Skip to content
MakerGain
← Back to project

Human handoffs are free: Fini puts its AI-support bill on “actually resolved”

Edited by MakerGain

Fini's AI support agent handles chat tickets inside a customer's existing helpdesk. Behavior trees and a deterministic executor read plan and account state node by node, then call customer system interfaces to perform actions such as canceling subscriptions. Human handoffs are not billed; by default, only end-to-end resolutions without human takeover count as billable tickets. The team cleared its first product and rebuilt after four weeks embedded with a fintech support team, choosing not to replace existing helpdesks and requiring human approval before publishing knowledge corrections proposed by background AI.

The official documentation illustrates the workflow with canceling a subscription. A customer types “I want to cancel my subscription.” Fini's agent takes it from there: identify intent, read the plan, account status, and billing date, check that the plan permits cancellation, ask the reason, call the company's own cancellation endpoint, then reply with the result and next billing date. The model is not improvising. The workflow is written as a behavior tree, with a deterministic executor walking the nodes in order.[1]

If a check fails—say the plan does not allow cancellation—or the customer explicitly asks for a person, the conversation goes to the human team with the knowledge used and steps taken. A failed call to the customer’s system is retried or escalated. That handoff is not billed.

Fini builds AI support agents that run on top of a customer's existing helpdesk, handle tickets across chat, email, and voice in 130+ languages, and actually take actions in customer systems. Billing defaults to resolved tickets rather than seats: Growth is $0.89 per ticket, including up to 3,000 a month; Scale is $0.69, including up to 10,000; Enterprise is $0.49, with no limit. Escalation to humans is free, reopening within 72 hours is not billed again, and greetings, spam, and abandoned sessions are excluded. There are no per-seat charges or charges for human teams. The company says customers with high ticket volumes or short handling times can request conversation-based pricing and switch at renewal.[2]

Image: Browser screenshot from the official page[2] on 2026-11-18. The three pricing areas were cropped and stacked vertically, retaining the original text and figures.

An AI-support company leaves “how many conversations we intercepted,” an easy marketing number, out of its bill. By default it charges only for tickets resolved end to end without human takeover. That pricing acknowledges that human handoff is part of the product, not its failure. It targets companies with support teams; daily users include support staff and knowledge reviewers.

You need to trace “resolved” before charging only for resolutions

Making handoffs free requires a way to determine what counts as resolved. Fini does not let the model choose the next step. Rules are behavior trees; the LLM works only inside nodes such as Check, Reply, and Form, while a deterministic tree walker decides which node runs first. The company says identical inputs produce the same execution path. Each conversation has an AI Steps trace recording the nodes run, their inputs, and their outputs in sequence. If an answer is wrong, the team can locate the failed step rather than hearing only “the AI decided to do that.”Introduction — Fini[1]

After a reply is generated, Guardrails checks prohibited words, confidential attribute values, disallowed links, or custom conditions. A match can trigger a rewrite rather than escalation. If the check itself errors, the company says the original reply may remain unchanged—may, not a default pass-through. Before launch, Test Suite replays real conversations to test draft Behaviors, rules, and articles. External Actions use recorded or mock responses during testing, without calling the customer's live systems.[1]

The knowledge side involves more division of labor. Help-center pages, files, Notion, Google Drive, Confluence, and Zendesk Help Center first enter Sources as candidates, then go through review and publication before becoming Articles. The company says it neither chunks documents nor uses embeddings, instead having the LLM reason over a knowledge graph. Background AI proposes knowledge gaps, source conflicts, and draft corrections from conversations. Those proposals enter a review queue by default and are not published without approval. Magic Articles, a separate category generated from source content, enter Review or publish directly after generation, depending on configuration.[3]

The automation boundary is drawn between proposing and approving. Someone on the customer's side must be available to approve knowledge. The company has not quantified that work, yet it is precisely what determines whether $0.89 is cheap.

After clearing the first version, the product took a “no moving house” shape

Fini began in June 2022. The Starter Story interview page, dated 2025-07-18, lists an Amsterdam team with 2 founders and 8 employees. In the Starter Story interview, whose page is dated 2025-07-18, cofounder Deepak says nobody used the first version after launch. They cleared it out and started over, spent 4 weeks inside a fintech support team, then rebuilt Fini around an “AI-native source of truth.”How I Built a $2.5M/Year AI Support App — Starter Story[4]

Two aspects of the rebuilt product fit that experience. First, knowledge is more than throwing documents into a model's search: ordinary sources require review before publishing, and changes proposed by background AI need human approval. That fits spending four weeks in a support team, though the interview does not explicitly attribute the rule to those weeks. Second, Fini does not ask customers to move house. It runs on existing Zendesk, Intercom, and Salesforce systems. Tickets, macros, queues, and human teams stay where they are; Fini reads and replies within them. Teams without a third-party helpdesk can use its built-in Email channel.[3]

It is an unflashy choice, but removes the thing customers least want to do: migrate. The cost is that Fini must adapt to someone else's ticket model. What it sells shifts from “replace your support system” to “insert a step into the workflow you already have.”

