How to add AI calling to HubSpot (without replacing your CRM)

What HubSpot already does — and where it stops

Check this before you buy anything. A meaningful number of teams that shop for AI calling already own most of what they need.

The native dialer. HubSpot's documentation states that the number of calling minutes you get depends on your subscription, that they reset on the first of the month, and that you cannot place new calls once the account runs out. Published tier breakdowns put the allowances here:

Minutes are pooled per account, not granted per seat, and you cannot combine hubs to raise the ceiling — additional minutes are a purchase. Figures from HubSpot's published plan documentation and product catalog as of July 2026.

Breeze. HubSpot's AI agents are real and worth knowing before you shop. The Prospecting Agent researches accounts, monitors buying signals, sources contacts through providers like ZoomInfo and Apollo, and drafts personalized outreach. On 14 April 2026 HubSpot moved it and the Customer Agent to outcome-based pricing: $1 per lead recommended for the Prospecting Agent and $0.50 per resolved customer conversation for the Customer Agent. The Customer Agent runs across nine channels, with a Voice/Calling channel in beta.

Where it stops.Two honest limits, and neither is a knock on HubSpot. First, the native dialer is built around a human clicking call from a record. Second, Breeze's voice channel is early and support-oriented — it answers more than it prospects, and the Prospecting Agent's finished output is overwhelmingly drafted email rather than a held phone conversation. If your requirement is a few hundred outbound calls a week in languages your team does not speak, that is a different machine. See AI cold calling software for the category, or multilingual voice AI for the specific capability.

Assist or autonomous: pick the shape first

“AI calling” names two products that install the same way and do opposite things. Deciding late is what turns a rollout into a rip-out.

• AI that assists. Power and parallel dialers, live coaching, automatic call summaries. A human still holds every conversation; the AI removes the dead time around it. This is the right buy when your reps are good and under-utilized.
• AI that calls. An agent places the call and holds the conversation itself. The right buy when the constraint is coverage — more calls, more hours, or more languages than your headcount can reach. It is also the buy that brings the artificial-voice rules in the compliance section into scope.

Veera is the second shape. If you want the first, HubSpot's power dialer or a HubSpot-native tool will serve you better and cost less — the shortlist of AI calling tools for HubSpot agencies covers both shapes side by side with real pricing.

The seven steps

In order. Steps 1 and 2 decide whether you buy at all; steps 3 through 5 are the integration; steps 6 and 7 are what keeps it out of trouble.

1. Step 1

   Audit what your HubSpot plan already does

   Check your calling minutes and Breeze entitlements before you buy anything. HubSpot meters its native dialer per account, not per seat, and the minutes reset on the first of the month: published tier breakdowns put Starter at 500 minutes per month with one phone number, Professional at 3,000 minutes with three, and Enterprise at 12,000 minutes with five. If your team places a low volume of human calls, HubSpot native plus Breeze may already be enough, and the honest answer is to add nothing.

2. Step 2

   Decide the shape: AI that assists, or AI that calls

   These are different products that share a category name. An AI-assist tool speeds up human reps with power dialing, live coaching, and automatic call summaries. An autonomous AI-calling layer places and holds the conversation itself. Buying the wrong shape is the most expensive mistake on this list, because the integration work is identical and only the outcome differs.

3. Step 3

   Connect over OAuth, not CSV and not a Zapier bridge

   Install the tool from the HubSpot App Marketplace or run its OAuth flow, which sends you to app.hubspot.com/oauth/authorize with a client_id, redirect_uri, scope, and state, and returns tokens scoped to one portal. OAuth is what makes the connection revocable, auditable, and per-portal — three properties a CSV export and a hand-wired Zapier bridge both lack. For an agency holding many client portals, this is the difference between an integration and a liability.

