HubSpot’s Breeze agents now live in Agent Hub, which replaced Breeze Studio in July 2026. Setting them up well is mostly data work: lock down AI access, write a data contract for every property an agent reads, dedupe to under 3%, test on a golden set, then launch one agent in review mode with a credit cap.
Naming first, because it changed twice in a year. Then each agent’s real job, its limits, the data it needs and what it costs. Names and rates are HubSpot’s as of October 2026.
| Agent | What it does well | What it doesn’t do | Data prerequisites |
|---|---|---|---|
| Customer AgentService · any Pro/Ent hub | Answers support questions on HubSpot-connected chat and email from the knowledge you give it, and hands off on rules you set. | Can’t answer from sources you haven’t loaded, can’t see Intercom or Zendesk conversations, and resolution lags 72 hours in reporting. | One question per KB article, no contradicting pages, pricing as CSV, handoff triggers defined. 50 credits / resolved conversation |
| Prospecting AgentSales · all paid tiers | Researches contacts and accounts, recommends leads and drafts personalized outreach inside plays you configure. | Doesn’t know your ICP unless it’s in properties and Context; can email the wrong persona if review is off. | Deduped companies, contacts associated to companies, ICP tier and persona as picklists, brand kit loaded. 100 credits / lead recommendation |
| Data AgentData · smart properties | Fills interpretive fields (use case, “sells to healthcare?”, ICP fit reason) from records, websites, web research or call transcripts. | Not a firmographic enrichment vendor; bills per answer even when it returns nothing. | Prompts reference properties by token; run on a 50-record sample first; scope the list, not the whole database. 10 credits / answer per record |
| Deal ProgressionSales or Service Pro/Ent + seat | Suggests next steps and deal updates from activity and conversations. | Can’t fix stage definitions that sales doesn’t follow; garbage stages produce garbage suggestions. | Documented stage exit criteria, required fields per stage, logged calls and emails. |
| Custom agentsAgent Builder · Pro/Ent | Multi-step, portal-specific jobs with clear inputs, including agentic workflows that run without a record. | Not a shortcut past pre-built agents; read-only until you enable CRM writes; cost varies per run by tokens. | A written input/output spec, a dedicated agent-owner user, per-agent monthly credit cap. Dynamic, token-based |
Sources: HubSpot product and credit documentation as summarized by MarTech (“HubSpot July 2026 updates,” Aug 2026) and Supered (“HubSpot Breeze AI in 2026,” Oct 2026); Customer and Prospecting Agent moved to outcome-based billing on April 14, 2026. Included credits: Starter 500, Professional 3,000, Enterprise 5,000, Data Hub Enterprise 10,000 per month; only the highest allotment applies. Plan availability changes during beta; confirm in your portal.
Everyone asking “who’s actually implemented Agent Hub and how’s it going?” gets the same answer: it works where the data model is clean and breaks everywhere else. These are the six failure points we see.
A blank Industry, a free-text “Company Size,” three meanings of MQL. Agents don’t flag ambiguity — they resolve it, sometimes by inventing the answer.
“It can still misunderstand a field, hallucinate a logical connection.” — HubSpot Community, Agent Hub launch threadDuplicate companies and contacts with no company (every gmail.com signup) split account history. The Prospecting Agent researches the stale record; the Customer Agent can’t see the account.
“Closed-won contacts show up as Direct Traffic.” — Exit Five, same weekThe Customer Agent “can only answer from the material you give it.” Two pricing pages that disagree become a “Knowledge conflict” and a handoff.
Coaching opportunities tag these as Knowledge gap vs. Knowledge conflict.Generative AI, CRM data and conversation data toggles ship on. Agent Builder agents default to “everyone can run.” One write-enabled custom agent can overwrite lifecycle stage.
Settings › Account Management › AI › AccessOne shared pool, no rollover. Data Agent bills per answer even when it fills nothing. Trigger-based agents keep running until the pool is empty — then every AI feature pauses.
Pro includes 3,000 credits = 30 Prospecting Agent leads.No golden test set, no accuracy sample, no cost per outcome. So when leadership asks how it’s going, the only answer is “it’s running.”
