the company website, reading the form answers, setting lifecycle stage and fit, assigning an owner and writing a short note for the rep. That is about 125 hours a month.

A custom build listens for new form submissions, pulls the contact and company records, reads the company website, and asks the model for a fit score, a one-line reason, the likely product line and a suggested owner based on your territory rules. Leads above the confidence threshold get their AI properties filled, an owner assigned and a task created. Leads below it, or that match an existing open deal, land in a saved view for the ops person to check.

On HubSpot's side this is about six API calls per lead (read contact and company, check for duplicates, update properties, add a note, create a task), roughly 9,000 a month, far inside even the Starter daily limit. If 75 percent of leads pass cleanly and the rest take a minute to review, the ops time drops to a few hours of exception checking. For comparison, running three data agent prompts per lead at 10 credits each would use 45,000 credits a month, about $420 after Professional's included credits at the pay-as-you-go rate, before any routing rules. Neither number is a reason on its own; the choice depends on how custom your rules are. Your own pass rate depends on form quality and how clean your company records are, which is why we test on a sample of real leads before quoting.

How do you keep HubSpot data safe when you add AI?

Safety comes down to scopes, where AI writes, and where your data goes. A private app token with every scope ticked is the most common mistake, because any bug or injected instruction can then touch every object in the account.

Treat incoming text as data, never as instructions. A form field, email body or chat message can contain text written to manipulate the model, so the service should only ever produce the fixed output schema and never call tools based on what a visitor typed. Check, too, whether your model provider plan keeps prompts out of training and how long it retains them. If records include protected health information, every vendor in the chain needs a signed Business Associate Agreement, which model vendors offer only on certain API or enterprise tiers. This is not legal advice.

• One private app per integration, with only the scopes that job needs, and tokens stored in a secrets manager.
• Write AI output to dedicated properties (for example AI fit score, AI summary) before anything overwrites a field reps rely on.
• Keep owner changes, lifecycle stage changes and customer-facing emails behind an approval step or a rule you can audit.
• Log every prompt, output, cost and write, keyed to the HubSpot record ID, so any change can be traced and rolled back.
• Test in a HubSpot sandbox or a test pipeline first, then switch on for one form or one team before the whole account.

How do you choose a partner for HubSpot AI integration?

Choose a partner who starts by asking which route fits, not one who sells a single tool for every job. If Breeze or the MCP connector covers your case, a good partner will say so and save you a build. When a build is right, the difference between vendors shows on the awkward records: a contact with two companies, a lead from a free email domain, a 429 at the end of the quarter.

We scope every HubSpot build after a free review of real records and a sample of the inputs, quote a fixed price, and hand over the code, prompts, private app and model accounts so you own all of it. A person stays in the loop on anything that changes ownership, money or customer messages. Ask these questions of anyone you consider, including us.

• Which route would you use for this job (Breeze, connector, workflow action or custom build), and why not the others?
• What scopes will the private app use, and which properties will the AI be allowed to write?
• What happens to a record the model is unsure about, and where does a person review it?
• How will you test accuracy before go-live, and on how many real records?
• How will the integration share API limits with our existing syncs and handle 429 errors?
• What will it cost to run each month in model usage and HubSpot Credits, and how is that capped?