4. Step 4

   Map the write-back: name the HubSpot object each call becomes

   Get a specific answer before you dial: does the call become a note, a task, a call engagement, or nothing at all? Tools that render a calling iframe inside HubSpot use the Calling Extensions SDK, where HubSpot creates an engagement on the record timeline and the app fills in the details; contacts and timeline scopes are the minimum for calls to log properly. Tools that call outside HubSpot write back through the CRM API instead. Ask us the same question and the answer is a date, not a yes: Veera calls outside HubSpot and the write-back is designed to land as a note or task engagement through the CRM API, but that push is built and verified against provider fixtures rather than switched on — it is flag-gated off until live sandbox parity is signed off. Plan the pilot around what a vendor can demonstrate in your portal today.

5. Step 5

   Map pipeline stages before the first call, not after

   Reading your deal stages requires the crm.schemas.deals.read scope, and mapping them is the step teams skip. An AI layer that cannot see your pipeline cannot move a deal when a call goes well, so the outcome lands as an orphan note and a human re-keys it — which is the exact manual work you bought the tool to remove. In Veera the reading half is live: it pulls your pipelines and stages and resolves the mapping locally. The writing half — pushing a stage change back into HubSpot — is built and verified but still flag-gated off, so do the mapping now and treat the push as activating.

6. Step 6

   Set consent, quiet hours, and AI disclosure before you dial

   The moment an AI voice places the call, the FCC’s February 2024 ruling puts you squarely under the TCPA’s artificial-voice rules. Confirm three things in the tool, not in a policy document: that you hold prior express consent for the numbers you are dialing, that quiet hours are enforced in the called party’s timezone rather than yours, and that the agent identifies itself. Veera enforces TCPA quiet hours per call and timezone-aware, so a contact in a different timezone is not dialed outside their local window.

7. Step 7

   Pilot on one portal, reconcile, then scale

   Run roughly twenty calls against a single client portal and check every one against the HubSpot timeline: did the contact update, did the outcome land on the right record, did the deal stage move, and did anything write to the wrong pipeline. Only then connect the second portal. A stage mapping that is quietly wrong costs one afternoon to catch in a pilot and weeks of dirty pipeline data to catch later.

How the connection actually works

Worth understanding at one level of detail, because it tells you which vendor claims are plausible.

OAuth, per portal. The flow sends an admin to app.hubspot.com/oauth/authorize with a client_id, redirect_uri, scope, and state; HubSpot returns tokens bound to exactly one portal. The tool resolves which portal it is holding via GET /oauth/v1/access-tokens/{token}, which returns the hub ID and the granted scopes. That portal ID is what scopes every subsequent read and write — the reason an agency can hold twenty client portals without them bleeding into each other.

Two paths for the call itself. A tool that puts a dialpad insideHubSpot uses the Calling Extensions SDK: HubSpot renders the app in a calling iframe, creates an engagement on the record's activity timeline, and exchanges data with the app to fill in the details. Contacts and timeline scopes are the minimum for calls to log properly this way. A tool that places calls outside HubSpot — which is what an autonomous agent does — skips the SDK and writes back through the CRM API instead, where engagements are ordinary v3 CRM objects.

Sync has real limits, and they matter at agency scale. HubSpot's list endpoints cap at 100 records per page and search caps at 200, but the hard one is that search returns at most 10,000 records total regardless of paging. Any tool syncing a large portal has to keep an updated-after watermark moving forward rather than walk history in one pass. Ask a vendor how they handle the 10k cap; the answer separates a real integration from a demo.

One scheduling note if you are building this yourself: HubSpot moved to date-based API versioning with the 2026-03 release, replacing the old v1–v4 scheme, so calling-settings endpoints now look like /crm/extensions/calling/2026-03/{appId}/settings.

What changes the moment an AI places the call

This is the section teams skip and regret. Handing the dial to an AI voice does not create a new legal category — it moves you into an old one that is aggressively litigated.

