What an AI agent can do inside HubSpot

Five ways an agent plugs into HubSpot

Which one you use decides what the agent can see, how fast it reacts and who it acts as. Most production agents combine the first three.

What it can read, write and react to

Every CRM object is open to an agent through the API, as long as the private app has the scope for it. This is the map we start every HubSpot agent from.

The API calls behind every agent

Whatever the agent's job, it makes the same kinds of call in the same order: find, read, write, connect, and sometimes act beyond the record. These are the endpoints that matter, with the limits that shape the design.

One real call: create a follow-up task

The task, its owner and its link to the deal go in a single request. hs_timestamp is the due date and the only required property; 216 is HubSpot's default task-to-deal association.

POST /crm/v3/objects/tasks
{
  "properties": {
    "hs_timestamp": "2026-11-04T09:00:00Z",
    "hs_task_subject": "Send 3-site pricing to Brightlane",
    "hs_task_body": "Dana asked for 3-site pricing on 22 Oct. No pricing email sent since.",
    "hs_task_priority": "HIGH",
    "hs_task_type": "EMAIL",
    "hubspot_owner_id": "81234567"
  },
  "associations": [{
    "to": { "id": "18342207551" },
    "types": [{
      "associationCategory": "HUBSPOT_DEFINED",
      "associationTypeId": 216
    }]
  }]
}

Scopes to ask for

Paths shown are the v3/v4 forms most integrations call. HubSpot also publishes date-versioned paths (for example /crm/objects/2026-09/deals/search). Check the developer docs for your version.

Three agents, call by call

How the calls above become a working agent. The first two are patterns we run on live client pipelines today; the outputs on the right show what lands in HubSpot or the inbox.

24 agents we can run in your HubSpot

Each one is a trigger, a judgement and a write back to HubSpot. We set it up, run it and tune it; your team gets the result. Pick the team, then the leak.

Breeze, or an agent we run for you?

HubSpot now ships its own agents under Agent Hub: Customer Agent, Prospecting Agent, Data Agent, Content Agent, Campaign Agent, Nurture Agent, Revenue Agent, Smart Deal Progression, Agent Builder. Use them where they cover the job. Have us run one where the job crosses tools, follows your own rules or needs someone watching every run.

Breeze names, tiers and credit prices change often. Check HubSpot's plan pages for your portal.

11 mistakes we learned on live pipelines

None of these are in the API docs as warnings. Each one cost us a wrong task, a missed deal or a confused rep before we fixed it.

1. 01

   No email scope, no email context

   Without sales-email-read the agent can't read one-to-one emails, so a deal with a reply yesterday looks untouched for a month. Check the scope before trusting any “no contact” flag.

2. 02

   Not every logged email is contact

   Some reminder and tracking tools log fake “incoming” emails days after a send. Count outgoing emails and held meetings as contact, and nothing else.

3. 03

   Search lags behind writes

   New and updated records take a few moments to reach search. Confirm a write by reading the record by ID, not by searching for it.

4. 04

   Search stops at 10,000

   One query returns at most 10,000 results and paging further fails. Slice big portals by date window or pipeline.

5. 05

   Deals owned by people who left

   Deactivated owners keep their deals. Read owners with archived=true, and give every portal a fallback owner, or tasks land with nobody.

6. 06

   Duplicate tasks

   Reps, meeting recorders and other tools create tasks too. Compare every new task with the deal's open ones before writing, or the rep gets three copies.

7. 07

   Meeting notes go stale

   The proposal promised in a meeting is often sent the next morning. Check what happened after the meeting before creating a task for it.

8. 08

   AI due dates aren't promises

   A due date the model picked is a suggestion, not something the client agreed to. Never write it into a note as an agreed date.

9. 09

   Stage IDs aren't labels

   Custom pipelines use numeric stage IDs. Map them through the pipelines API, or the agent reasons about “1094523” instead of “Proposal sent”.

10. 10

    Rate limits are shared

    The daily cap is shared by every app on the portal. Use batch endpoints and back off on a 429 so the agent never starves your other integrations.

11. 11

    A token is not permission

    Having API access doesn't mean the business approved writes. Keep a read-only, draft or auto flag per portal and change it only on the owner's say-so.

How much to let it do

Trust is earned per agent and per portal. Every agent we run sits at one of three levels, and moves up only on evidence.