The short answer: AI customer service is not priced by tokens alone. The budget usually combines human seats, bot licences, message or voice usage, knowledge-base implementation, system integration, and ongoing operations. For a 10-seat, 100,000-conversation department in China, our editorial model lands at RMB 205,500–395,500 in year one. If AI independently resolves 50,000 of those conversations, the full-cost figure is about RMB 4.1–7.9 per AI-resolved conversation. That is a model, not a quote, and it does not promise a replacement ratio or go-live date.
This guide is for teams comparing online support, support bots, and LLM-based customer-service agents. Data cut-off: 25 October 2026. China list prices come from Tencent Qidian and Tencent Cloud Contact Center (TCCC); international comparisons use Intercom and Zendesk public pricing. All figures are list prices before tax, discounts, implementation services, and channel fees. For the knowledge-base layer, see what an enterprise RAG knowledge base really costs.
A seat price is not the total cost. A bot licence often costs more than the seats. A demo without a real question set and transfer rules is not acceptance.
How is AI customer service priced?
Vendors usually mix seats, bot instances, message volume, voice minutes, and resolutions. Ask for every unit before comparing a monthly headline price.
Support software is commonly sold as “people + bot + usage”. Human seats are annual per-user licences. Text or LLM bots are annual per-instance licences. Messages, voice minutes, and outbound calls are metered. International products increasingly add an outcome or automated-resolution fee. A low seat price can hide a much larger bot or usage line.
| Pricing unit | Public example (25 Oct 2026) | Good fit | What buyers miss |
|---|---|---|---|
| Human online seat | Tencent Qidian online seat, Professional: RMB 3,000/user/year; minimum 3 users | Escalations, complaints, sales follow-up | Training, scheduling, QA, attrition, recruitment |
| Text bot | Tencent Qidian text bot: Standard RMB 5,800/instance/year; Professional RMB 28,000/instance/year | Rules-based FAQs and menu-style support | Knowledge-base setup, transfer rules, message allowance |
| LLM support bot | Tencent Qidian LLM support bot, Standard: RMB 60,000/instance/year (10 bots, 5 concurrent, 3m characters, 40,000 messages in year one) | Generative answers, knowledge Q&A, agent assist | Message packs, concurrency, knowledge-base expansion, implementation |
| Message/session pack | Tencent Qidian: 100,000 messages RMB 5,500; 500,000 messages RMB 25,000 | High bot-handled volume | First-year allowance, overage price, whether failed sessions count |
| Voice agent | Tencent Cloud TCCC Standard: RMB 1,999/month for 8,000 minutes and 3 concurrent calls; overage RMB 0.25/minute | Phone support, callback, appointments | Telecom fees, numbers, SIP integration, ASR errors |
| Per resolution | Intercom Fin: USD 0.99 per outcome; Zendesk: USD 1.50 per committed automated resolution, USD 2.00 pay-as-you-go | Outcome-aligned buying | Definition of “resolved”, minimum commitment, channel fees, tax |
China prices are RMB list prices; Intercom and Zendesk are USD, annual billing. No exchange-rate conversion is applied. Product scopes differ, so the lowest unit price is not a total-cost ranking.
Where does the money go?
Most of the budget sits in human cover, bot licences, usage, implementation and integration, and ongoing knowledge operations. Token costs are a small line item.
An AI support agent must connect to web chat, WeChat, apps, phone, or ticketing; read orders, logistics, and after-sales policy; restrict answers to permitted data; and hand off to a human when the request is risky. That is a system and an operating process, not just a model.
- Human seats: complex complaints, technical diagnosis, and commitments on refunds or discounts still need trained people. A seat licence is only part of the employment cost.
- Bot licence: text and LLM bots are usually annual per-instance licences; concurrency, knowledge capacity, and message allowances determine whether peak demand is covered.
- Usage: messages, voice minutes, outbound calls, and channels such as WhatsApp or SMS may be metered separately and priced differently by region.
- Implementation and integration: document cleanup, intent/flow design, and CRM/ERP/ticketing/order APIs often take more time than the subscription itself.
- Operations and QA: knowledge updates, bad-case reviews, transfer rules, compliance sampling, and agent training are recurring costs, not a one-off project fee.
What does a 10-seat, 100,000-conversation service cost?
Using public China list prices and editorial assumptions, a 10-seat department handling 100,000 AI-fronted conversations costs about RMB 205,500–395,500 in year one, before tax and hardware.
Assumptions: 10 online seats; 100,000 conversations enter the AI support layer each year; 30% are transferred to a human; AI independently resolves 50% (50,000). The vendor lines use Tencent Qidian Professional seats and the Standard LLM support bot, plus a 100,000-message pack. Implementation, integration, and operations are editorial estimates, not vendor quotes.
| Cost line | Calculation | Year-one amount (RMB) |
|---|---|---|
| Human online seats | 10 seats × RMB 3,000/seat/year | 30,000 |
| LLM support bot | Standard RMB 60,000/instance/year × 1 | 60,000 |
| Message pack | Additional 100,000 messages | 5,500 |
| Knowledge-base setup | Document cleanup, flow design, launch training | 30,000–80,000 |
| Channel/CRM/order/ticketing integration | APIs, permissions, joint testing | 20,000–100,000 |
| Operations, QA, knowledge updates | 0.5–1 FTE, sampling and bad cases | 60,000–120,000 |
| Year-one total | Full-cost basis | 205,500–395,500 |
Editorial model, not a quote. Excludes tax, telephony, hardware, private deployment, discounts, and channel fees. Actual cost depends on document quality, system count, concurrency, and transfer rate. Dividing the full cost by 50,000 AI-resolved conversations gives RMB 4.1–7.9; the denominator is not total conversation volume, and this is not evidence that AI is cheaper than human-only support.
