How an AI Support Agent Resolves a Ticket, Step by Step

The recipe: one ticket, start to finish

Here is the path of one customer question through an AI support agent, built from the vendors' own descriptions. None of the vendors publishes how many minutes each step takes, so the card says so rather than guessing.

Step 1 to 4 sources: Intercom's description of Fin over chat and email, its knowledge sources and its models;³ step 3's token billing is Crisp's;¹² step 5 and 6 follow Intercom's and Zendesk's billing rules.¹⁴ Intercom's "under an hour" setup claim is for Fin on an existing helpdesk.²

Before the first ticket: what the AI learns from

An AI support agent answers from the material you connect, not from general knowledge about your company. Intercom says Fin "can learn from a variety of public and private knowledge sources, including Help Center articles, internal support content, PDFs, and webpages", and that you can target content by a customer's plan, location or brand. You also set a tone of voice and write Guidance: instructions on how to answer and which policies to follow.³ Tidio describes the same start: "Feed Lyro with your support content", after which Lyro "answers customer questions using only your content".¹⁰ Crisp notes that very long or broad knowledge sources cost more to run, because more content may need processing for each reply.¹²

Steps 2 and 3: find, then write

Most of these agents use retrieval-augmented generation (RAG): first search your content for the passages that match the question, then have a language model write the answer from those passages. Intercom lists Fin's sequence as: understand the context and clarify the question, search for the right information, apply your guidance and policies, then generate the response. It names the models behind it: a retrieval model to find potential answers, a reranker to put the best content first, a summary model for the customer's issue, a model that detects when to escalate to humans, and a model to understand the customer's reply.³

On Crisp, this is the step that drives cost. Hugo's credits are token-based, so message length, conversation depth, how much knowledge is read, the language and the model chosen all change what one conversation costs. Crisp's own estimate is $0.05 to $0.10 per full Hugo conversation, which it calls "only an estimate".¹²

Steps 4 and 5: when the ticket comes to you

The handoff is the step that still needs you, and vendors treat it differently on the bill. On Intercom, if a customer asks for a person or shows frustration, or a workspace escalation rule applies, Fin escalates and you aren't billed. If Fin hands off because a Procedure you configured says to, that is a "Procedure handoff" and costs $0.99, like a resolution.¹ On Zendesk, a conversation where the AI helped but a human finished the job is an "Assisted escalation", which doesn't draw on your allowance.⁴ On Help Scout, a customer who types that they need more help or presses "I still need help" isn't counted as a resolution.⁷ Tidio counts any conversation in which Lyro replied at least once, and says Lyro "alerts your team instantly" when a customer needs a human.⁹¹⁰

Step 6: when a resolution is counted

Each vendor waits before deciding a ticket is resolved. Zendesk evaluates at the end of the conversation: 72 hours after the first email, 2 hours after the first chat message by default (up to 72 hours on request), or at hang-up on a call. An LLM then reads the conversation; if it confirms the request was resolved, it is a Verified resolution and draws on the allowance; if not, it is a Contained resolution and doesn't.⁴

Intercom counts a resolution when, after Fin's last answer, the customer confirms it helped ("Ok thanks") or leaves without asking for more. A greeting isn't an answer, a clarifying question that goes unanswered isn't billed, and if the customer later comes back to the same conversation for more help, the resolution is deducted, even across billing periods.¹ Help Scout likewise doesn't charge when AI Answers only greets the customer or asks a follow-up question.⁷ Freshdesk's email session is a fixed 72-hour window from the first email.¹¹

When the meter runs out

Each product lets you cap spending, and what happens at the cap decides how much lands back on you. Intercom offers usage reminders and hard limits.¹ Help Scout disables AI Answers until the next cycle once your monthly limit is reached, and the help widget falls back to its other support options.⁷ Zendesk's "Don't allow overage" setting pauses AI agents at the allowance limit and routes more requests to human agents; the default allows overage, billed monthly.⁵ Crisp's Hugo stops answering by default once included credits are used, and new conversations go to human operators until credits reset or pay-as-you-go is turned on.¹² Freshdesk's 500 included sessions are a one-time allowance per account; after that, sessions are bought in packs.¹¹

Vendors also publish resolution rates, such as Tidio's "Our customers average a 75% resolution rate with Lyro AI" and Zendesk's "Resolve up to 80%+".¹⁰⁶ Both are Vendor-stated, each vendor counts a resolution its own way, and neither is an independent measurement, so they aren't compared here.