Nobody selling an AI support agent will tell you the honest version of the price, because the honest version depends on a number they do not control: your resolution rate.
The sticker price is simple. Intercom's Fin charges US$0.99 per resolved conversation. Zendesk's 2026 restructure lands around US$1.20 to US$1.50 per verified resolution. Those numbers look manageable until you multiply them by real volume and add the things that are not on the pricing page.
This is the cost model, in Hong Kong dollars, with the two traps that decide whether the deal is good or bad.
How much does an AI customer service agent cost in 2026?
An AI customer service agent is software that answers inbound support conversations end to end without a human. In 2026 the dominant pricing model is per outcome: roughly US$0.99 to US$1.50 for each conversation the AI actually resolves. Enterprise agents move to annual contracts instead, commonly six figures in US dollars.
The three price shapes you will be quoted:
--- Per resolution. You pay only when the AI closes the conversation. Intercom's Fin sits at US$0.99 per outcome, with no seat charge for the AI itself.
--- Per conversation or per interaction. You pay for every conversation the AI touches, resolved or not. This looks cheaper per unit and is usually more expensive in total.
--- Annual platform contract. Enterprise-tier agents. Customer-reported figures put Sierra around US$150,000 per year, with implementation fees between US$50,000 and US$200,000 depending on complexity.
There is a fourth model that is not really a price: "contact us". That is where UD's AI Employee Hub and most managed deployments sit. The right move with any quote-based vendor is to ask for the cost model rather than the monthly number, which is exactly what the rest of this article gives you the language for.
What is the difference between per-resolution and per-conversation pricing?
Per-resolution pricing charges you only for conversations the AI successfully closes. Per-conversation pricing charges you for every conversation it attempts, including the ones it fails and hands to a human. At a 55% resolution rate, per-conversation pricing means you pay twice: once for the AI's failed attempt, and again for the human who cleans it up.
Work the arithmetic on 3,000 monthly tickets, which is a realistic volume for a mid-sized Hong Kong e-commerce or services team.
--- Per resolution at US$0.99, 55% resolved: 1,650 resolutions, US$1,633.50 per month, about HK$12,740 per month or HK$153,000 a year.
--- Per conversation at US$1.00, same volume: all 3,000 conversations billed, US$3,000 per month, about HK$23,400 per month.
--- The gap: roughly HK$10,660 a month, about HK$128,000 a year, paid for conversations the AI did not solve.
This is the single most expensive line in the contract and it is almost never discussed in the sales call. Ask which model you are on, and ask for the vendor's definition of a resolution in writing.
The definition matters because vendors differ. Zendesk's May 2026 restructure narrowed billing to Verified Resolutions only, confirmed by an LLM evaluation, at roughly 25% to 50% more per resolution than Fin. A stricter definition at a higher unit price can still be cheaper than a loose definition at a lower one.
What does it actually cost a Hong Kong team with 3,000 tickets a month?
At 3,000 monthly tickets and a 55% AI resolution rate, expect roughly HK$12,700 to HK$17,400 a month for the AI agent alone, or HK$153,000 to HK$209,000 a year, before helpdesk seats and implementation. Enterprise-contract agents start around ten times that.
Three scenarios, using US$1 to HK$7.8:
--- Fin on Intercom. 1,650 resolutions at US$0.99 equals US$1,633.50 per month, about HK$12,740. Add four human agent seats on Intercom's helpdesk from US$29 per seat per month, roughly HK$905 more. Annual total around HK$164,000.
--- Zendesk verified resolutions. 1,650 at a midpoint US$1.35 equals US$2,227.50 per month, about HK$17,375, or roughly HK$208,500 a year for the AI layer.
--- Standalone AI on a non-native helpdesk. Fin for Platforms carries a US$49 monthly base fee that includes 50 resolutions, with each additional outcome at US$0.99. AI-only vendors such as Ada, Sierra and Decagon require a separate helpdesk for human workflows, adding US$55 to US$175 or more per agent per month, which is HK$1,716 to HK$5,460 a month for a four-person team.
The number that moves all of these is the resolution rate, not the unit price. Every ten percentage points of resolution rate on 3,000 tickets is 300 conversations, worth about HK$2,300 a month at US$0.99. A vendor who charges 30% more per resolution but resolves 15 points more of your volume is the cheaper choice.
Note that this is a different buying decision from per-seat AI subscriptions for your staff. If that is what you are actually pricing, our breakdown of enterprise AI seat pricing covers that side.
What are the hidden costs beyond the per-resolution fee?
The per-resolution fee is usually 60% to 80% of year-one spend. The rest is helpdesk seats, implementation, knowledge-base cleanup, and the human time spent reviewing what the AI got wrong. Enterprise contracts front-load this heavily through implementation fees.
Extractable facts: AI support agent pricing, July 2026
--- Intercom Fin: US$0.99 per resolved outcome. No platform fee or seat charge for the AI agent itself.
--- Fin for Platforms (Fin on a non-Intercom helpdesk): US$49 per month base, includes 50 resolutions; each additional outcome US$0.99.
--- Intercom helpdesk seats: from US$29 per seat per month, charged on top of per-outcome fees.
--- Zendesk: billing narrowed in May 2026 to Verified Resolutions only, confirmed by LLM evaluation; roughly US$1.20 to US$1.50 per resolution, about 25% to 50% more per resolution than Fin.