• AI voices are “artificial” under the TCPA. In Declaratory Ruling FCC 24-17(CG Docket No. 23-362), adopted 2 February 2024 and released 8 February 2024, the FCC unanimously confirmed that AI technologies generating human voices fall inside the Telephone Consumer Protection Act's restrictions on “artificial or prerecorded voice” — 47 U.S.C. § 227. Those calls need the called party's prior express consent, must identify the entity responsible for the call, and must offer an opt-out where the message is telemarketing.
• Quiet hours are the recipient's, not yours. 47 C.F.R. § 64.1200(c)(1) prohibits telephone solicitations to a residential telephone subscriber before 8:00 a.m. or after 9:00 p.m. in the called party'slocal time. (The FCC presumes wireless subscribers who join the do-not-call registry are residential — 2003 TCPA Order, 18 FCC Rcd 14014 — so do not read “residential” as exempting mobile numbers.) The TCPA provides $500 per violation, trebled to $1,500 for willful violations, and quiet-hours class actions have climbed sharply since late 2024. An AI dialing a list across four timezones will find this rule for you — see TCPA quiet hours.
• Whether consent switches quiet hours off is contested — do not bank on it. The defined term does the work here. § 64.1200(f)(15) excludes from the definition any call made to a person with that person's prior express invitation or permission, and the quiet-hours rule reaches only solicitations. In King v. Bon Charge, No. 25-cv-00105-SB (D. Del. 30 Apr 2026), Judge Bibas took that route and dismissed a quiet-hours claim: a plaintiff who knowingly handed over her number to get marketing had given the permission that removes the messages from “solicitation,” and unlike the do-not-call rule, the quiet-hours rule layers on no heightened signed-writing standard. That is one district court, it binds nobody, and courts are expected to diverge — the same opinion let the do-not-call claim proceed. Treat it as a defense that may work, not a green light: enforce quiet hours in the recipient's timezone and let the carve-out be an argument you never need.
• Disclosure is now statutory in the EU. EU AI Act Article 50 applies from 2 August 2026 and requires that people be informed they are interacting with an AI system, in a clear and distinguishable manner, no later than the first interaction.
• No general US rule makes an AI caller say “this is an AI” — and the statute most often cited does not reach calls at all. What federal law requires of an artificial-voice call is identification, not an AI script: 47 C.F.R. § 64.1200(b)(1)–(2) compels the message to state the business responsible at its beginning and to give a callback number. The FCC proposed an AI-disclosure requirement in FCC 24-84(CG Docket No. 23-362, released 8 August 2024) and has not adopted it as of July 2026, so it binds nobody. California's SB 1001 reaches only bots communicating online, and a phone call is not an online communication — it does not govern your calls, whatever a vendor's marketing says. The state duty that does bite is Utah's AI Policy Act (SB 149, effective 1 May 2024, as amended by SB 226 effective 7 May 2025), which requires disclosure whenever a consumer asks. The full map is in AI disclosure laws.
• Email opt-outs are a separate regime. If your sequence touches email, the CAN-SPAM Act requires a working opt-out, and RFC 8058 defines the one-click List-Unsubscribe=One-Click mechanism mailbox providers now expect. Under GDPR Article 17, a contact can require erasure of their personal data.

What Veera enforces in code. TCPA quiet hours are enforced per call and timezone-aware, evaluated against the recipient's local window rather than the sender's. Opt-outs are honored with one-click unsubscribe and suppression applied before the next send. Erasure is honored on request. These are specific, tested behaviors — not a claim that Veera makes your outreach compliant. Consent, targeting, and what the agent says remain yours, and no vendor can take them.

How Veera does it

Concretely, so you can hold it against the steps above.

