How Enterprise Teams Are Using AI in HubSpot

Most enterprise organisations now have AI features switched on somewhere in their CRM. Far fewer can point to a measurable outcome from them.

That gap has widened, not narrowed, as HubSpot has moved from assistive AI — tools that help a person work faster — to agentic AI: agents that carry out multi-step tasks on their own.

The capability is real and worth having. What determines whether you get value from it is rarely the technology, and almost always whether your data, permissions and processes were ready to support an agent acting on them.

This article covers what HubSpot’s AI actually does today, from the assistant through to agents and Agent Hub, what has to be true of your CRM before any of it works, where enterprise teams are getting real results, and the governance that an agentic deployment requires.

What AI is actually available in HubSpot today?

As of mid-2026, HubSpot’s AI spans three layers: an assistant you interact with directly, a set of autonomous agents that complete tasks, and Agent Hub (in beta) — the place where those agents are managed.

It is worth understanding each, because they carry very different governance implications.

Breeze Assistant

Breeze Assistant is the conversational companion built into HubSpot.

You ask it to draft an email, summarise a CRM record, answer a question or pull up data, and it responds. It is useful, low-risk and broadly familiar to anyone who has used a chat assistant — the key point for governance is that a person stays in the loop on every action.

Breeze agents

Agents are different in kind, not just degree. Rather than responding to a prompt, an agent runs an entire task — handling a support conversation end to end, or researching an account and drafting outreach — with configurable guardrails and approval steps.

That autonomy is where the leverage is, and where the risk is: an agent takes actions in your CRM, so the governance question shifts from “who can see this data?” to “what is allowed to act on it?”

The pre-built agents most relevant to enterprise teams, described by what they do:

Customer agent. Resolves support conversations across chat and email, drawing on your knowledge base, tickets and connected content, and hands off to a person — with context — when it cannot resolve or when the customer asks.
Prospecting agent. Monitors the CRM for high-intent accounts, researches decision-makers and prepares personalised outreach. Rebuilt in 2026 to handle more of the prospecting lifecycle rather than just drafting.
Data agent. Enriches and maintains CRM records to improve data quality — the agent whose job is, in effect, to keep the others fed with good data.
Deal progression. Surfaces AI recommendations for next steps to move deals through the pipeline, turning activity data into a management signal for sales.
AEO agent. Tracks how your brand appears when buyers ask AI tools such as ChatGPT and Gemini, and recommends content to improve that visibility — a sign of how central AI-answer visibility has become to demand generation.
Content and Social agents. Draft marketing content — landing pages, blogs, social posts — in your brand voice, informed by CRM data, for human review and approval.
Knowledge Base agent. Analyses support tickets and conversations to find gaps in your help content — useful where the same questions keep generating tickets.

Beyond these, an Agent Marketplace offers more specialised agents (for example deal-loss analysis, customer-health scoring and RFP handling), and availability, naming and packaging continue to change as the line-up matures.

Agent Hub

Agent Hub (in beta, formerly “Breeze Agents”) is the single place these agents are turned on, monitored and built. It shows which agents are active across your customer journey, the outcomes they are driving, and what is available but not yet switched on.

It is included for Professional and Enterprise customers; custom agents consume HubSpot Credits on a usage basis, which — as covered later — makes cost a governance item in its own right.

Agent builder and agentic workflows

Two capabilities inside Agent Hub matter most for enterprise teams.

The agent builder lets you create custom agents on a single canvas using your own prompts, knowledge sources and the data already in your CRM, without code, so an agent can be built around a process specific to your business rather than an off-the-shelf task.

And agentic workflows, including a “Run Agent” action (in beta), let you trigger an agent inside a multi-step automation — for example, kicking off account research and outreach automatically when a deal reaches a certain stage.

This is the point at which agents stop being individual tools and become part of your operating process, which is exactly why the governance around them has to be deliberate.

Underpinning all of it is an AI context layer, the business information agents draw on, from company messaging, tone and ICPs down to knowledge vaults holding specific product, policy or support detail.

That context layer is, in effect, HubSpot acknowledging in the product what the next section argues: agents are only as good as what they are given to work from.

Why is my CRM data not ready for AI?

If agent output feels unreliable, the cause is almost always the data it reads rather than the model. HubSpot’s own guidance is blunt on this point: agents automate what is already in your CRM, so data-quality problems should be fixed before activation, not after.

