HubSpot Service Hub AI, Job by Job
HubSpot Service Hub AI clusters into three jobs: drafting for agents, answering for customers, and summarising for everyone. They aren't equally good.
What HubSpot Service Hub AI offers
HubSpot groups its AI under the Breeze name, and the packaging moves around, so treat feature names as less stable than the capabilities behind them.
Some of this is included in your tier and some consumes credits. Get that split in writing before you build a business case, because usage-based AI pricing is where support budgets surprise people.
Where AI genuinely pays off
The wins are less glamorous than the demos and considerably more reliable.
Summarising. A forty message thread becomes three sentences before an escalation. This saves real minutes, it fails safely, and an agent checks it in the same glance they would've spent scrolling anyway.
Drafting the boring reply. Not the sensitive ones. The reply that explains the same policy for the ninth time this week, where the agent edits two lines and sends. Fastest payback of anything on the list.
Turning resolved tickets into articles. Your best knowledge base content is already sitting in closed tickets, written badly and buried. Getting a draft out of it removes the excuse that nobody has time to write.
First-line answers to genuinely documented questions. Password resets, opening hours, where to find an invoice. Narrow, factual, already written down somewhere.
Where it does not
An answering agent is only as good as your published content, which means most disappointing rollouts are a content problem wearing an AI costume. If your knowledge base is nine articles and three of them are out of date, the assistant will be confidently wrong and your customers will learn to skip it within a fortnight.
It also handles account-specific questions poorly. "Why was I charged twice" needs your billing system, not your documentation. Route those to a human early rather than letting the bot try twice first, because a customer who has been failed by a bot arrives at the human already annoyed.
And be careful with sentiment scores. They are interesting in aggregate and a poor basis for any individual decision.
Measuring deflection honestly
Almost every deflection number in this industry is measured wrong. The common version counts conversations the assistant handled without escalation, which measures giving up as a success.
Three honest measurements instead:
Compare CSAT on assisted conversations against human ones as a guardrail. If deflection rises and satisfaction falls by more than a little, you have moved cost to your customers rather than removing it.
A sensible rollout order
Start with the internal features, because they carry no customer risk. Summaries and drafting, for two weeks, with agents free to ignore both.
Then fix content. Take your top ten ticket topics and make sure each has an accurate, current article. This step is the whole project and it's the one people skip.
Only then turn on the customer-facing assistant, scoped to those ten topics, with an obvious route to a human on every reply. Widen it when the numbers earn it.
Frequently asked questions
What does the AI actually do on a ticket?
Three jobs. A HubSpot AI ticket summary condenses a long thread, HubSpot AI chatbot support answers customers from your content, and HubSpot Breeze Service Hub drafting helps the agent write the reply.
Which tiers include HubSpot Service Hub AI?
Assistive features appear broadly, while the customer-facing agent and deeper analysis sit on higher tiers and can consume credits. Confirm the split for your quoted tier.
Does HubSpot AI cost extra?
Some capabilities are included and some are metered. Ask specifically what consumes credits, at what rate, and what happens when you run out mid-month.
Can AI detect duplicate tickets in HubSpot?
Not natively. The assistive features work within one record or one conversation, and comparing an arriving ticket against the whole open queue is a different problem.
What deflection rate is realistic?
It depends almost entirely on how well documented your top topics are. Measure tickets created per thousand sessions before and after rather than trusting a headline figure.
The problem AI drafting can't fix
A perfect reply sent twice to the same person is still twice the work. Ticket Merger catches that on Zendesk and Freshdesk now, with HubSpot coming soon.
Start free trial14-day free trial. No credit card required.