AI customer delivery enquiry handling.

Which delivery enquiries AI can answer alone

Status questions, once the enquirer is verified. Everything else gets a draft and a person. The split below is the one we would start with; your own history will move the lines.

How an enquiry moves through the system

Inputs

Emails, web chat and contact forms; the order record (contact details, items, carrier, tracking number); carrier tracking events; the proof of delivery record where one exists.

Systems

Your helpdesk (Zendesk, Intercom or similar), the store or order system (Shopify, WooCommerce or Cin7), carrier tracking APIs, and read-only access to POD images.

Steps

Classify the intent; find the order; verify the enquirer against the order’s contact details; fetch the latest scan; pick the reply type from the table; send or draft; tag and log.

Human review

For the first four weeks agents approve every reply. After that, status replies auto-send and a team lead samples 10% weekly; every other category stays draft-only.

Outputs

Sent status replies, approval-ready drafts, tagged tickets, and loss or damage cases opened with tracking history attached.

Measures

First response time, share resolved without an agent, reopen rate within 48 hours, auto-replies that should have escalated (target zero), and disclosure complaints.

Risks

Details sent to the wrong person, delivery dates invented, stale tracking quoted as current, and customer messages that try to instruct the system.

Illustrative example: an online retailer’s month

Illustrative example: a Wellington online homewares retailer shipping 4,000 orders a month through two courier companies. Assumptions: 15% of orders prompt a delivery enquiry (600 a month); 65% are status, 20% delivery window or failed delivery, 15% loss, damage, redirect or complaint; agents currently take 3, 4 and 10 minutes for those groups. All figures are synthetic.

The pricing model matters. Intercom lists its Fin agent at $0.99 per outcome, and Zendesk charges for AI agents per automated resolution and bills in USD, EUR, GBP or BRL, not NZD (both as listed on 19 October 2026). At about a US dollar each (check the currency on your quote), 390 automated status replies cost roughly US$390 a month to save about 19 agent hours. A custom build swaps that for token costs of cents per reply plus build and maintenance.

The New Zealand rules your replies must respect

Who owns a lost delivery. Where a supplier is responsible for delivering or arranging delivery, the Consumer Guarantees Act 1993 guarantees the consumer receives the goods at the agreed time, or within a reasonable time, and treats the carrier as the supplier’s agent (section 5A). A retailer’s assistant must never tell a consumer to sort it out with the courier.

What it may promise. The Fair Trading Act 1986 bars unsubstantiated representations in trade (section 12A): a delivery promise with no reasonable basis breaches it even if the parcel happens to arrive on time. Dates come from the carrier’s estimate, attributed as such.

Who it may tell. Under privacy principle 11, personal information should be disclosed only on reasonable grounds, such as disclosure to the person it is about. An address, door photo or delivery time sent to the wrong email is a privacy breach; if serious harm is likely, the Privacy Act 2020 requires notifying the Commissioner as soon as practicable (section 114).

Where a person must take over

• Address changes and redirects after dispatch, which carry fraud and personal safety risk.
• Declaring a parcel lost, and any refund or replacement decision.
• Damage reports, which need the evidence from proof of delivery extraction.
• Anyone distressed, or asking about someone else’s delivery.
• Wider delays, which the ops desk sees first in its shipment exception summaries and should announce once, not ticket by ticket.

Controls to have in place before auto-replies go live

• The identity rule is written down: which fields must match before any delivery detail is shared.
• The assistant has read-only access to orders and tracking; it cannot change addresses or issue refunds.
• Every date in a reply is quoted from the carrier with its scan time (“last scanned in Palmerston North at 6.12am”).
• Fifty real past emails per escalation category have been tested and none auto-sent.
• Instructions inside customer messages are treated as text, not commands; see recognising prompt injection.
• Reopen rate and the weekly sample are reviewed by a named team lead.

The drafting habits behind good replies are covered in using AI for customer support drafts, and store connections in AI and Shopify. For a build, see AI customer support systems and the wider AI for logistics in New Zealand guide.

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