HubSpot AI Playbook: Where Breeze Starts and What It Costs

HubSpot's AI has a name, Breeze, a currency, HubSpot Credits, and a wall. The wall is the part worth planning around, and it is in the same place as every other wall in this product: at Professional, $90 a seat a month on annual billing, with a mandatory one-time onboarding fee of $1,500.

This playbook covers what the AI layer contains at each tier, what it costs, which workflows are realistic per role, and one significant thing the vendor does not publish.

The Tier Map, Which Matters More Than The Credits

Buyers ask how many credits are included. The more consequential question is which AI exists at their tier at all, because credits cannot unlock a capability the tier does not have.

The line to notice is the one between Starter and Professional. Everything that behaves like an agent, as opposed to a summary or a basic builder, starts at Professional. So a team evaluating HubSpot specifically for its AI is not evaluating a $7 product. It is evaluating a $90 one with a $1,500 fee attached, and the honest entry price should be written that way in the budget.

What The Credits Are, And What We Cannot Tell You

HubSpot includes 500 credits at Starter, 3,000 at Professional and 5,000 at Enterprise. The allowance is per account rather than per seat, which has a consequence most buyers miss: a ten-person team on Professional shares 3,000 credits, so the effective allowance per person is 300, not 3,000.

What the vendor does not publish on the pricing page is the consumption rate per action. Without that rate, no allowance can be converted into a quantity of useful work, and any article that tells you 3,000 credits is enough for a ten-person team is guessing.

The practical consequence shows up in the review corpus. One reviewer describes following HubSpot's own setup guidance into an action that began importing roughly 2,000 emails, stopping after 250 with the month's credits exhausted and a thirty-day wait. We cannot verify an individual account and we are not presenting it as vendor behaviour. We are presenting it as the shape of the risk that an unpublished consumption rate creates, and as the reason the question below belongs in your evaluation.

Ask for the credit cost of one named workflow at your volume. For example: summarise 200 calls a month, or enrich 2,000 contacts once. Then ask what happens when the allowance is exhausted mid-month: does the feature stop, does it bill, or does it queue.

Provenance And Terms: The Gap In This Playbook

We are going to be explicit about an absence rather than fill it.

Our sources for this cluster are HubSpot's Sales Hub pricing page, its startup and nonprofit programme pages, its real-estate industry page, its partner-tier page, and four of its documentation articles. None of them states which models power Breeze, whose they are, whether customer data is used for training, whether an opt-out exists, or what the retention period is.

Those are the four questions that decide whether an AI feature is usable by a company with data obligations, and we do not have vendor answers to them from the pages we read. So this playbook does not score HubSpot on AI terms, and it does not infer them.

If you handle regulated data, treat those four questions as prerequisites rather than details, and get them answered from the Data Processing Agreement and the subprocessor list rather than from a sales conversation. The review corpus contains a cluster of data-handling complaints, 13% of the sample, mostly about unsolicited marketing and deletion requests rather than about AI training, which is a different issue but a reason to read the data terms rather than skim them.

What The Vendor Claims, Labelled As Such

HubSpot's startup programme page makes specific performance claims for the Customer Agent: that it "resolves up to 90% of conversations automatically and delivers 85% faster ticket resolution"^[1]. The same page frames the agents as more than assistance, stating that "Unlike traditional AI features that simply assist, Breeze AI Agents function as autonomous team members that work across your entire customer journey"^[2].

Those are vendor claims. We quote them because they are what the vendor is committing to in public, which makes them useful as a benchmark you can hold them to, and because the numbers are specific enough to test in your own deployment. They are not measurements we hold.

The counter-signal in our data is thin but not absent: 5% of the reviews we tagged raise AI quality specifically, and the most detailed of them, a three-star review, describes the assistant insisting on processing large tasks in batches, so that work that could finish in minutes before the end of the day instead occupied most of the next one. One customer's experience, and a concrete enough complaint to put to a sales engineer during a trial.

Workflows By Role, Within The Published Capabilities

These are built only from capabilities the vendor lists, with the tier each needs.

For a sales development representative, Professional. Prospecting Agent for research and outreach drafting, with sequences capped at 1,000 sends per user per day from Professional rather than 500. The pairing that matters is Prospecting Agent plus lead form routing, also Professional, so that inbound and outbound share one queue.

For an account executive, Professional. Notetaker plus meeting prep and recap, so the record is written without manual logging, and Deal Summaries, which are available from Starter. Conversation intelligence turns calls into searchable material rather than recordings nobody revisits.

For a sales manager, Professional. Conversation intelligence plus call transcription and coaching, at 750 hours a month, which is the capability most likely to repay the tier on its own: coaching from transcripts rather than from ride-alongs. Coaching Playlists, also Professional, are where that becomes repeatable.

For an operations or revenue operations person, Professional or Enterprise. Data Agent against the Smart CRM, with custom reports at 100 on Professional and 500 on Enterprise, and ten million versus a hundred million events per query. Agent Builder with event triggers at Professional is where AI stops being a button someone presses and becomes part of a workflow.

For a support lead, Professional. Customer Agent across chat and the other channels, with conversational bots gaining functionality at this tier. At Enterprise, bots additionally accept code snippets for custom actions in the Messenger integration.