The important finding is not the headline total. The bot licence and operations dwarf the seat subscription. Ten seats cost RMB 30,000; one Standard LLM bot licence costs RMB 60,000. If the knowledge base and integrations are weak, the apparent seat saving is consumed by rework and transfers.
Why is per-resolution pricing not automatically cheaper?
Per-resolution pricing links the bill to outcomes, but the unit price times resolution volume can exceed a seat subscription. Model resolution volume, transfer rate, and channel scope before comparing totals.
Intercom Fin charges USD 0.99 per outcome. Zendesk AI agents charge per automated resolution: USD 1.50 committed and USD 2.00 pay-as-you-go. Intercom counts an outcome when the customer confirms resolution, does not ask for more help, or Fin completes a workflow. Zendesk counts an automated resolution when AI resolves the request without escalation. The conditions differ; put the definition in the contract.
| Option | Assumptions | Annual fee (USD) | Notes |
|---|---|---|---|
| Intercom Essential + Fin | 10 seats; 10,000 outcomes; annual billing | 3,480 + 9,900 = 13,380 | USD 29/seat/month; Fin USD 0.99/outcome |
| Zendesk Suite Team + AI agents | 10 seats; 10,000 automated resolutions; annual billing | 6,600 + 14,100 = 20,700 | USD 55/agent/month; 600 included; 9,400 at USD 1.50 |
| Zendesk pay-as-you-go | Same, overage at USD 2.00 | 6,600 + 18,800 = 25,400 | No committed volume; higher unit price |
Public pricing accessed 25 Oct 2026; annual billing, before tax and add-on channels. Plan features differ, so the totals are not a direct ranking. 10,000 is the AI resolution volume, not total conversations.
“Pay only for success” still carries risk. Minimum commitments, what counts as resolved, whether a repeat contact is re-billed, whether a handoff is charged, and whether channel fees are separate can all change the effective unit cost. At high volume, outcome fees can dominate the bill; at low volume, seats and minimums can be heavier.
How much human work can AI customer service replace?
There is no universal replacement rate. FAQs and order lookups suit automation; complaints, technical diagnosis, and commitments on refunds or pricing still need people. Test with real conversations.
“The AI can answer” is not the same as “the AI can resolve”. A fluent answer is easy; closing the loop without a repeat contact, a wrong order, or an unauthorised promise is harder. Segment by scenario:
| Scenario | AI fit | Human role | Main risk |
|---|---|---|---|
| FAQs, hours, policy explanations | High | Sampling and knowledge updates | Outdated policy; answer without a source |
| Order, logistics, refund status | High, with system APIs | Exceptions and cross-system checks | Permission, data lag, wrong order |
| After-sales troubleshooting | Medium | Diagnosis confirmation, replacement decision | Wrong diagnosis; missed safety issue |
| Complaints, compensation, price negotiation | Low | Human-led; AI summary only | Unauthorised promise; escalation |
| Phone peaks and appointments | Medium | Callback and exception handling | ASR errors, dialect, background noise |
| Technical proposals and quotes | Low–medium | Sales/engineer confirmation | Parameter hallucination; delivery promise |
Editorial guidance, not an industry benchmark. Replacement rates must be tested on your own historical conversations, product structure, and transfer rules.
How should you accept an AI customer service system?
At minimum, measure independent resolution, answer accuracy, transfer rate, seven-day repeat contacts, CSAT, full cost per resolution, and permission compliance. “Sounds human” is not acceptance.
Define the metric and denominator before the POC. Otherwise a vendor can report answer rate, intent accuracy, or “customer did not reply” while the team still spends heavily on humans.
| Metric | Suggested definition | What to verify |
|---|---|---|
| Independent resolution | Conversations resolved by AI without a human transfer and without a repeat contact on the same issue within 7 days / all conversations | Break down by intent; do not report only an aggregate |
| Answer accuracy | Sample answers consistent with policy, order, and product facts | Double review; keep the transcript and citation |
| Transfer rate | Human transfers / all conversations | Split by channel, time, and intent |
| 7-day repeat contact | Same customer and issue within 7 days / resolved conversations | Detect false resolution |
| CSAT / complaint rate | Report AI and human separately | Do not blend them into one average |
| Cost per resolution | (Software + usage + implementation amortisation + operations) / successful resolutions | State the denominator and full-cost basis |
| Permission and compliance | Unauthorised reads, promises, or personal-data leaks | Test by role; keep audit logs |
| Stale-answer rate | Answers citing expired policy or withdrawn products / sampled answers | Effective dates, versions, incremental sync |
Set thresholds by business context; there is no universal “90% is good”. For refunds, pricing, health, or safety, require human confirmation.
