AI customer support systems.

What an AI customer support system does

An AI customer support system does four jobs: it classifies each incoming message, routes it to the right queue, drafts or sends an answer from your own help content, and escalates whatever it should not handle. The model is the smallest part. The work is in the ticket categories, the knowledge it answers from, and the rules for when it stops.

It is different from a website chatbot. A support system works the queue your team already works (email, web forms, chat and social messages) inside the ticketing tool you already pay for. If you want a customer-facing assistant on your website, see AI chatbot development; for phone lines, see AI voice agents.

The customer support AI assistants we build

“AI assistant” covers very different systems, and a New Zealand business usually needs one or two of them, not all four. We build each as a customer support AI assistant connected to the help desk, knowledge base and order or account systems you already run:

Most first projects start with agent-assist drafting, because a person checks every answer while the assistant learns your tone and policies. Customer-facing answering comes later, once accuracy on your own questions is measured.

Who it is for

It fits when most of these are true:

• A few question types make up most of the volume. Where is my order, how do I reset my login, can I change my booking, can you resend the invoice.
• The answers already exist in writing. Help articles, policies, macros and good past replies.
• Tickets arrive in a system with an API. Zendesk, Intercom, Freshdesk, HubSpot, or a Microsoft 365 or Google Workspace shared inbox.
• Someone owns support quality and can approve a list of what the system may answer on its own.

It fits poorly when every ticket is bespoke advice, when volume is low enough that a tidy macro library would do the job, or when the answers live only in one experienced person’s head. In that last case, write things down first. Our lesson on building customer FAQs from support tickets shows how to do it from the queue you already have.

What we deliver

Ticket categories

Drawn from a sample of your real tickets, with volumes, so you can see which types are worth automating and which are not.

Routing rules

Category, urgency, language and customer tier, written back to your ticketing tool as tags and assignments.

Answer source

A cleaned, searchable index of help articles, policies and approved past replies, plus a list of gaps for your team to fill.

Grounded drafts

Replies that cite the article or policy they used, shown to agents inside the ticket.

Auto-send allowlist

The specific question types allowed to send without review, each with its measured accuracy on your tickets.

Escalation design

Triggers, the hand-off message, and the context a person receives.

Test set and dashboard

Labelled tickets used to check every change, and weekly accuracy, reopen and escalation rates by type.

How the work runs

1. Discover: read the queue. We pull a sample of recent tickets, masking personal details where we can, sort them into types and sit with the people who answer them. You see which types are frequent, which can be answered from written sources, and which should never be automated.
2. Prove: draft-only shadow mode. The system routes and drafts but sends nothing. Agents accept, edit or reject each draft, and every decision becomes test data. This is where missing articles, vague policies and deceptively simple questions turn up.
3. Prove: open the allowlist. Question types that reach the accuracy bar you set move to automatic sending one at a time, with a daily sample still checked by a person.
4. Scale: monitor and extend. We track reopens, escalations and complaints by type, pull back anything that drifts, and add new types only when the evidence supports it. Ongoing care can sit with your team or with our managed AI operations.

Escalation is the product

Escalation rules decide whether customers trust the system. Every build has explicit triggers, and the person who picks up the ticket gets a summary, the customer’s history and the draft the system would have sent.

What it costs to build and run

Integration and review effort drive the cost far more than model fees. The main drivers are how many systems the build reads from and writes to (ticketing, orders, CRM, accounting), how much help content needs writing or fixing before answers can be grounded, the number of channels and languages, and how long draft-only mode runs before you trust automation. Model usage per ticket is usually small; review time in the first months is not. Our guides to AI automation pricing and what custom AI projects cost set out the arithmetic.

One caution when comparing products. Some vendors bill per resolution. Intercom lists its Fin AI agent at $0.99 per outcome and counts a resolution when “no further help is requested after the last AI answer” (Intercom, accessed 19 October 2026). A customer who gives up also requests no further help. Check reopen rates and repeat contacts yourself before trusting any resolution figure, ours included.

New Zealand rules that shape the design

• Privacy. Tickets are full of personal information. The Privacy Commissioner’s starting point is that the Privacy Act “applies to everyone using AI tools in New Zealand” (OPC), and its AI guidance expects human review before acting on AI outputs. We record where every ticket is processed: Anthropic’s own API, for example, offers only US or global processing (Anthropic).
• Consumer law. Services supplied to consumers carry a guarantee of reasonable care and skill under the Consumer Guarantees Act 1993, and the Fair Trading Act 1986 bans misleading conduct. An AI reply that misstates your returns policy is still your reply.
• Sounding human. In Xero’s May 2026 survey of its NZ customers, 55% of small and medium businesses worried that AI in customer communications would make their brand feel less human (Xero). Saying plainly when a reply is AI-drafted, and keeping a person one click away, answers most of that worry.

Where AI support goes wrong

• Confident answers from stale content. If the help centre is out of date, the system repeats it fluently. Assign an owner for every source before launch.
• Policy edge cases. “Can I return a sale item that was a gift?” looks routine and is not. Shadow mode exists to find these before customers do.
• Instructions hidden in customer messages. A ticket can contain text written to steer the model. Prompt injection is first on the OWASP Top 10 for LLM Applications 2026, so customer text is treated as data, and the system cannot issue refunds or change accounts without a person. See our guide to prompt injection testing.
• Hidden decline. Deflection can rise while satisfaction falls. Track reopens and complaints by question type, not just total volume handled.

Work this draws on

• Property management agent for APL Property: an internal agent that triages email, keeps records and follows up for a regional NZ property firm, while staff decide what goes out.
• PainPal: Levy Street worked with BetterWorx on this AI workmate for aches and strains. It answers from expert-reviewed guidance and offers extra support from ergonomic specialists when symptoms suggest something serious.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.