HubSpot's AI agent lineup has grown quickly, and it keeps shifting: an agent that's in beta today might be generally available in a quarter, and the marketplace adds new ones regularly. HubSpot's Breeze Marketplace has more agents than we can usefully cover in one post; these are the ones we get asked about most and have hands-on experience deploying. Here's a practical breakdown of each: what it does, where we've seen it earn its keep, and where it still needs a human.
A HubSpot AI Agent uses natural language processing to interpret an input, whether that's an inbound message, a support ticket, or a CRM record, and takes a specific, predefined action based on it: answering a question, flagging a record, drafting an email, briefing a rep. Some do this through live back-and-forth conversation; others work in the background and only surface once there's something to act on. What they have in common is interpreting language and context rather than running a fixed set of if-this-then-that rules.
The Key Distinction: AI Agents can interpret intent and generate natural-sounding responses, but they operate within boundaries you define. They can answer questions about order status, book meetings based on availability, route conversations to the right team, and surface knowledge base articles. They cannot handle objections, read emotional cues, or make strategic decisions about how to progress a relationship.
People often confuse AI Agents with workflows, chatbots, or full automation platforms. Workflows execute predefined sequences based on triggers. Chatbots follow scripted paths based on button clicks or keywords. AI Agents interpret natural language and decide which action to take based on conversational context. The difference matters when you're deciding which tool to deploy.
Breeze Assistant isn't an agent. It's a general-purpose AI helper that answers questions and drafts things alongside you inside HubSpot, rather than an agent that acts on its own towards a specific outcome. The clearest use is asking it plain-language questions about your CRM data and getting an answer back, instead of exporting data or searching through segments and reports to find what you're after.
RevOdyssey example: one client had just finished a project building out customer health scores and clean revenue reporting, pulling together ticket data, CSAT, invoicing status, and engagement into a single model. Before that work landed, a question like "which accounts are overdue on invoices this month" or "how many open tickets are breaching SLA right now" meant someone stopping to build a filtered view or a one-off report. Once the underlying properties were reliable, the same questions could be asked directly of Breeze Assistant in plain language, with an answer back in seconds rather than a report cycle. The catch, and it's a real one, is that Breeze Assistant was only trustworthy because the health score and reporting groundwork had already been done properly. Point it at inconsistent or half-populated properties and it will answer confidently from bad data. Breeze Assistant isn't an agent. It's a general-purpose AI helper that answers questions and drafts things alongside you inside HubSpot, rather than an agent that acts on its own towards a specific outcome. The clearest use is asking it plain-language questions about your CRM data and getting an answer back, instead of exporting data or searching through segments and reports to find what you're after.
What it doesn't do: spot patterns across your entire dataset unprompted, or tell you what a number means for your strategy. It answers the question you ask; it doesn't yet tell you which question you should have asked.
Our take: this is the quiet workhorse most people underrate because it's not technically an agent. Less flashy than the customer-facing agents below, but often the fastest to show value for an ops-minded team, because it removes friction rather than replacing a whole process.
The best way to evaluate a HubSpot AI agent is by who it talks to and how costly a mistake would be, not by how new or talked-about it is.
Three tiers:
Internal-facing: the agent talks to your own team, not a customer or prospect. Mistakes are cheap: someone catches a bad answer before it goes anywhere.
Customer/prospect-facing, reviewed: the agent drafts or responds, but a human is still in the loop before anything ships.
Buyer-facing, autonomous: the agent acts with less oversight, closer to revenue. Mistakes are expensive and visible.
That's the framework we use below, and it's the one we'd recommend using internally when you're deciding where to start.
The Data Agent answers specific questions about a contact, company, or record by analysing the signals already sitting in your CRM, including ticket content and call notes, and outputs that as something you can act on: a flag, a score, a routing decision.
RevOdyssey example: we set the Data Agent to monitor incoming support tickets for language referencing injury, the type of content a triage team needs to catch immediately rather than find three tickets deep in a queue. The agent reads each ticket, determines whether that kind of language is present, and if so sets the ticket priority to urgent automatically and notifies the team. What used to depend on whoever picked up the ticket noticing the right words now happens on every single ticket, consistently, before a person even looks at it.
What it doesn't do: decide on the right response once a ticket is flagged, that's still a human call, and it can't catch something described in language it hasn't been set up to recognise. It's read-only intelligence, not judgement: it tells you this ticket matters, not what to do about it.
Our take: this is where the Data Agent earns its keep, turning unstructured text into a structured, actionable signal at a volume no team could triage manually by eye. It's a clearer fit for "check every one of these for X" than for "tell me something interesting."
The Company Research Agent pulls together a target company's website, recent news, and existing CRM history so a rep walks into a call already briefed, instead of doing that research themselves beforehand.
What to watch for: it's a briefing tool, not a strategy tool; it tells you what's happening at the account, not what you should do about it.
Our take: a genuinely low-risk one to test, since it's informational rather than customer-facing. Good candidate for an early pilot precisely because a wrong or incomplete brief is far less costly than a wrong customer-facing message.
Why these two are grouped together: both only ever face your own reps. A wrong answer wastes ten minutes; it doesn't damage a customer relationship.
The Customer Agent handles inbound customer queries directly, resolving straightforward questions from your knowledge base and CRM data, and handing off to a human when it hits the edge of what it knows.