Five conditions cause most of the trouble:

Fragmentation. A customer record split across systems — several CRMs after an acquisition, marketing data divorced from pipeline — means an agent reads a partial record and acts on a confident partial picture.
Duplication. Multiple records for the same company or contact mean summaries, scores and agent decisions are computed on a fraction of the real history.
Inconsistent definitions. If “qualified” or “closed won” means different things across regions, anything an agent infers from those stages inherits the inconsistency and reports it as fact.
Field sprawl and decay. Hundreds of properties, many unused or near-duplicates. An agent weighs whatever is present, including the fields nobody maintains.
Ungoverned unstructured data. Notes, emails and call transcripts are exactly what agents are good at reading, which is also why sensitive information sitting in free-text becomes an exposure the moment an agent can summarise across records.

The uncomfortable implication is that poor data does not merely limit agent value, it manufactures plausible-sounding errors that are harder to catch than an obvious gap, and those errors then propagate into forecasts, routing and customer messages.

We wrote about the underlying principle of why your data needs to work before your AI does.

What has to be true before agents work in your CRM?

A workable readiness baseline, in the order we would sequence it:

One trustworthy customer record. Deduplicated, with a defined system of record for each object and clear rules for what wins in a conflict.
A designed, documented data model. Objects, properties and lifecycle stages that mean the same thing everywhere, with owners. Retire unused fields before an agent starts weighting them.
Curated knowledge sources. Agents draw on knowledge bases and knowledge vaults; a retrieval-grounded agent is only ever as good as the content behind it, so budget ongoing knowledge upkeep, not a one-off load.
Data classification and field-level restriction. Know which fields hold sensitive or regulated data and restrict them before enabling any agent that reads across records.
Permissions that reflect the org. An agent operates within the access it is given. Over-broad permissions become an information-disclosure problem the moment an agent can act on them.
A measurement baseline. Capture current performance — response time, forecast accuracy, resolution time — before deployment, or you will have no way to evidence the effect.

Where are enterprise teams seeing real value?

Where those conditions are met, these are the applications delivering measurable results — each tied to the agent doing the work, and to the precondition that determines whether it works.

Service resolution — Customer agent

Automated triage, response and resolution reduce first-response time and free specialists for genuinely complex cases. This tends to return value fastest, because support data is comparatively clean and the outcome is easy to measure.

The governance requirement is human review on anything carrying contractual or regulatory weight, and a well-maintained knowledge base behind the agent.

Sales productivity and pipeline

The prospecting agent recovers selling time by handling research and first-draft outreach; deal-progression recommendations turn activity data into a next-best-action signal.

The precondition for the pipeline side is strict: consistent stage definitions and reliable close-date discipline across every region, or the recommendations reproduce your data inconsistencies with unearned confidence.

Marketing operations at scale

Drafting, variation and segmentation compress production cycles for teams working across many regions and languages. The discipline is editorial: agents draft, humans approve, and brand and regulatory review stay in the workflow.

Our AI in marketing guide covers the channel-level detail.

Data quality and enrichment

Automated enrichment and de-duplication improve targeting precision and keep the records other agents depend on trustworthy, valuable for ABM in particular.

Verify enrichment accuracy against your own known-good records before trusting it for routing or scoring. See HubSpot Breeze Intelligence for how this works in practice.

AI-answer visibility — AEO agent

As buyers increasingly ask AI tools rather than search engines, visibility in those answers has become a demand-generation concern.

HubSpot’s AEO agent tracks and recommends against it inside the platform; getting genuinely cited, though, is a content and authority problem rather than a toggle, which is the subject of our AEO and AI search work.

What governance do HubSpot agents require?

This is the section most often written after an incident rather than before deployment. Agents act, which raises the stakes above assistive AI.

Seven controls we would treat as non-negotiable in an enterprise or regulated environment:

Approval steps and guardrails are configured on each agent, so autonomy is bounded — an agent proposes or acts within limits you set, rather than running unsupervised.
Human-in-the-loop on consequential output — anything customer-facing, contractual, regulated or feeding a forecast gets a reviewer.
Agent-action auditing — use the audit trail of which records an agent changed and which actions it took, both to investigate incidents and to evidence compliance. Treat this as mandatory, not optional.
Field-level restriction of sensitive data before enabling any agent that reads across records, so an agent cannot surface or act on what a user should not see.
Access review on the assumption that an agent amplifies whatever permissions exist. Over-permissioning that was tolerable manually is not tolerable when an agent can act at scale.
Cost governance — agents run on a usage-based credit model, and default settings can auto-escalate spend during volume spikes. Set credit limits and monitor consumption so cost stays a decision, not a surprise.
Regulatory assessment where personal data is processed — potentially a data protection impact assessment and a documented lawful basis, particularly for anything approaching automated decision-making about individuals.