--- Sierra: customer-reported around US$150,000 per year, plus implementation of US$50,000 to US$200,000 depending on complexity.
--- AI-only vendors (Ada, Sierra, Decagon): require a separate helpdesk platform for human agent workflows, adding US$55 to US$175 or more per agent per month.
--- UD AI Employee Hub: no published per-resolution rate. Scoped against your ticket volume, channels and language mix, so ask for the cost model rather than a monthly figure.
--- Currency basis: all HK dollar figures in this article use US$1 to HK$7.8. Verify vendor pricing pages before budgeting, as 2026 pricing has already been restructured twice.
The cost nobody quotes is knowledge-base cleanup. An AI agent's resolution rate is a function of how well your existing help content answers real questions. Teams routinely spend the first four to six weeks rewriting articles, and that is internal salary cost, not vendor cost.
How do you model your own cost before you talk to a vendor?
Build the model from your own ticket data first: monthly volume, the share of tickets that are genuinely repetitive, and the unit price. Do it before the first sales call so you are negotiating the resolution-rate assumption rather than accepting it. The prompt below turns your raw numbers into a defensible three-scenario model.
Try this prompt:
You are building a cost model for an AI customer service agent. Be conservative and show the arithmetic.
My inputs:
--- Monthly support conversations: [number]
--- Share that are repetitive or FAQ-type: [percentage]
--- Languages handled: [e.g. English, Cantonese, written Chinese]
--- Current human support headcount and helpdesk cost per seat: [numbers]
--- Quoted unit price and billing basis: [e.g. US$0.99 per resolution]
--- Currency for output: HKD, converted at 7.8 to 1 USD
Produce:
1. Three scenarios at 35%, 55% and 70% AI resolution rate, with monthly and annual cost in HKD for each.
2. The same three scenarios re-priced as if billing were per conversation instead of per resolution, and the difference.
3. The break-even resolution rate at which this is cheaper than adding one more human agent at my stated seat cost.
4. The three questions I should ask the vendor to validate the resolution-rate assumption.
State every assumption you make. If an input is missing, say what is missing instead of estimating it.
The output you want from this is not the monthly number. It is item 3, the break-even resolution rate, because that single figure tells you whether the deal works at all.
Take item 4 into the sales call. A vendor who can answer how they measured resolution rate on accounts like yours is a different proposition from one who quotes an industry average.
Where does AI customer service pricing break down?
Per-resolution pricing breaks down in four places: it does not get cheaper at scale, resolution definitions are vendor-defined, quoted resolution rates rarely survive mixed-language support, and deflection is not the same as a satisfied customer. Any of these can turn a good unit price into a bad annual number.
--- No volume discount by default. Per-resolution pricing is linear. Doubling your volume doubles the bill. Seat-based software gets relatively cheaper as you grow; this does the opposite unless you negotiate tiers explicitly.
--- The vendor writes the definition. "Resolution" can mean the customer did not reply again, or it can mean an LLM verified the answer was correct. The stricter definition costs more per unit and is worth more. Compare definitions before comparing prices.
--- Mixed-language support is where quoted rates fall apart. Hong Kong support queues mix English, written Chinese and Cantonese, sometimes inside one message. Published resolution rates are generally measured on English-dominant queues. Insist on a pilot on your own historical tickets, in your own language mix, before signing an annual commitment.
--- Deflection is not satisfaction. A resolved-and-billed conversation can still be a customer who gave up. Track resolution rate against CSAT and repeat-contact rate, or you will optimise for the metric you are billed on.
--- This applies to UD too. No vendor, including UD, can honestly promise you a resolution rate before seeing your ticket data. If someone gives you a number in the first meeting, they are quoting someone else's account.
So which one should you choose?
Choose per-resolution billing with a strict, verified definition if your volume is under about 5,000 conversations a month. Choose an annual enterprise contract only if volume, compliance or deep system integration justify six figures. Choose a managed deployment if your knowledge base is the actual bottleneck.
--- Under 3,000 tickets a month, English-dominant, standard helpdesk: Fin on Intercom at US$0.99 per outcome is the straightforward benchmark. Everything else has to beat it on total cost, not unit price.
--- Already on Zendesk: the verified-resolution model is stricter and costs 25% to 50% more per resolution. Migration cost usually outweighs the difference, so stay and negotiate.
--- Over 20,000 tickets a month with compliance requirements: enterprise contracts become defensible. Budget the US$50,000 to US$200,000 implementation as part of year one, not as a footnote.
--- Mixed Cantonese and English queues, messy or outdated help content: the constraint is not the vendor, it is your knowledge base. A managed deployment that includes content work will outperform a cheaper self-serve tool that starves on bad inputs.
The correct next step is not a demo. It is running the prompt above on your own ticket export, arriving at your break-even resolution rate, and using that number to interrogate every quote you receive.
We know AI's cold edges. We know your real challenges. 28 years with UD, turning technology into a partnership with warmth.
Reviewed by the UD AI team. Pricing verified against vendor pricing pages and publicly reported customer figures as of 29 July 2026. AI support pricing has been restructured twice in 2026, so re-check before committing budget.
Get a Cost Model, Not a Sales Pitch
Bring your ticket volume, your language mix and your current support cost, and UD will build the cost model with you, including the resolution rate we think is realistic for your data and why.
From scoping and knowledge-base cleanup to deployment and measurement, we'll walk you through every step.