In that same interview, the first paying customer after the rebuild paid $500 a month and saw 70% resolution within weeks. That is the participant's account, with no definition of resolution in the interview. For an outside reader, it is a starting point, not a conclusion.

The business on the bill, and the work outside it

Fini's customers lean toward regulated sectors such as fintech, banking, insurance, and healthcare. The interview says customers including Atlas, Column Tax, and TrainingPeaks handle 500,000+ support tickets a month through Fini. Deepak also reports automating 80% of support issues, saving 50% in costs, and improving CSAT by 10%. These are company or founder claims, without an audited measurement basis in the source.[4]

Revenue also needs careful scope. The Starter Story title says “$2.5M/Year,” while the same page's sidebar says “$190K revenue/mo.” Neither explains the measurement period or calculation. These cannot simply be read as MRR, much less used to infer profit. Fini has not disclosed inference costs, gross margin, or profitability.

The sales terms put some underperformance risk into payment conditions, but human handling and maintenance costs remain with the customer. Enterprise customers can first run real traffic free for 90 days, without pilot charges; the plan begins after written targets are met. Separately, the 90-Day Money-Back Guarantee says eligible new Enterprise customers pay $0 if 90% resolution is not reached within 90 days, conditional on providing 1,000 real tickets for validation. The official FAQ says onboarding takes 14 days and full autonomy 30, while the Enterprise pilot area says “Live in 30 days.” Both descriptions coexist; the contract matters at signing.[2]

Voice is priced separately per answered call and is available on every plan. The first 10,000 calls each month cost $0.89 each; calls 10,001–50,000 cost $0.59; and 50,000+ cost $0.35. Alternative per-minute rates are $0.22/$0.18/$0.14, including speech recognition, AI inference, speech synthesis, and telephony. The official pricing page lists unit rates, included allowances, billing periods, and overage methods, but does not list Growth and Scale's monthly platform base fees. A customer's minimum monthly spending therefore cannot be calculated from the public page.[2]

Channels have boundaries too. Competitor Carly's comparison page, which says prices were checked on 2026-10-01, lists Fini's channels as chat, email, website widget, and voice, without text messaging. It positions its own Carly Phone as a direct alternative when customers primarily call and text.8 Best Fini AI Alternatives in 2026[5] That is a competitor's account. The Fini official pages read for this article did not list SMS, so whether it can handle texts still lacks confirmation from Fini.

One final point needs the contract to settle it. The official pricing page emphasizes charging only for end-to-end resolution without a human takeover, while the official guide adds a further condition: no customer reply within 72 hours. How that sentence appears on the bill matters more than any demo.[2] · [6]

Making human handoffs free means forgoing charges for unresolved conversations; human handling and maintenance costs stay with the buyer. How much revenue Fini gives up depends on how the contract defines “resolved.” Before signing, that is the word worth watching.

Continue reading

AI support
Teach the bot to admit uncertainty first: a side project bringing in about \$500 a month

Andy and his two cofounders at London's HeyDeacon built Deacon for founders and small teams still answering support themselves, starting with their own use. When it cannot answer, it admits that and asks visitors for an email address. Only after an address is actually provided is the conversation marked unread and the owner emailed; human replies can enter the knowledge base. The team focuses on the workflow around the model. Knowledge needs manual recrawling or publishing-script triggers to update, while changing models does not require rebuilding the product.

AI support
The database crashed in his first week: Zipchat wants every store's support ready out of the box

Zipchat is an AI support agent for ecommerce. When Ruslan took over in January 2024, the team struggled to account for its MRR and customer sources, and the database started crashing in the first week. He first fixed the books, using Mantle to make MRR visible and collecting feedback on replies. His blog reports a 25% MRR increase about two months later, reaching US$20,000 for the first time. The product claims 30-minute setup, but brand rules, helpdesk connections and answer corrections still require store-by-store configuration.

AI commerce
He asked if people would pay before building LeadFlow. He started by making cold email easier to set up

Before building LeadFlow, 21-year-old German university dropout Adrian Berisha asked people in his marketing network and independent founders whether they would pay for a cold email tool that was extremely simple to set up and could get emails into the inbox. After hearing that people would pay, he put campaign configuration, mailbox setup, AI personalization, reply analysis and CRM handoff in one workflow. He used his own product to run cold email campaigns, booked meetings over LinkedIn and helped early customers get set up through a done-with-you service.

AI images
Pykaso turned “make the next image look like the same person” into a business

Pykaso AI fits between defining a virtual character and publishing content. It helps creators make subsequent images look like the same person. The team bootstrapped the company and combined training and generation in one platform from its first version. To promote it, the team pursued organic growth on social media and reached out to creators. Archived pages differ on how Pykaso measured usage and how it charged.