Technology Insights

AI Customer Service Cost: Requirements Review, PoC and Production Launch

Based on custom projects in Taiwan in 2026, the AI customer service requirements review is free, a single-use-case PoC costs about NT$300,000 to 800,000, and a production launch about NT$800,000 to 2.5 million, while cross-department, multi-system projects are quoted by scope; AI model API and cloud hosting are billed separately by usage. This article breaks down where the money goes at each stage, compares the cost structure of platform subscriptions and custom integration, and covers how to keep the budget under control.

By Founder & CEO, JoinXPublished 8 min read
AI Customer Service Cost: Requirements Review, PoC and Production Launch

Key takeaways

  • Custom AI customer service runs in four stages: a free requirements review, PoC about NT$300,000 to 800,000, production launch about NT$800,000 to 2.5 million, and cross-department, multi-system projects quoted by scope.
  • Beyond development, AI model API and cloud hosting are billed by actual usage, and maintenance after launch is a separate contract.
  • Platform subscriptions are usually priced per agent seat or per successful AI resolution, which is cheap upfront but harder to predict as volume and integrations grow.
  • Custom integration costs more upfront and suits companies that need to look up orders or member data, or whose data cannot leave a designated environment.
  • The most effective way to control the budget is a PoC on one channel with a set of real questions, expanding only after the accuracy target is met.
Contents

The cost of AI customer service depends mainly on three things: how much data the knowledge base needs to organize, how many service channels and existing systems it connects to, and how strict the acceptance criteria are. The model itself is usually not the biggest cost. Below, using the cost ranges JoinX publishes on its AI customer service page, we break down each stage and compare where platform subscriptions and custom integration differ in cost.

How much does AI customer service cost?

Based on custom projects in Taiwan in 2026, the requirements review is free, a single-use-case PoC costs about NT$300,000 to 800,000, a production launch about NT$800,000 to 2.5 million, and cross-department, multi-system projects are quoted by scope. AI model API and cloud hosting are billed separately by usage.

StageCost rangeIncludesTimeline
Requirements reviewFreeUnderstand your workflow and data, judge whether AI fits, suggest the first use case1–2 meetings
Single-use-case PoCNT$300,000–800,000One channel, limited data, tested under human supervision4–8 weeks
Production launchNT$800,000–2.5 millionPermissions, integration, human handoff, monitoring, testing and deployment2–4 months
Cross-department, multi-systemQuoted by scopeMultiple knowledge bases and AI agents, cross-system workflows and full governance4+ months

The ranges are a starting point for budgeting, not a fixed quote. Where a stage lands within its range depends on data volume, number of channels and acceptance requirements.

Where does the money go at each stage?

Most of the money goes to work outside the AI itself: organizing the knowledge base, designing human handoff, integrating existing systems and testing accuracy. The later the stage, the larger the share of system integration and permission design.

Requirements review: confirm whether AI fits

The requirements review is free and usually takes 1–2 meetings. We look at your current service workflow, common questions and which documents hold the answers, judge whether AI fits, and suggest the first use case, so the company knows which questions suit AI and which do not before investing in a PoC.

PoC: measure accuracy with real questions

We build the knowledge base and answer flow for one channel and a set of real questions, and measure accuracy, citation accuracy and handoff rate. Where the cost lands depends mainly on how hard the documents are to organize: consistent formats and current versions keep it low, while documents scattered across departments with mixed versions mean more work.

Production launch: permissions, handoff and monitoring

We connect the website, app or LINE Official Account, set up human handoff, permissions and monitoring, and integrate order or member systems where needed. AI customer service that looks up orders and members costs more than one that only answers common questions, because every system connected needs its own permission design and testing.

Cross-department, multi-system: quoted by scope

When several departments each have their own knowledge base, tasks run across systems, or AI agents need to create orders or change bookings, scope varies widely, so we quote by the number of modules and systems. These projects usually run in phases, each with deliverables you can accept.

What are the monthly costs beyond development?

After launch there are three main recurring costs: AI model API, cloud hosting and maintenance. The first two are billed by actual usage; maintenance is a separate contract covering knowledge base updates, answer rule adjustments and system monitoring.

Cost itemBillingWhat drives the amount
AI model APIBy usage (volume of text processed)Monthly conversations, length of documents cited per answer
Cloud hosting and databaseBy usage and specificationKnowledge base size, concurrent users, whether private deployment is needed
MaintenanceSeparate contractKnowledge base update frequency, number of channels and systems monitored

When estimating, use the past 3 months of service records to project monthly conversations, then allow for peak usage. NIST's AI Risk Management Framework also lists ongoing monitoring after launch as a governance priority, so maintenance is not something to cut.

How do platform subscriptions and custom integration differ in cost structure?

Platform subscriptions need little upfront investment and are priced per seat or per successful resolution, which suits teams with standard workflows and no deep integration. Custom integration costs more upfront but connects to existing systems, keeps data in a designated environment, and is easier to predict as usage grows.

