I'm implementing it · 5 min read
Decide before you configure — HubSpot AI in what order, at what cost
You're the admin or the ops lead about to turn HubSpot's AI on for real, and you need the order to configure things in and what each step actually costs, in the operators' own words.
If you have five minutes
- The tempting order is: turn it on, see what it does, fix problems as they surface.
- Every take in this archive argues against that order, for a specific reason each time — permissions, cost, trust, or the fact that a good scoping conversation and a good build are not the same step.
Decide before you configure
The tempting order is: turn it on, see what it does, fix problems as they surface. Every take in this archive argues against that order, for a specific reason each time — permissions, cost, trust, or the fact that a good scoping conversation and a good build are not the same step. So the sequence below runs decision first, configuration second, at every stage.
The decisions, in order, with what it costs to skip them
| Decision | What's actually being decided | What it costs to get it late |
|---|---|---|
| Who already has permission to do what | Breeze Assistant inherits the user's existing permissions rather than granting new ones, so the access review has to happen before you turn on a capability like payment-link creation, not after | Someone gets a capability they shouldn't have, quietly, because their existing permissions were already too broad |
| What the credit budget is | Custom agents stopped being free on Jul 23, 2026; anything that runs an agent at volume — especially inside a workflow — now has a running cost attached | An agent workflow ships, works, and then the bill for running it at scale is a surprise, not a plan |
| Whether an output needs to cite its source before anyone trusts it | "Hard to trust the responses if you're not telling me where they came from" — decide up front which surfaces need citations before you roll them out broadly | People quietly stop using a surface that doesn't show its work, and you don't find out until adoption is already dead |
| Whether you're planning or generating | The agent builder asks good scoping questions and can still hand back something very not right — the planning conversation has to happen separately, before the build | You ship a first draft as if it were the finished thing, and the fix costs more than doing the planning conversation would have |
| Who owns the process gap an agent finds | Action Gap surfaces a missing action; the actual finding underneath it is usually a process gap a human owns, not a bug in the agent | The coaching moment gets treated as an agent problem, gets reported as a bug, and the real process issue never gets fixed |
| What happens to a connector on its way out | Retiring connectors give you a hard date, not a grace period that can slip; migration has to start well before the sunset date | The connection removes itself on schedule whether you migrated or not |
The order things go in
- Permissions first. Audit who can already do what before you enable any Breeze Assistant capability that acts on that permission conversationally — a payment link, an invoice correction, a quote approval. The capability doesn't hand out new power; it removes friction from power someone already has, which means the review has to happen on the permission, not on the feature toggle.
- Credit budget second. Before enabling an agent to run inside a workflow at volume — via Run Agent Workflows Action or a custom agent using webhooks as one of its tools — set a budget the same way you'd budget a metered API. Both webhooks-in-custom-agents and the credit change landed the same week, on purpose, and the design question from that point forward is what each run costs, not just whether it can work.
- The citation and sourcing bar, third. Decide which surfaces need to show where an answer came from before you turn them loose on a wider audience. The Slack rollout only became worth using once it started citing sources — treat "does it cite its source" as a go/no-go gate for any new AI surface, not a nice-to-have that arrives later.
- A separate planning step, fourth. Before anyone uses the agent builder to generate a real agent, have the scoping conversation somewhere else first — a doc, a meeting, a prompt written by hand. One-shotting the build is the documented failure mode; planning first is the documented fix.
- Enablement, fifth. Don't hand someone a finished agent or a finished report and expect them to know what to do with it. Build and share a few starting prompts that open the right conversation in the right place in the product — the deals page and a contact record put Breeze Assistant in different frames of reference, and that matters more than it looks like it should.
- Migration windows, ongoing. Track every sunset date on the record (connectors: Sept 26, 2026; classic theme: Aug 31, 2026, bundled into the new-look Breeze Assistant beta) as a hard deadline, not a soft one — the connection removes itself whether you planned for it or not.
What each step costs, in the operators' own words
"A capability expansion and a pricing change landing in the same week is not a coincidence." Credit: Chris Carolan, Use Webhooks in custom agents — Watch the take.