Gartner: 40%+ of agentic AI projects will be canceled by end of 2027.Measure the inputs before the agent and the outcome after it. If the data rows are in the weak band, no prompt will save the output rows.
| Metric | Weak | Acceptable | Strong |
|---|---|---|---|
| Customer Agent resolution rateNo qualifying human handoff within 72h | < 40% | 50–65% | 70%+ |
| Duplicate rate, contacts & companiesAgents reason over records, not people | > 7% | 3–7% | < 3% |
| Fill rate on properties an agent readsPer agent, per input property | < 70% | 70–85% | > 85% |
| Contact re-verification cadenceData decays 25–30% a year without enrichment | Never | Annually | Quarterly or rolling |
| Prospecting Agent launch volumeNew contacts per play per day | Uncapped | 5–10 for 2–3 weeks | Scaled on meetings booked |
| Golden-set accuracy before go-live30–50 real questions or records | Not tested | ≥ 80% | ≥ 90%, re-run on every change |
Sources: resolution rates — HubSpot CEO on the Q2 2026 call (72%), reported by CX Today, Aug 2026; partner case study, 59% across 2,732 conversations (Hypha, 2026). Duplicate and fill-rate thresholds — ZoomInfo data-quality KPI framework (<3% duplicates and >85% completeness acceptable; >7% and <70% red flags), 2026; middle bands are JayOh’s working range. Decay — ZoomInfo, 2026. Launch volume — Vantage Point, 2026. Golden-set gates are JayOh’s implementation standard, not a third-party benchmark.
Five of the nine steps happen before you configure an agent. That ratio is the point: the agent is the last 30% of the work.
Agent Hub now shows every templated, Marketplace and custom agent in one place — including agents retired on July 23, 2026 that still run but get no updates. Two agents touching the same account is how prospects get double-tapped.
More › Agents › Agent Hub › Agents. Log every agent: owner, trigger, objects read, objects written, runs per month.The master generative-AI toggle, CRM data and customer conversation data all default on. Automation-triggered runs use the agent owner’s permissions, so a super-admin-owned agent sees everything.
Settings › Account Management › AI › Access: decide CRM data, conversation data and files data per use case. Turn off model training if legal requires it; Sensitive Data mode opts you out permanently.Users & Teams: grant Data Agent access, Prospecting agent admin (configure plays) vs. assignment (enroll contacts), and Customer agent editor to named owners only.Assign access › “Owners and super admins can edit; everyone can run” at most. Default lets super admins edit and all users run.This is the agent’s input spec. Every property it reads must have one definition, one source of truth and an owner. Anything not on the contract is noise the agent will still try to use.
Settings › Properties › Export all properties. Build a data dictionary: label, internal name, field type, allowed values, source of truth, owner, current fill rate.Agents reason over records. A contact with no company has no account history, no ICP tier and no owner — so the agent fills the gap with a guess or skips the best lead.
Data Management › Data Quality › Manage duplicates for contacts and companies. Merge until the duplicate rate is under 3%.Settings › Objects › Companies: confirm “Create and associate companies with contacts” is on.Below 70% completeness, an agent is guessing on roughly every third record. Above 85%, its output becomes reviewable at a glance.
More › Data Management › Data Agent › Create smart property, prompt with + Insert property token, source = property data, company website, web research or call transcripts.Data Agent: Fill Smart Property workflow action. Budget 10 credits per record per prompt.Customer Agent resolution rate is mostly a content problem. Contradictory sources produce handoffs; missing sources produce knowledge gaps; both show up in coaching.
Service › Customer Agent › Knowledge: one question per KB article. Retire superseded pages instead of leaving both.Analyze › Coaching opportunities weekly and fix every Knowledge conflict first.Autonomy is a setting, not a default you discover later. Every agent that touches customers or writes data needs a human checkpoint and a spend ceiling.
Testing a Customer Agent or an Agent Builder agent consumes no credits. There’s no reason to learn accuracy from customers.
Test › Email or Live Chat › run each question, read Testing Insights (reasoning, triggers, cited sources), thumbs up/down, then Improve response.Test agent with runtime inputs; review the Outputs tab and estimated credits per run.“How’s it going?” needs a number. Cost per outcome — per resolved ticket, per meeting booked, per correct field — is the only metric that survives a renewal conversation.