RevOdyssey example: a support team of three, handling a mix of portal-submitted tickets and a knowledge base, was heading into a known seasonal spike and needed Level 1 deflection before volume doubled. Rather than switching the Customer Agent on against the whole knowledge base, we started with a historical ticket review to find the top 5-10% of repeat issues, then built the agent's scope around just those, deliberately narrow. The knowledge base itself was restructured so each FAQ answer sat behind its own unique URL, giving the agent a single, dynamic source to pull from instead of a lightly-organised article dump. It went live on a 30/60/90 basis, expanding coverage only once the team trusted the answers it was giving. The result was meaningful ticket deflection heading into peak volume, without the early wobble that comes from an agent let loose on a knowledge base that's too broad for it.
What it doesn't do: it can't handle a frustrated customer with any real empathy, and it can't troubleshoot a genuinely novel problem that isn't already documented somewhere. It's also only as good as what you've given it to work from: a thin knowledge base means confident, incomplete answers, not no answers at all.
Our take: this is the one most businesses should start with, but only after auditing what's actually in your knowledge base. If you can't name fifty questions it could reliably answer today, it's not ready to go live.
The Prospecting Agent researches accounts, watches for buying signals, sources the right contacts, and drafts personalised outreach for a rep to review. It's built to replace the hours a rep spends on research before they ever write the first email.
RevOdyssey example: A client running Sales Hub alongside a data enrichment tool had just launched a new product, which meant a wave of new companies landing in the CRM with no obvious way to prioritise them. Rather than pointing the Prospecting Agent at everything at once, we set it up against buyer intent and lead score together, then split the output into three tiers: top accounts routed to reps for a manual, white-glove approach; mid-tier accounts got an AI-drafted email that a rep reviewed and edited before sending; lowest-tier accounts were flagged as candidates for more autonomous handling later, once the team had built confidence in the drafts. Nothing sent automatically at any tier to start with. The effect was less time lost to research and list-building, without the generic, obviously-automated outreach that tends to come from switching an agent fully loose too early.
What it doesn't do: decide whether an account is actually worth pursuing, read tone or seniority well enough to adjust its approach, or build the trust that turns a reply into a meeting. Every draft still needs a rep's eyes before it goes out; this is a research assistant, not an autonomous SDR.
Our take: strong fit for small sales teams without a dedicated research or SDR function. Weak fit if your lead data is thin or your lifecycle stages are inconsistent: the agent will happily research the wrong accounts with total confidence.
The Customer Health Agent evaluates account health, flags customers showing risk signals, and drafts an outreach email ready for a CS manager to send. In effect, it's trying to do the early-warning work a CS team usually does manually by scanning usage data and support history.
What to watch for: like anything reasoning about "health," it's only as reliable as the signals you feed it: support ticket volume, product usage, contract dates. If those data points are inconsistent across your portal, the health score it produces will be too.
Our take: worth piloting on one account segment before trusting it broadly. Don't let it fully replace a CS manager's own judgement yet.
Why these three are grouped together: all touch a customer or prospect, but nothing goes out the door without a person reviewing it first. That review step is the real safety mechanism.
The Closing Agent sits on a quote and answers a buyer's questions about it directly, using content you've designated as a source: pricing detail, contracts, FAQs, whatever's relevant. When it knows the answer, it answers immediately, day or night, instead of a buyer waiting until your rep is next online. When it doesn't, it hands straight back to the rep rather than guessing.
What to watch for: this is the one with the most on the line if it goes wrong, since it's buyer-facing at the point closest to revenue. "Approved content" needs to actually be tightly scoped and current, or it risks giving a prospect an answer that contradicts what your rep already told them.
Our take: the highest-value agent on this list, and the one with the clearest upside: a buyer getting an instant, accurate answer at 9pm on a Sunday is a real advantage no rep can match. That upside only holds if the knowledge behind it is solid, so we'd sequence it after you've got at least one other agent live and trusted, rather than as your first move.
Why it's alone: it's the only agent here acting on a live deal with minimal human review before the buyer sees it.
HubSpot AI Agents are genuinely useful for specific, repetitive, query-based interactions where resolution paths are clear and well-documented. They're not a silver bullet for customer experience, sales efficiency, or operational drag. They require significant upfront investment, ongoing maintenance, and realistic expectations about what they can and cannot do.
FAQ:
Do I need a specific HubSpot tier to use these agents? Yes, in practice. Breeze Assistant is free on every plan, including the free CRM, but the autonomous agents covered here need a paid Professional or Enterprise subscription, and specific agents sit behind specific Hubs: Customer Agent needs Service Hub, Prospecting Agent needs Sales Hub. Pricing has also moved to an outcome-based model rather than a flat fee, so you pay per resolved conversation, per qualified lead, or per answer, on top of your seat cost. That model has shifted more than once this year, so treat the exact numbers as a starting point and confirm current pricing before budgeting.
Can I use more than one agent at once? Technically yes, but we'd recommend against it early on. Get one agent working reliably, understand what "working" looks like for your business, then add the next. Running several at once before you've validated any of them makes it hard to tell what's actually driving results.
What's the difference between a Breeze Agent and a Breeze Assistant? Assistant is a general-purpose AI helper that works alongside you inside HubSpot, drafting, summarising, answering questions as you work. Agents are built to run whole tasks end-to-end with less oversight, each specialised for a specific job like prospecting or support resolution.
Where do new agents come from, and will this list stay accurate? HubSpot adds and graduates agents through the Breeze Marketplace fairly regularly, and status can change from beta to general availability without much notice. Treat this as a snapshot rather than a permanent list, and check HubSpot's current agent directory before making a final decision.
If you're still building your AI foundations, that's exactly where we help businesses focus first. If you're not sure which agent fits where you are right now, let's talk through it.