Am I charged when the AI hands off? It depends who asked for the human

The finding first. Two of the largest vendors that bill per resolution, Intercom and HubSpot, only stop the meter when the customer asks for a person. If one of your own agents reads the AI's answer, sees it is wrong and steps in before the customer asks, the conversation can still be billed as resolved by the AI.

HubSpot's help centre says so in as many words. For its Customer Agent, priced at $0.50 per resolved conversation since 14 April 2026, a conversation counts as resolved once the agent has answered from a content source or performed an action, unless there is a qualifying hand-off. Only hand-offs the visitor starts qualify: asking for a person, or triggering a hand-off phrase or guideline you configured. Assigning the conversation to a person, by hand or through a workflow, does not by itself stop it being counted as resolved. The status is set 72 hours after the customer's last message, and after that, messages from a live agent or negative feedback do not change it. Intercom's rule for Fin is the same in substance, and we covered it in an earlier piece.

HubSpot's own pages describe the rule three ways, and only the last one is precise.

Credits page

The Customer Agent uses 50 credits, $0.50, "when it resolves a conversation without help from a human".

Pricing FAQ

A resolution means "the conversation was not handed to a human rep for 72 hours", with a link to the help centre for how the agent resolves conversations.

Help centre

"Please note: only visitor-initiated handoffs count as qualifying handoffs for resolution purposes." Assignment by a person or a workflow does not count.

The three agree only if "help from a human" means help the customer asked for. A reader of the credits page would reasonably assume a conversation your agent took over is not billed. The help centre, updated on 25 September 2026, says otherwise. HubSpot's billing terms, its product and services catalogue, price the Customer Agent at 50 credits to "Resolve one conversation" and do not define resolve, so the help centre is the definition the pricing FAQ points to.

Help Scout, Zendesk and Gorgias take the opposite view: any human involvement takes the conversation off the AI bill. So the honest answer to "am I charged when the AI escalates?" is a question back. Who started the hand-off?

Four kinds of hand-off

Vendors' pricing pages talk about "the hand-off" as one event. Their help centres, read closely, distinguish four.

The customer asks

"Can I speak to someone?" Every vendor that bills per resolution treats this as free. It is the easy case, and the one the marketing describes.

The AI gives up

It cannot answer, or it detects frustration, and escalates under its own default behaviour. Usually free on per-resolution plans.

Your rules send it

A topic you told it never to handle, a refund above a limit, a procedure that ends with "pass to the team". Here the vendors split.

Your agent steps in

Nobody asked. A person on your team reads the AI's reply, sees it is wrong or about to go wrong, and takes over. This is the rescue, and it is the case the pricing pages are quietest about.

A fifth outcome looks like a hand-off and is not one: the customer gives up and leaves. That is the assumed resolution problem, and on most per-resolution plans it is billed. It is covered in the pieces linked above, so this one sticks to conversations that reach a person.

What eight vendors charge for each kind

From each vendor's own pages. Intercom, Zendesk, Help Scout, Crisp, Helply and eesel were read on 23 September 2026; HubSpot and Gorgias on 27 September.

A ninth vendor, Freshdesk, is left out of the table because its terms do not address any of the four cases. Its Freddy AI Agent is sold in packs of 100 sessions for $49, and a session is every exchange with a customer in 24 hours on chat or 72 hours on email.

Three patterns come out of it.

Any human, no charge. Help Scout, Zendesk and Gorgias keep every hand-off off the AI bill, including the rescue. Zendesk goes furthest: since May 2026 it has a named tier, assisted escalation, for conversations where the AI collected details or authenticated the customer before a person finished the job, and that tier does not count against your allowance. Gorgias needs a footnote rather than a caveat: every conversation carries a ticket fee on your helpdesk plan, and the automation fee is added only when the AI finishes it without a person. A hand-off costs the ticket fee alone, so "free" here means free of the AI charge.

Only the customer can stop the meter. Intercom and HubSpot bill per resolution too, but a hand-off counts only if the customer starts it, directly or through a hand-off phrase you configured. Intercom adds a second exception in the other direction: a hand-off that ends a procedure you configured is billed as an outcome at $0.99, while the same hand-off written as a workspace rule costs nothing.

Every attempt is billed. Crisp, eesel and Helply bill for the work, so a hand-off costs whatever the AI spent before it. None of them pretends otherwise, and for that they deserve more credit than the per-resolution vendors that describe hand-offs as free and bill some of them.

Why a vendor would bill the rescue

The rule looks mean, and it is worth stating the reason for it properly. If any message from a person cancelled the charge, a team could cancel every charge: let the AI answer, then have someone type "Glad that helped" under each reply. A per-resolution bill with no protection against that is a bill the customer sets. Requiring the customer to ask is a crude but unarguable signal that the AI did not finish the job.

The trouble is what it does to the people using it. On Intercom's forum in 2025, a customer described watching stressed customers work through wrong troubleshooting steps because stepping in first would have been billed, and we quoted that thread in an earlier piece. The rule protects the vendor's bill by teaching your team to wait.