Analyze › Reports for resolution, handoff and deflection; Source Usage for underperformers. Remember the 72-hour lag on resolution.Performance tab → meetings booked ÷ credits spent.More › Agents › Agent Inbox): review run history and credit cost. Expand volume or remove review only after two consecutive weeks above threshold.Pre-built if the job is support FAQs, outbound research or enrichment. Custom only for a portal-specific, multi-step job with clear inputs — and only after the pre-built one falls short in testing.
If you can write it as if/then, it’s a workflow — deterministic and free per run. Use an agent when the step needs judgment: classify, summarize, research, draft.
Firmographic facts (headcount, revenue, HQ) come from one enrichment source. Interpretive fields (“sells to healthcare?”, use-case category, ICP fit reason) are where Data Agent earns its 10 credits.
The agent only sees channels connected to HubSpot. If support lives in Intercom or Zendesk, consolidate the inbox first or scope the agent to HubSpot channels only.
None of these are about prompts. They’re about scope, inputs, autonomy and money — the four things that decide whether an agent survives its first quarter.
+ Insert property token in every prompt. Never describe a field in prose and hope the agent finds it.Each of these is common, avoidable and expensive. The right-hand column is what it costs you, in credits or in trust.
Uses HubSpot’s published credit rates as of October 2026: 50 per resolved conversation, 100 per recommended lead, 10 per Data Agent answer. 1 credit = $0.01.
One per agent. If you can’t fill a line, that’s the work to do before you turn it on. Paste it into a doc, a ticket or your Asana board.
AGENT READINESS SPEC Agent name: ____________ (e.g. Customer Agent - Billing FAQ) Agent type: Customer | Prospecting | Data | Deal Progression | Custom Business outcome: ____________ (one sentence, one verb) Owner (accountable): ____________ Success KPI: ____________ target: ____ measured: weekly INPUTS (properties the agent reads) - industry Company dropdown source: enrichment target >85% now ___% - numberofemployees Company number source: enrichment target >85% now ___% - lifecyclestage Contact dropdown source: workflow target 100% now ___% - icp_tier Company dropdown source: scoring model target >85% now ___% - ________________ ______ ______ source: ________ target ___ now ___% OUTPUTS (properties / actions the agent writes) - ________________ allowed: Y / N human approval: Y / N rollback: ________ IDENTITY Duplicate rate (contacts / companies): ___% / ___% target <3% Contacts with no associated company: ___% target <5% KNOWLEDGE & CONTEXT Sources loaded: ____________ Known conflicts removed: Y / N Brand kit + ICPs in Agent Hub Context: Y / N ACCESS & GUARDRAILS AI settings reviewed (CRM / conversations / files): Y / N Who can edit / run: ____________ Autonomy mode: Review | Approval on tool calls | Autonomous Volume cap: ____ per day Monthly credit limit: ____ credits alerts at 75 / 85 / 90% Handoff triggers: ____________ TESTING Golden set size: ____ (30-50 real cases) Golden set accuracy: ___% gate >=80% to launch Estimated credits / run: ____ GO / NO-GO [ ] Inputs above fill-rate target [ ] Duplicates under 3% [ ] Golden set passed [ ] Caps and alerts set [ ] Weekly review slot booked (owner + RevOps) Decision: ____ Date: ____ Expand autonomy after: 2 clean weeks
Straight answers, using HubSpot’s current product names and rates as of October 2026.
A consultant who owns the outcome and developers who execute in one-week sprints. Read-only first, sign-off before anything touches production.
Agent inventory, AI access review, data dictionary for every input, dedupe under 3%, association fixes, picklist conversion and a fill-rate baseline.
Knowledge curation, guardrails, credit caps, a 30–50 case golden set built from your real tickets or records, then a narrow launch with a human checkpoint.
Weekly outcome review, coaching-queue fixes, cost per outcome on a dashboard, and autonomy expanded only on two clean weeks.
We engineer growth systems. JayOh builds the RevOps architecture, GTM execution and data infrastructure that turn messy revenue operations into predictable pipeline for B2B SaaS and PE-backed teams.
