Let the busywork answer itself

Breeze is HubSpot umbrella name for its AI layer, which includes Copilot for assistance inside the interface and agents that do defined jobs such as answering customer questions or researching prospects. Some capabilities are included in existing tiers and some require credits or add-ons. Part of scoping is telling you what your current subscription already covers, because a fair amount of what brands ask us to build is already in the plan they pay for.

Grounding and refusal. The agent is restricted to your knowledge base, policies, and product data as its source, so it answers from your content rather than general knowledge. Then we define prohibited topics where it must escalate instead of attempting a reply, particularly anything involving a specific order total, a refund decision, or a policy exception. An agent that says it will get a human is working correctly, not failing.

Yes, and we set it up that way deliberately. Disclosure is a legal requirement in some jurisdictions, it is becoming standard practice, and hiding it backfires the moment the agent gets something wrong. In practice customers accept AI handling routine questions when the handoff to a human is fast and obvious. What they resent is a bot that loops them without an exit, which is what the escalation rules exist to prevent.

It depends on how concentrated your repeat questions are, and we would rather measure it than promise a percentage. Brands where order status, shipping times, and returns dominate the queue see the most deflection because those answers are factual and groundable. Sizing, compatibility, and anything requiring judgement deflect much less. The audit looks at your actual ticket topics and tells you the realistic ceiling before you commit.

It gets caught by the review loop, logged, and fed back into either the grounding content or the escalation rules. During rollout the agent drafts and a human approves, so errors are caught before the customer sees them. Once autonomy increases, we sample a percentage of conversations for quality. The failure mode to avoid is nobody owning review, because then bad answers go out for months and only surface as complaints.

Within limits, and this is where we are deliberately conservative. Reading data, drafting replies, updating properties, and creating tasks are low risk. Writing back into Shopify or an ERP, issuing refunds, or changing an order are high risk and we generally build those as a proposed action a human confirms rather than something the agent executes alone. The cost of a wrong write into a system of record is much higher than the time it saves.