01225 591 597hello@spotdev.co.uk
UK Business Hours: 9am - 5pm
AI Transformation16 Sept 2026Read: 17 min read

AI by function, for a company that runs on HubSpot

What HubSpot's own AI already does for sales, service, marketing and operations, what independent studies show, and where a custom build starts.

John Kelleher
John Kelleher
AI by function, for a company that runs on HubSpot

John Kelleher is a Claude Certified Architect (Foundations and Professional) and leads SpotDev, a Claude Registered Partner and OpenAI Select Partner.

You already pay for HubSpot, and HubSpot now ships with AI agents. Before you fund a build for sales, service, marketing or operations, you want two answers: what does the CRM already do, and has anyone independent measured whether AI works in that function at all. Vendor pages answer the first generously and the second not at all.

This post takes each function in turn and gives three things: what the independent evidence shows (including where there is none), what HubSpot provides natively, and the point at which a custom build starts. The function list itself comes from Anthropic's Enterprise AI Transformation Guide (Oct 2025), which names engineering, legal, marketing, finance and customer support as the usual pilot areas and adds sales, revenue operations and HR in its own worked examples. Those examples are Anthropic's internal experience, not published with independent detail, so we do not use them as evidence. The guide's advice to focus on "where AI can deliver meaningful improvement" is sound; this post says where that is, with the basis stated.

Two related posts already own the arguments this one leans on. How to roll out AI department by department covers sequencing and links to the individual workflow spokes. Breeze versus custom Claude agents covers the build-or-buy argument in full. Neither is repeated below.

Start with what HubSpot's own AI already does

HubSpot's AI surface as of 20 Sep 2026 has three parts.

Agent Hub is HubSpot's home for the agents running across the account. It carries five pre-built agents: the AEO Agent (visibility in AI-powered search), the Data Agent (enriching and maintaining CRM data), the Prospecting Agent (researching target companies and drafting outreach), the Deal Progression Agent (recommending next steps on open deals) and the Customer Agent (answering and resolving support conversations). HubSpot's knowledge base does not group them under any category label, so neither do we. Agent Hub was renamed from Breeze Agents earlier in 2026; the current product page describes it as the home for every agent running your go-to-market, without the "formerly" clause it once carried.

Breeze Assistant is the conversational assistant inside the app: it drafts content, summarises records, prepares for meetings and completes tasks on request. Breeze Intelligence is a separate product covering data enrichment and buyer intent.

HubSpot Credits are how most of this is metered. They sit in one account-level pool, sized by the highest subscription tier you hold, and they are consumed by the Customer Agent, the Prospecting Agent, the Data Agent, Data Studio syncs, Breeze actions inside workflows and buyer-intent tracking. The pool resets monthly and unused credits do not roll over. Usage is visible under Account and Billing, then Usage and Limits. Since April 2026 the Customer Agent and the Prospecting Agent have moved to outcome-based pricing: you pay per resolved conversation or per lead recommended for outreach, rather than a flat fee. HubSpot quotes the per-outcome prices in dollars and they move, so check the current HubSpot page rather than a blog.

On tier availability, be careful with what partner blogs tell you. HubSpot's own product page states the Prospecting Agent is available on Starter, Professional and Enterprise; the knowledge base does not state per-agent tier gates for the others. If a proposal tells you an agent "needs Professional", ask for the HubSpot page that says so.

Two further facts matter for a finance director. First, HubSpot publishes its own performance figure for the Customer Agent: a 65% resolution rate and 39% faster resolution across more than 8,000 customers who have activated it. That is HubSpot's own reported figure, not an audited one, and the base is customers who switched it on, not a sample of all accounts. Second, HubSpot publishes no pilot methodology and no AI ROI guidance of its own. It gives you per-agent reporting (for the Customer Agent: resolution rate, deflection rate, human handoff rate, handle rate and visitor feedback, with resolution judged 72 hours after the agent's last response) and an ROI calculator. How to run the pilot, and how to decide whether it worked, is left to you.

Sales

What the evidence shows. There is no peer-reviewed, large-sample field study of generative AI in sales. The closest is a working paper by Fang and colleagues (Oct 2025) covering one online retail platform, where effects across seven AI-assisted workflows ranged from no detectable impact to a 16.3% increase in sales. The adjacent evidence is stronger but comes from another function: a study of 5,179 customer-support agents found an AI assistant raised issues resolved per hour by 14% on average and 34% for novice and low-skilled staff, with minimal effect on experienced agents. The mechanism is the one sales AI relies on, but nobody has measured it against quota or win rate.

What HubSpot gives you. The Prospecting Agent researches accounts and drafts outreach, priced per lead it recommends. The Deal Progression Agent recommends next steps on open deals. Breeze Assistant drafts emails and summarises records before a call, and Breeze Intelligence supplies enrichment and intent signals. For a team whose problem is thin research and slow follow-up, that is a serious baseline.

