HubSpot as the memory. Claude as the reasoning. A human as the brake.
Custom AI agents that read HubSpot for context and only write after a human approves. Knowledge agents, follow-up watchers, meeting briefs, deal-stage mappers, conversation summarisers. We run a fleet of these on our own operations.
Our three rules for AI agents in HubSpot
A contractor with admin access and no judgement is not a colleague. It is a liability.
- 01 —
Approval-first by default
The agent proposes. A human authorises in one click. Then the write happens. Autonomy gets earned on a per-agent, per-action basis. It is never the starting position.
- 02 —
Scope-limited tokens
Each agent has a HubSpot private app token scoped to exactly the objects and properties it needs. A knowledge agent does not get write scope. A summariser does not get billing scope. Blast radius is bounded by design.
- 03 —
Every action is auditable
Every prompt, every tool call, every retrieval, every authorisation, every write. Logged with reasoning. If the agent does something surprising, you can trace exactly why. That is what makes approval-first viable.
The agents we run on ourselves
Five agents in production. None of them autonomous by default.
Our internal operations brain is a fleet of small Claude-powered agents reading and writing to HubSpot. Each one earns its trust before it earns its autonomy. The follow-up watcher will probably stay approval-first forever (the cost of a wrongly-sent nudge is real). The meeting-brief generator runs unsupervised (the cost of a brief no-one reads is zero).
Everything we ship for clients, we run for ourselves first. We have a strong opinion that agents which cannot explain themselves should not have access to your CRM.
Agent patterns we have shipped
- Knowledge agent. Answers product, pricing and contract questions against a private Q&A base. Cites the source entry every time. No invented specifics.
- Follow-up watcher. Scans deal conversations daily. Default position is "do not follow up". Only flags the ones that genuinely warrant a nudge, and drafts the nudge for human approval.
- Meeting-brief generator. Pre-call brief assembling deal context, conversation history and similar past deals. Runs autonomously because the failure mode is harmless.
- Deal-stage mapper. Infers correct pipeline stage from signals in the conversation. Proposes the change for approval. Never demotes a deal without explicit override.
- Conversation summariser. Long thread into a three-line summary with a suggested next step. Posted as a HubSpot note for the rep to read.
How it’s built
Anthropic Claude. HubSpot API directly. A small retrieval layer per agent. AWS Lambda for compute.
Claude handles the reasoning. Its tool-use behaviour and long-context properties suit HubSpot agent work well. We hit the HubSpot API directly with scope-limited private app tokens. Retrieval is built specifically for each agent: a knowledge agent reads a Q&A index; a meeting-brief generator reads the deal’s conversation history; a follow-up watcher reads the timeline.
We deliberately avoid frameworks that obscure what the agent is doing. Every prompt is in the repo. Every tool call is logged. Every retrieval is inspectable. The approval UI is simple: a proposal queue, an authorise or reject button, and a record of both. Approving an agent action takes a second because you can see exactly what the agent proposed and what it based the proposal on.
Got an AI agent idea that needs HubSpot context?
Thirty minutes on a call. We will walk through the agents we run on ourselves, talk through whether your idea is a one-week experiment or a six-week build, and tell you honestly when the answer is "buy Breeze and don’t build anything".
Book a discovery callOr back to the Integrations hub · Tom Leyden · tom.leyden@redyellowblue.com.au · +61 413 432 185