ItemPlatform subscriptionCustom integration
Upfront costLow; usually ready once activatedBilled by stage; see the cost table above
PricingPer agent seat, per successful AI resolution or by usageOne-time development fee; model API and hosting by usage
Integration with existing systemsDepends on the platform's integrations; special systems need extra developmentOrder, member, booking and other systems connected as needed
Data locationThe platform's cloud environmentA designated environment, including private deployment
Long-term costGrows with conversation volume and seatsGrows mainly with model usage; development is not paid again
Best forStandard workflows, quick trialsOrder and member lookups, complex permissions, data that must stay in-house

Examples of published prices on official websites (checked November 4, 2026, in US dollars):

  • Zendesk: The Suite Team plan is US$55 per agent per month billed annually (US$69 billed monthly). AI agents are billed per automated resolution; beyond the plan's included allowance, each costs US$1.50 on a committed plan or US$2.00 pay-as-you-go (Zendesk pricing page).
  • Intercom: Fin AI Agent costs US$0.99 per successful outcome (Intercom pricing page).

Pricing per successful resolution is very economical at low volume, but the cost grows with usage: at US$0.99, 10,000 successful resolutions a month cost about US$9,900, or about US$119,000 a year, before agent seat fees. Actual prices are those on each platform's website, and plans and prices may change at any time.

When should you choose a platform, and when custom work?

A platform subscription is the better value when questions are mostly common FAQs, no internal company data needs to be looked up, and the service team wants to try quickly. Custom integration fits better when you need to look up orders or member data, have special permissions, cannot store data in a third-party cloud, or have volume high enough that per-resolution pricing stops making sense.

You can also mix the two: use a platform to learn what customers ask most, then hand the parts that need system integration to custom development. When comparing, use the same period, for example 3 years, and add up development, subscription, usage and maintenance fees rather than looking only at first-year spending.

If your AI customer service will later run tasks across systems, such as creating orders or changing bookings, see the permission and audit design in enterprise AI agents and RAG.

How do you keep an AI customer service budget from overrunning?

The most effective approach is to narrow the first phase: one channel, a set of real questions and a clear accuracy threshold. Expand only after the PoC meets the target; if it does not, stop there and limit the loss to the PoC cost.

  1. Pick the channel with the most questions. For example the website chat or LINE Official Account, rather than launching every channel at once.
  2. Organize the knowledge base first. Remove outdated and duplicate documents, and prioritize FAQs with clear answers.
  3. Define the accuracy threshold. Use past real questions as the test set and state what accuracy counts as passing.
  4. Write human handoff into the workflow. Refunds, complaints, billing and other questions that carry responsibility go straight to staff, not to AI.
  5. Start with read-only access. When integrating systems, read first without writing, and keep human confirmation for any changes.
  6. Review conversation logs after launch. Fill knowledge base gaps from wrong answers and handoffs, then decide whether to add channels.

What AI conversation projects has JoinX built?

JoinX's AI conversation projects include an AI agent interface for insurance, AI search added to an existing platform, and JoinX GPT, which answers questions on our website every day. What they share is that AI answers from existing data and hands anything it cannot answer to people.

  • Insurance company: In 2026 we designed the conversational interface and workflow of an AI agent for insurance sales so frontline staff would actually use it, with a person confirming at the end.
  • GoodParty: We added AI search to the platform JoinX built and continues to maintain, reusing the existing data and admin back end without replacing any system. See the GoodParty case study.
  • JoinX GPT: The AI assistant at the bottom right of joinx.co answers questions from the website's services, process and articles, and hands anything it cannot answer to sales.

Financial institutions adopting AI customer service also need governance documents aligned with the FSC's AI guidelines; see AI adoption in financial services.

FAQ

What is the minimum cost of AI customer service?

For a custom project, the first step is a free requirements review; a testable version from a single-use-case PoC costs about NT$300,000 to 800,000. If you only need to answer common questions without system integration, you can also start with a platform subscription.

How are the monthly costs of AI customer service calculated?

There are three main items: AI model API, cloud hosting and maintenance. The model API is billed by the volume of text processed, hosting by usage and specification, and maintenance under a separate contract. Use the past 3 months of service records to project conversation volume, then allow for peaks.

Which is cheaper, a platform subscription or custom development?

It depends on the period and volume. With low volume and no system integration, a platform subscription is cheaper; with high volume, order and member lookups, or data that cannot be stored in a third-party cloud, custom integration is usually easier to control over time. Compare all costs added up over a 3-year period.

What if the AI customer service PoC misses its target?

First find out why. Most often the knowledge base documents are incomplete or outdated, so fill the gaps and test again; if the questions themselves need human judgment, such as complaints and refunds, route them straight to staff. The purpose of a PoC is to confirm feasibility before investing in a production launch.

Is there a maintenance fee after AI customer service goes live?

Yes. Products, rules and FAQs change, so the knowledge base must be updated, answer rules adjusted for wrong answers, and tests rerun when the model version changes. Maintenance scope and fees are set in a separate contract based on the number of channels and systems.

Sources

  • Zendesk plans and pricing (checked November 2026)Zendesk
  • Intercom Pricing (checked November 2026)Intercom
  • AI Risk Management Framework (AI RMF 1.0, 2023)National Institute of Standards and Technology (NIST)

These are the regulations and sources cited at the time of writing. Rules may be revised; always check the latest official announcement.

About the author

Jack Lee

Founder & CEO, JoinX

Jack founded JoinX in 2016 and leads a team of 80+ people across Taiwan and Japan. JoinX has delivered 520+ projects for 220+ clients, including 56 mid-to-large systems.

Related articles