For anyone on Starter. Deal Summaries and a basic Agent Builder. That is the honest list. Build one small automation with the builder and treat the rest of the AI story as something you have not bought.

Where The AI Layer Is Genuinely Well Placed

Two observations in HubSpot's favour, both structural rather than promotional.

The AI sits on the same database as everything else. Because contacts, companies, deals and tickets are objects in one Smart CRM with associations between them, an agent acting on a deal has the company, the contacts and the activity history available without integration work. Products that bolt AI onto a CRM through an integration do not start from there.

And the metering is at least visible. A published allowance with an unpublished rate is not transparency, but it is better than the common alternative in this category, which is an AI capability with no published figure of any kind. In our wider reading of this field, one competitor publishes a usable per-session rate and most publish nothing you can multiply.

Where It Is Not

The per-account credit pool scales badly with headcount. The allowance does not rise with seats, so AI capacity per person falls as the team grows, which is the opposite of how a per-seat product usually behaves. A twenty-person Professional team has 150 credits per person.

Two of the more interesting capabilities are marked Beta by the vendor, HubSpot Work and Custom Signals. Beta is an honest label and it belongs in a plan as a risk, not as a feature.

And the agents are gated at the same tier as the reporting, the forecasting and the automation, so you cannot buy the AI layer cheaply by accepting a thinner CRM. HubSpot's ladder does not offer that trade.

Conversational Bots Are Tiered Differently From The Agents

The bots are the oldest part of HubSpot's automation story and they are gated on a different pattern from the agents, which is worth knowing if customer-facing chat is your use case.

Conversational bots exist on every tier, including free, with what the vendor describes as limited functionality on Free and Starter and additional functionality from Professional. The Facebook Messenger integration follows the same shape and goes one step further: basic at Starter, branching logic and advanced reporting at Professional, and code snippets for custom actions at Enterprise.

So a free account can put a bot on a website today. What it cannot do is branch that bot on logic, report on it properly, or extend it with custom actions. The progression is unusually gradual by HubSpot's standards, where most capabilities are simply absent and then present.

Live chat itself is on every tier, with HubSpot branding removed from Starter upward, which matters more for a customer-facing bot than for internal tooling.

The Sequence Allowances That Pair With The Prospecting Agent

An AI that drafts outreach is only as useful as the sending capacity behind it, and that capacity is published.

Sequences are capped at 5,000 per account on every tier, which is not the constraint. Sends are: 500 per user per day on Free and Starter, rising to 1,000 per user per day from Professional.

Pair that with the Prospecting Agent, which starts at Professional, and the two limits line up deliberately. The tier that gives you an agent to draft outreach also doubles what you can send. For a four-person team that is 4,000 sends a day, far more than any sensible outbound operation should use, so the binding constraint on an AI-assisted outbound motion at Professional is credits rather than sends.

That is a useful conclusion because it tells you which number to negotiate. Nobody needs more sends. The 3,000 shared credits are where an outbound team will feel the ceiling.

What The AI Layer Adds At Enterprise

Enterprise is $150 a seat a month against Professional's $90, and on the AI side the difference is mostly volume rather than new agents.

Call transcription and coaching doubles, from 750 to 1,500 hours a month. Credits rise from 3,000 to 5,000. Events per query go from ten million to a hundred million, which matters for any agent working across a large data set. HubSpot Work moves from limited to full automation, still marked Beta. And the Messenger integration gains code snippets for custom actions.

The named agents, Data Agent, Prospecting Agent and Customer Agent, are the same ones you get at Professional. So if the AI layer is the reason for the purchase, Enterprise is a capacity decision rather than a capability decision, and the question to answer first is whether 3,000 credits and 750 transcription hours are actually your constraint.

A Two-Week Plan For Testing The Claims

The vendor's claims are specific, which makes them testable. If you have a trial or a first month, this sequence produces your own numbers rather than theirs.

Week one, measure the baseline. Count how many calls a manager actually reviews in a week, how long a rep spends logging activity after a meeting, and how many support conversations arrive. Write the three numbers down before switching anything on.

Week one, instrument the credits. Run one agent on one workflow at a known volume and record the consumption. This is the number the vendor does not publish, and after a week you will have it for your own usage pattern, which is the only version that matters.

Week two, repeat the baseline measurements. Same three counts, same conditions.

The comparison answers the only question worth asking, which is whether the AI replaced work you were already paying for. It also gives you a consumption rate to extrapolate, and a figure to put against the claim of 90% automatic resolution if support is your use case.

Do not try to test everything. One agent, one workflow, two weeks, three numbers.

Four Questions To Settle Before You Commit

1. The rate. How many credits does one named workflow of ours consume at our volume, and what happens at exhaustion?
2. The models. Which models power Breeze, whose are they, and where is that stated in writing?
3. The data. Is customer data used to train them, is there an opt-out, and what is the retention period?
4. The Beta items. What is the support and availability commitment for HubSpot Work and Custom Signals while they carry that label?

The first is a budgeting question. The second and third are the ones we could not answer from public pages, which is why they are here rather than in the article above.