We looked for a HubSpot customer describing the same thing on HubSpot's community forum and found none. That may say more about how few customers have read the help centre than about how many are affected.

There are other ways to stop the gaming, and three vendors in the table use them. Zendesk has a language model read each finished conversation before it counts. Gorgias uses a 72-hour window in which any human involvement cancels the charge. In the same Intercom thread, Peter Moriarty, a Fin customer, suggested the obvious one: watch the rate of human overrides per account and investigate the outliers, rather than billing everyone for the few who might cheat.

Designing hand-off rules for the bill you are on

You cannot change a vendor's definition. You can change where your hand-offs come from, and on most plans that decides the bill.

• On Fin, put hand-offs you want free in workspace escalation rules, and keep them inside a procedure only when the work Fin does before handing over is worth $0.99. Turn on escalation when Fin detects frustration: Intercom's own help centre recommends it, and those escalations are not billed.
• On HubSpot, the lever is your hand-off phrases and guidelines. A hand-off the visitor triggers through one of them qualifies and stops the count; a teammate assigning the conversation to themselves does not. Topics you know it gets wrong belong in a guideline that hands off before it answers.
• On a usage plan such as Crisp, eesel, Helply or Freshdesk, hand off early. Every turn the AI spends before giving up is a turn you pay for, so a quick, clean hand-off is the cheapest outcome after a real resolution.
• On any plan, tell your team in writing that stepping in is right when the AI is wrong, whatever it costs. Then measure what it costs, so the decision is made with a number rather than a feeling.

How to find your own rescue penalty

An afternoon with your own data.

1. Export a month of conversations the AI took part in where someone on your team also replied.
2. Sort them by who started the hand-off: the customer asked, the AI or a rule sent it, or a teammate stepped in without being asked. Most help desks record which, and the conversation itself usually shows it.
3. Check each group against your bill. Intercom's performance funnel, HubSpot's resolution reporting and Zendesk's allowance dashboard all let you open the conversations behind the count.
4. Count the unasked rescues that were billed and multiply by your unit price. That is your rescue penalty.
5. Then look at the rescues that were not billed because the team waited for the customer to ask. Note how long they waited. That delay is the part of the price your customers pay.

Before you sign with anyone, ask four questions and get the answers in writing. If the customer asks for a person, is the conversation billed? If the AI escalates on its own? If one of my rules sends it? If my agent steps in first? A vendor that answers all four on a public page has told you more than most do.

Where we stand

Varley sells AI support, so it competes with every product above, and this section is the pitch. Everything before it can be checked on the vendors' own pages.

Varley's published definition treats a hand-off as free at any stage, however much work came before it, and counts any conversation a person on your team joins, at any point and for any reason, as not a resolution. That includes the rescue. It leaves us exposed to the gaming described above, and we accept that: we would rather find an account that games it in the data than bill every team that steps in to help a customer.

Work Varley does alongside your team, such as drafting a reply your agent then sends, is priced per action on the published price list, and a draft your team rejects is free.

Right of reply

None of the companies named here was approached before publication. Every claim is taken from their own public pages, read on the dates given in the sources. If anything on this page is wrong or has changed, write to support@dragapp.com and the correction will be published here with a date, including if it is inconvenient for us.

Changes to this piece

27 Sep 2026

First published.

Sources

• Analyze your customer agent's performance, HubSpot Knowledge Base, updated 25 September 2026. Read 27 September 2026.
• Set up and customize the customer agent's handoff process, HubSpot Knowledge Base, updated 26 August 2026. Read 27 September 2026.
• HubSpot Credits, HubSpot. Read 27 September 2026.
• AI customer agent, HubSpot, including the pricing FAQ. Read 27 September 2026.
• HubSpot product and services catalog, HubSpot Legal, the credits rate sheet. Read 27 September 2026.
• HubSpot's Customer Agent and Prospecting Agent: now you pay when the task is complete, HubSpot, updated 13 April 2026. Read 27 September 2026.
• Fin AI Agent outcomes, Intercom help centre. Read 23 September 2026.
• About automated resolution tiers, Zendesk help, edited 24 August 2026. Read 23 September 2026.
• AI Resolutions pricing, Help Scout docs. Read 23 September 2026.
• How you're billed for using Gorgias, Gorgias docs. Read 27 September 2026.
• Gorgias AI Agent pricing, explained, Gorgias, updated 28 May 2026. Read 27 September 2026.
• Hugo pricing and billing explained, Crisp help centre. Read 23 September 2026.
• Purchase and manage Freddy AI add-ons, Freshdesk support, updated 4 August 2026. Read 27 September 2026.
• Helply pricing, including its FAQ. Read 23 September 2026.
• eesel AI pricing, including its FAQ. Read 23 September 2026.
• Fin's flawed assumed resolved and pricing design, Intercom Community, started 22 January 2025, including the reply by Peter Moriarty. Read 23 September 2026.