How should you run the POC and structure the contract?
Run a two-to-four-week shadow POC on real channels, then pay by stage. Write the resolution definition, overage price, data ownership, transfer rules, SLA, and exit into the contract.
- Scope: one channel, one product line, or one problem class. Do not connect every channel and SKU at once.
- Shadow mode: AI suggests, a human confirms and sends. Record the suggestion and the edit; do not let it make commitments in the first phase.
- Same question set: test 300–500 real historical conversations against AI and human agents, comparing resolution, accuracy, and handling time.
- Read-only first: connect order and logistics queries first. Refunds, address changes, and coupons come later with human approval.
- Contract: define “resolved”, minimum commitment, overage, channel fees, data/knowledge export, model portability, SLA, handoff rules, and exit.
- Payment: tie stages to knowledge-base acceptance, POC, launch, and stable operations. A demo pass is not acceptance.
Is Lingyi a fit?
Standard support SaaS is faster for pure online support. Consider Lingyi’s AI foundation and custom integration only when private deployment, permission isolation, and multi-system workflows are required.
Lingyi does not sell a standard contact-centre SaaS. If you need web/WeChat chat, ticketing, and agent management, evaluate a specialist product first; it is usually faster and cheaper. If the agent must run inside a private network, connect to OA/CRM/ERP with granular permissions, and use private models, LingHarness can serve as the AI foundation. LingWork covers internal office workflows, not a contact-centre replacement. Channels, ticketing, and agent management still need a separate product or custom integration.
We do not promise a fixed replacement ratio, accuracy, or go-live date. A controlled POC with real conversations is the honest way to measure resolution, transfer rate, and unit cost before expanding. For a cost model or POC scope, talk to us. For the knowledge layer, see RAG knowledge-base costs; for internal office agents, see office AI agent selection.
FAQ
Is AI customer service billed by tokens?
Not only. China support products commonly charge per seat, bot instance, and message volume, with model tokens included in the bot licence. International products often combine seats with a per-outcome fee. Ask how licence, usage, implementation, channels, and tax are priced separately.
What does a 10-seat service cost per year?
Ten Tencent Qidian Professional seats alone are about RMB 30,000/year. Adding the Standard LLM bot at RMB 60,000/year, message packs, knowledge-base setup, integration, and operations produces an editorial first-year range of RMB 205,500–395,500. This is a model, not a vendor quote.
Is per seat, per message, or per resolution cheaper?
There is no universal answer. Low-volume, human-heavy support is simpler per seat. High bot volume is easier to control with message packs or annual licences. Per-resolution pricing fits clearly measurable outcomes but carries minimum commitments. Model the same volume and overage risk before comparing totals.
How much human work can AI replace?
Do not use a vendor’s headline percentage as your plan. FAQs and order lookups have more automation potential; complaints, technical diagnosis, refunds, and pricing commitments still need people. Measure independent resolution, transfer rate, and seven-day repeat contacts before changing headcount.
Is private deployment always more expensive than SaaS?
Not always, but it adds servers, model deployment, backup, security, and operations. It can be justified when data cannot leave the intranet, permissions are complex, or several internal systems must be connected. For standard online support, SaaS or a hybrid is usually faster.
How long does it take, and how do we accept it?
There is no universal timeline and we do not promise one. Scope, knowledge quality, APIs, and transfer rules dominate. Start with a two-to-four-week shadow POC and accept it on 300–500 real conversations; then expand by stage. Measure resolution, accuracy, transfers, repeat contacts, CSAT, cost, and compliance, not just the demo.
Sources and caveats
This is public-source research and an editorial model, not a vendor quote, and it is not based on interviews or first-party testing. RMB and USD are not converted. China prices are list prices; international prices are public annual-billing prices. Tax, discounts, implementation, channel fees, and minimum commitments can change the total. The 30% transfer rate, 50% AI resolution rate, and implementation/operations ranges are assumptions used to show the cost structure. Actual results depend on the knowledge base, APIs, permissions, channels, and operations; this article does not promise savings or a replacement headcount.
- Tencent Cloud, Tencent Qidian customer-service billing (updated 3 Jul 2026): online seat RMB 3,000/user/year; text bot RMB 5,800/28,000 per instance/year; LLM bot RMB 60,000 per instance/year; message and QA packs.
- Tencent Cloud, Cloud Contact Center billing (updated 22 Sep 2026): voice agent RMB 1,999/month for 8,000 minutes and 3 concurrent calls; overage RMB 0.25/minute; human seats and telecom fees.
- Intercom, Pricing (accessed 25 Oct 2026): Fin USD 0.99 per outcome; Essential USD 29/seat/month annual; outcome definition and minimum commitment.
- Zendesk, Pricing (accessed 25 Oct 2026): Suite Team USD 55/agent/month annual; automated resolutions USD 1.50 committed / USD 2.00 pay-as-you-go; 5 included per agent/month.
The Chinese version of this guide is at AI 客服要多少钱.