Where a custom build starts. When the work needs to read systems HubSpot cannot see (your quoting tool, your delivery data, your finance system), when outreach has to follow rules HubSpot's agent cannot be given, or when the approval line has to sit somewhere HubSpot's configuration will not put it. AI for sales sets out where that line sits and why.

Customer success

What the evidence shows. This is the best-evidenced function in the set. The Brynjolfsson, Li and Raymond study, published in the Quarterly Journal of Economics in 2025, is the one to cite: 5,179 agents, a 14% average rise in issues resolved per hour and 34% for the least experienced, with the tool doing little for the experienced. Two consequences follow. A team of long-tenured agents should expect less. And the study measured an assistant helping a human, not an agent replacing one, so it says nothing about a fully automated front line.

What HubSpot gives you. The Customer Agent answers and resolves support conversations from your knowledge sources, priced per resolved conversation. HubSpot's own reported figure is 65% of conversations resolved and 39% faster resolution across the 8,000-plus customers who activated it. Its reporting (resolution, deflection, handoff, handle rate and visitor feedback) is enough to run a before-and-after on your own tickets, which is worth more than the vendor's aggregate.

Where a custom build starts. When the answer depends on data outside HubSpot (an order status, a contract term, a usage figure), when the agent needs to take an action rather than give an answer, or when what it is allowed to say has to be constrained more tightly than a knowledge base allows. AI for customer success covers where HubSpot's own agent stops and what the build looks like beyond it.

Marketing

What the evidence shows. As with sales, no peer-reviewed field study exists for marketing specifically. The nearest independent evidence is on professional writing: a randomised trial of 444 professionals found that access to a chat assistant cut time on writing tasks by 0.8 standard deviations and raised quality by 0.4 standard deviations, with the largest benefit going to weaker writers. That is a fair basis for expecting faster drafting. It says nothing about conversion, pipeline or brand fit, and no independent study does.

What HubSpot gives you. Breeze Assistant generates and rewrites content inside the tools your team already uses. The AEO Agent works on how your brand shows up in AI-powered search. Breeze Intelligence shortens forms and tracks buyer intent. Content generation draws on the same credit pool as everything else.

Where a custom build starts. When the job is explanation rather than production: attribution that says why a channel moved, content briefs built from actual segment behaviour in the CRM, or approval gates that HubSpot's editor will not enforce. AI for marketing sets out what is engineered and what stays with people.

Business operations

What the evidence shows. The most useful study here is also the most cautionary. Dell'Acqua and colleagues ran a pre-registered experiment with 758 consultants: on tasks inside the tool's competence, those using AI completed 12.2% more tasks, 25.1% faster, at over 40% higher quality; on a task chosen to sit outside it, they were 19 percentage points less likely to reach the correct answer. Operations work mixes both kinds of task, and the participants could not reliably tell which was which. Scope every operations use case to a bounded task for that reason. The writing evidence above applies to the drafting and summarising side of admin work too.

What HubSpot gives you. The Data Agent enriches and maintains CRM data. Breeze actions inside workflows let you put a model step into an existing automation, metered from the credit pool. Breeze Assistant handles the summarising and drafting. There is no separate operations agent; at this company size, operations AI and CRM-hygiene AI are usually the same work.

Where a custom build starts. When the process crosses systems: documents arriving by email that need to become records in HubSpot and your finance system, receivables chasing that reads the ledger, or management information that joins CRM data to data HubSpot does not hold. AI for business operations lists the systems we connect and where the human line sits.

Finance, legal, HR and engineering

HubSpot has no product for any of these four. What follows is the evidence only.

Finance has one good field study. Choi and Xie followed 277 accountants across 79 small and mid-sized firms and found AI adopters served 55% more clients each week, moved about 8.5% of their time from data entry to client communication and quality assurance, and cut the monthly close by 7.5 days. It is a working paper, forthcoming in the Journal of Accounting Research, on US accounting practices rather than in-house finance teams, so read it as directional. For a HubSpot-run company, finance AI lives in the accounting platform, fed by a well-built integration from the CRM; HubSpot supplies the deal and invoice data and does not do the analysis.

Legal has nothing in our verified evidence set. We could not find an independent field study of current-generation models on legal work that met our sourcing standard, and the vendor case studies that circulate are volume figures we could not check. In practice, legal AI in a company of this size runs alongside HubSpot: it reads a contract attached to a deal and writes a status back. That is a thin integration and we would rather say so.

HR is the thinnest of all. There is no independent study of AI in HR administration at any scale. HubSpot holds no HR data. If AI appears in HR at this company size, it is a general internal knowledge assistant over your document store, which is not a CRM project.

Engineering has the strongest evidence of any function. Field experiments at Microsoft, Accenture and a Fortune 100 company, 4,867 developers in total, found a 26.08% increase in completed tasks among those using an AI coding tool, with less experienced developers gaining most. An earlier controlled task found the AI-assisted group finished 55.8% faster, though that paper remains a preprint. For your business this shows up not as a HubSpot feature but as faster delivery of the integrations, portals and workflow actions that connect other systems to the CRM.