For the platform-level security detail behind these controls, see HubSpot AI security: what CTOs and CIOs need to know and our HubSpot security and compliance guide.

How should you sequence an enterprise AI rollout in HubSpot?

The pattern that works is narrow, evidenced and governed — in that order.

Assess readiness honestly. Audit data quality, knowledge sources, permissions and classification before selecting agents. This usually reorders the roadmap, because the highest-value agent often sits on the least reliable data.
Remediate the data that the first agent depends on. Not the whole estate — the subset your chosen agent reads from. This keeps the programme moving while fixing what matters.
Pilot one agent with a measured baseline and a contained blast radius. The customer agent is a common first choice: clean-ish data, a clear metric, and a bounded scope.
Put governance in before scaling, not after. Approval steps, auditing, access review and credit limits are far cheaper to establish once than to retrofit across a scaled deployment.
Treat adoption as a change programme. Agents people don’t trust go unused, and trust depends on transparency about limits and a route to flag bad output. Our HubSpot change management guide covers the mechanics.

How do you know whether it is working?

Adoption metrics — how many users touched an agent — tell you little.

Tie evaluation to the operational outcome the agent was chosen for: first-response and resolution time for the customer agent; qualified pipeline and selling time recovered for prospecting; production cycle time for content; data completeness and duplication rates for the data agent.

Compare against the pre-deployment baseline, which is why capturing that baseline is part of readiness rather than an afterthought.

Getting the foundations right

The organisations getting real value from AI in HubSpot are not the ones that switched on the most agents.

They are the ones that fixed their data model, curated their knowledge sources, tightened permissions, piloted one agent against a measured baseline, and governed it: approvals, auditing and cost limits, before scaling.

HubSpot has done the hard part of building capable agents; readiness and governance are where the enterprise work now sits.

We work with enterprise teams on exactly that sequence — data readiness, agent selection, governance design and phased rollout across HubSpot.

Our AI utilisation review assesses where you stand and what to fix first, or you can talk to our team directly.

Frequently asked questions

What is HubSpot Agent Hub?

Agent Hub (in beta, and previously called Breeze Agents) is where HubSpot’s AI agents are turned on, monitored and built.

It shows which agents are active across your customer journey and the outcomes they are driving, lets you activate pre-built agents, and lets you build custom ones. It is included for Professional and Enterprise customers, with custom agents consuming HubSpot Credits.

What is the difference between Breeze Assistant and Breeze agents?

The assistant is interactive — you ask it to draft, summarise or answer, and it responds, with a person in the loop throughout.

Agents are autonomous: they run a whole task, such as resolving a support conversation or researching and contacting a prospect, within guardrails and approval steps you configure. The distinction matters for governance because agents take actions in your CRM.

Which AI agents does HubSpot offer?

As of mid-2026, the pre-built agents include the Customer, Prospecting, Data, Deal progression and AEO agents, plus Content, Social and Knowledge Base agents, with further specialised agents in an Agent Marketplace and the option to build custom agents.

Availability and packaging change frequently, so confirm the current line-up against HubSpot’s documentation.

Do HubSpot AI agents cost extra?

Agent Hub is included for Professional and Enterprise customers, but most agents run on a usage-based HubSpot Credits model, so cost scales with use.

Some agents are priced on outcomes. Because default settings can auto-escalate spend during volume spikes, set credit limits and monitor consumption as a governance control.

How do you prepare CRM data for AI agents?

Deduplicate to one trustworthy record per customer, standardise property and lifecycle definitions across regions, retire unused fields, curate the knowledge sources agents draw on, and classify sensitive data so it can be restricted at the field level.

Agents automate what is already in the CRM, so this remediation should come before activation, not after.

Are HubSpot agents safe to use on sensitive data?

Only with controls in place first. Restrict sensitive fields at the field level, review permissions, keep humans in the loop on consequential actions, use agent-action auditing, and assess regulatory obligations where personal data is processed.

An agent acts within the access it is given, so that access has to be governed before it is switched on.