"Getting agents out of Breeze Studio triggers and into Workflows is the difference between a demo and a system." Credit: Chris Carolan, Run Agent Workflows Action — Watch the take.
"But Breeze can only do what the user already has permission to do, so this does not hand anyone new power, it removes the manual clicking from power they already had." Credit: Casey Hawkins + Chris Carolan, Breeze Assistant Support for Payment Links Creation — Watch the take.
"it asks genuinely good scoping questions, but the thing it built was very not right, with multiple issues." Credit: Casey Hawkins + Chris Carolan, Describe it. Breeze Assistant builds it in Agent Builder — Watch the take.
"As a HubSpot admin or manager, your job needs to be get people comfortable having conversations with Breeze." Credit: Casey Hawkins + Chris Carolan, Prompt Sharing in Breeze Assistant — Watch the take.
"if there's a gap to be found like this, it often relates to there being a real process gap" Credit: Chris Carolan, New Customer Agent Coaching Opportunity: Action Gap — Watch the take.
"Let's not support old architecture that breaks and is hard to maintain in an AI-native world." Credit: Chris Carolan, Some Breeze Assistant Connectors are being removed — Watch the take.
"Hard to trust the responses if you're not telling me where they came from." Credit: Chris Carolan, Breeze in Slack: A better first impression, clearer help, and cleaner citations — Watch the take.
What this pillar doesn't decide for you
Whether any of this pays for itself is a different question, with its own page. This pillar names what HubSpot's AI is, what shipped, and what it costs to run responsibly. The credits question above — what does this cost each time it runs — is the setup half of that argument, not the return half. The return half, including AI-assisted and AI-retained revenue attribution, lives on the AI ROI page and this pillar links across to it rather than re-arguing it here.
Where to next
Where this comes from
- Use Webhooks in custom agents, covered on air Jul 20, 2026. private beta; the credit change and the capability expansion arriving together
- Run Agent Workflows Action, covered on air Jul 20, 2026. live; agents out of Breeze Studio triggers and into Workflows
- Breeze Assistant Support for Payment Links Creation, covered on air Jul 21, 2026. live; why the access review comes before the feature toggle
- Describe it. Breeze Assistant builds it in Agent Builder, covered on air Jul 24, 2026. public beta; the documented failure mode, and the documented fix
- Prompt Sharing in Breeze Assistant, covered on air Aug 20, 2026. public beta; enablement as starting prompts in the right place in the product
- New Customer Agent Coaching Opportunity: Action Gap, covered on air Aug 28, 2026. live; who owns the gap an agent surfaces
- Some Breeze Assistant Connectors are being removed, covered on air Aug 19, 2026. upcoming sunset; the hard date and the reasoning behind it
- Breeze in Slack: A better first impression, clearer help, and cleaner citations, covered on air Sep 3, 2026. live; sourcing as a go/no-go gate rather than a later nicety
What we do not know yet
- Half of what's above is still beta. Use Webhooks in custom agents is private beta. Describe an app, deploy it to your portal is private beta. Describe it. Breeze Assistant builds it in Agent Builder is public beta. Meet Agent Hub itself is public beta. A decision made against a beta feature is a decision made against something that can still change shape before general availability — plan the configuration work accordingly, and re-check status before you commit budget to it.
- No validated cost figure exists in this archive. Every quote above establishes that a cost now exists and that it should be planned for; none of them states an actual credit price, a per-run cost, or a total budget figure. That number has to come from your own portal's live billing, not from this door.
- The connector migration path itself is untested here. The take on connectors being removed names the deadline and the reasoning; it does not walk through what migrating off one of the eight retiring connectors actually involves. If you're running one, the record's own advice is direct: come talk about it, because that's a more useful conversation than a changelog entry.
Mid-build on an agent, or running one of the connectors that is going away?
Come walk through it on the show and we will work it out on air, weekday mornings. The record's own advice on the connectors leaving is exactly that — a conversation is more useful than a changelog entry.