Veera connects to HubSpot over OAuth and requests six scopes — no more than the job needs: oauth, crm.objects.contacts.read, crm.objects.contacts.write, crm.objects.deals.read, crm.objects.deals.write, and crm.schemas.deals.read. Contacts move in both directions today, and deals are read from HubSpot — the schema scope is what lets Veera read your pipeline stages and map them. Calls are placed outside HubSpot rather than through the Calling Extensions SDK, so the outcome and summary are designed to land on the HubSpot record as a note or task engagement through the CRM API — that write-back is built and verified against provider fixtures, but it is not switched on yet. It ships dark behind a flag that stays off until live sandbox parity is signed off. The same is true of pushing a deal-stage change back into your pipeline, and of activity sync in either direction — one flag gates both halves, so with it off Veera does not read your HubSpot activities any more than it writes to them. We would rather you read that here than discover it in a pilot.

HubSpot remains the system of record throughout. Veera is not a CRM you migrate to and does not ask you to leave HubSpot — it is the calling layer that keeps HubSpot fed. The same adapter pattern backs the GoHighLevel integration, so an agency running both does not learn two mental models.

What runs today, stated plainly:

• Voice calling is live — autonomous outbound in 42 languages, with mid-call instruction to steer a live conversation and supervisor takeover when a human should finish it.
• Two-way contact sync — contacts move both ways. Deals and your pipeline stages are read from HubSpot and mapped; pushing deal changes back into HubSpot, and activity sync in either direction (including call notes), are separate capabilities that are built and activating rather than usable today, described below.
• Lead discovery and scoring — verified-lead discovery and AI lead scoring that ranks contacts hot, warm, or cold with a written reason.
• In-call WhatsApp document delivery — if a prospect asks for the brochure mid-conversation, the agent sends it during the call. See in-call WhatsApp delivery.
• Multi-tenant agency workspaces with per-client isolation and roles, so each client's HubSpot connection and contacts stay in their own workspace.
• Free to start. No seat minimum.

Building, not live yet. The HubSpot write-back this guide spends most of its time on — the call outcome and summary as a note or task engagement, and deal-stage changes pushed back into your pipeline — is built and verified against provider fixtures but ships dark behind a flag that stays off until live sandbox parity is signed off. Two-way contact sync and reads of your deals and pipeline stages are live; the deal push and activity sync in both directions are built and activating, not usable today. SMS, WhatsApp, and email as standalone outreach channels are in the same state — on the roadmap, not available today, and the one live send exception is in-call WhatsApp document delivery during a voice call. Multi-step sequences, the unified inbox, analytics dashboards, and calendar booking are also activating. We list them so you can plan, not so you can count on them this week.

Four ways this goes wrong

• Buying a CRM to get a feature. If adding AI calling means migrating off HubSpot, the calling feature is not what you are buying — and it will not repay the reporting history and workflows you leave behind.
• Never confirming where calls land. “Integrates with HubSpot” is compatible with writing nothing to the timeline. Make the vendor name the object.
• Dialing on the sender's clock. Quiet hours follow the recipient's timezone. A tool that enforces a window in your timezone is not enforcing § 64.1200(c)(1) at all.
• Connecting every client portal on day one. Pilot one, reconcile it against the timeline, then scale. A wrong stage mapping replicated across twenty portals is a data-cleanup project.

If you are weighing this across a client roster rather than one team, AI outreach and CRM for agencies covers the agency-specific requirements, and What is an AI Business Aide? defines the category this tooling sits in.

Frequently asked questions

Related

• AI calling tools for HubSpot agencies — the shortlist, with real pricing for each option
• AI outreach & CRM for agencies — the agency-category hub
• AI cold calling software — the calling-category hub
• What is an AI Business Aide? — category definition
• TCPA quiet hours — the 8 a.m.–9 p.m. rule, in the recipient's timezone
• AI disclosure laws — what an AI caller must actually say, by jurisdiction
• Is AI cold calling legal? — the direct answer, with the statutes
• Multilingual voice AI
• Mid-call instruction
• Supervisor takeover
• Veera vs. Aircall — head-to-head with the deepest HubSpot-native phone system
• State of Outbound Business AI 2026 — full vendor matrix