Function by function: the summary

Function Strongest independent evidence HubSpot native capability Custom build starts when
Sales None peer reviewed; one retail-platform working paper, none to 16.3% Prospecting Agent, Deal Progression Agent, Breeze Assistant, Breeze Intelligence The agent must read non-HubSpot systems or follow rules HubSpot cannot be given
Customer success 5,179 agents: +14% average, +34% for novices Customer Agent with per-agent reporting; HubSpot's own reported 65% resolution Answers depend on external data, or the agent must act, not just answer
Marketing None peer reviewed; writing RCT of 444: time down 0.8 SD, quality up 0.4 SD Breeze Assistant content, AEO Agent, Breeze Intelligence Explanation over production: causal attribution, CRM-informed briefs, enforced approval gates
Business operations 758 consultants: +12.2% tasks, +25.1% speed, but 19 points worse outside the frontier Data Agent, Breeze workflow actions, Breeze Assistant The process crosses systems (email intake, ledger, MI joins)
Finance 277 accountants, 79 firms: +55% clients, close 7.5 days shorter None Immediately, and it lives in the accounting platform, fed from HubSpot
Legal None in our verified set None Runs alongside HubSpot; reads the deal, writes a status
HR None at any scale None Not a CRM project
Engineering 4,867 developers: +26.08% completed tasks None (developer tooling, not a product) Applies to the delivery of your HubSpot work, not to HubSpot itself

Choosing which function goes first

The order is not fixed, but the evidence and the native coverage above narrow it: pick the function where independent evidence exists, HubSpot's own agent gives you a measurable baseline, and the workflow touches your customers least. The full sequencing argument, including what a quarter contains and how to choose the first department, is in How to roll out AI department by department.

Frequently asked questions

Can HubSpot's built-in AI replace a custom AI agent?

Sometimes, yes. If the workflow lives entirely inside HubSpot, the data it needs is already on the record, and the agent only has to answer or recommend rather than act, the Customer Agent, Prospecting Agent or Deal Progression Agent may be all you need, and the outcome-based pricing makes the trial cheap. A custom build earns its place when the agent must read other systems, take actions, or follow rules HubSpot's configuration cannot express.

What AI use cases exist for each department in a HubSpot-run business?

Sales: account research, outreach drafting, next-step recommendations, pre-call summaries. Customer success: first-line resolution from your knowledge sources, handoff with context, ticket summaries. Marketing: content drafting, AI-search visibility, form enrichment and intent. Operations: data hygiene and enrichment, model steps inside workflows, summarising. HubSpot covers a native slice of each; the department-by-department post links the individual workflow spokes for the rest.

Which department should adopt AI first?

Choose on three tests: independent evidence exists for the type of work, HubSpot's own agent gives you a measurable baseline, and a failure would not reach a customer. On those tests customer success and internal operations usually come before sales and marketing, where no peer-reviewed field study exists. The sequencing detail is in the department-by-department post rather than repeated here.

What does HubSpot's AI not do that a custom build covers?

It does not read systems outside HubSpot, so anything depending on your finance ledger, delivery platform or quoting tool is out of scope for the pre-built agents. It does not take actions beyond its defined remit, and it does not let you place the approval line wherever your risk appetite requires. It also publishes no pilot methodology, so the measurement design is yours either way.

How much does it cost to add AI to an existing HubSpot setup?

HubSpot's native agents are metered through HubSpot Credits from one account-level pool, with the Customer and Prospecting Agents priced per resolved conversation or recommended lead. For a custom build the cost is driven by data access, write permissions, integration depth, review load and seat count, not by the model. We do not publish a range; a diagnostic produces a scoped number for your account.

Is HubSpot's own Customer Agent resolution figure reliable?

It is HubSpot's own reported figure across more than 8,000 customers who activated the Customer Agent, and it is not independently audited. Treat it as an upper reference point, not a forecast. The independent evidence for AI in support work is a 14% average productivity gain for assisted human agents, with the benefit concentrated in newer staff. Your own resolution and handoff reports after a defined trial period are the number that matters.

The next step

If you run on HubSpot, the first job is to establish what the native agents already cover for the function you have in mind, and to set up the before-and-after on your own data before switching anything on. The second is to find out whether the workflow you need crosses into a build. A diagnostic does both: it inventories your HubSpot AI coverage, identifies the workflows that need more, and returns a scoped plan. The readiness assessment on our AI implementation page covers the data and access questions that decide the answer. For the HubSpot-specific picture, see HubSpot AI.

Sources

John Kelleher

John Kelleher

Author
John is the founder and the Chief Executive at SpotDev.
wwwin

Stay Updated with Our Latest Insights

Get expert HubSpot tips and integration strategies delivered to your inbox.

Related Articles

Continue exploring our insights and expertise