Yes. A small courier can let AI answer where-is-my-parcel messages, as long as every reply is built from the live tracking status and an approved template. A person still handles damage, loss, "marked delivered but not received", compensation, repeat chasers and anything involving a vulnerable customer or a high-value parcel.
The mistake that costs couriers money is letting the AI guess. A model with no tracking data invents an arrival window. A model with data but no rules promises a refund to calm an angry customer. So the job has two halves: give it the scan history for that one parcel, and fence it with rules about what it may say and when it must stop and pass the message to someone.
Sort a week of messages before automating any of them
Export or copy a week of inbound messages from every channel you use (email, web form, text, WhatsApp, voicemail transcripts) and put each one in a bucket. A handful of question types usually make up most of the pile, and the risky ones are a small slice. Your own count shows where AI will save time and where it would only add risk. If the pile is large and untidy, the approach in AI ticket triage helps you tag and route it first.
Label every bucket with one of three routes: send (the AI's reply goes out on its own), draft (the AI writes, a person approves) or stop (a person answers, the AI only summarises). The table below is a starting point for a local courier.
| Message type | Typical wording | Route |
|---|---|---|
| Parcel in transit | "Where is my parcel?" | Send from template |
| Out for delivery | "What time will it get here today?" | Send; no time unless the system holds a window |
| Delivery attempted | "The driver says I wasn't in" | Draft at first, send once checked |
| Held at depot | "Where do I collect it?" | Send from template |
| Address or safe-place change | "Can you leave it with a neighbour?" | Draft; dispatcher approves |
| Marked delivered, not received | "It says delivered but nothing is here" | Stop |
| Damaged or missing contents | "The box was crushed" | Stop; AI collects the details |
| Complaint about a driver | "Your driver was rude" | Stop |
| Price or booking | "How much for tomorrow?" | Draft from the rate card |
Which tracking statuses the AI may quote, and what each triggers
Every courier system words its statuses differently, so start by writing yours down. The AI should quote the latest status in plain words, add nothing, and know which statuses are never answered without a person. The table maps the common ones. Treat the last column as the part you adjust to your own service levels.
| Status in your system | What the AI may say | What it must not say | Hand to a person when |
|---|---|---|---|
| Booked, label created | We have the booking and have not collected yet | A collection time that is not in the data | Collection is overdue past your cut-off |
| Collected | Collected at [time] from the sender | A delivery day the route plan does not show | The sender says it was never collected |
| In transit or at a depot | Last scan at [depot], [time] | "It will arrive tomorrow" without a planned date | No scan for longer than your threshold, say one working day on a next-day service |
| Out for delivery | On the vehicle today; window if the system holds one | An exact arrival time | The route is delayed or cancelled, or the customer is about to leave |
| Delivery attempted | Attempted at [time], not handed over; redelivery options | A reason ("nobody was in") or any blame | The customer says no attempt was made |
| Held at depot | Where and when to collect, what to bring | Storage fees that are not on your rate card | The customer cannot collect in person |
| Delivered | Delivered at [time] to [recorded place or name] | That the customer must have it | Always, if they say it did not arrive |
| Exception, returned, unknown | A delay we are checking | Causes, guesses or new dates | Always |
The rule that does the most work is "quote the status, add nothing". A typical failure (illustrative) is a model turning the plain status "Delivery attempted" into "the driver couldn't find your address". Nothing in the data says so. The customer replies that the address is fine, and a small query becomes a complaint. The fix is one line in the instructions: use the status wording, and give no reason unless a dispatcher note supplies one.
Six reply templates for the commonest tracking questions
Each template below is written to fit a single 160-character text once the blanks hold typical values, so check yours: a long depot name can push a message into two billable segments. The AI chooses the template from the status, fills the blanks from the scan history and changes nothing else. For email, keep the same facts, put the tracking number in the subject line and finish with how to reach a person.
In transit
Hi [first name], parcel [tracking no.] was last scanned at [depot] at [time] on [date] and is on its way to you. Reply if you need a person.Out for delivery, with a window in the data and without one:
Hi [first name], parcel [tracking no.] is on the vehicle for delivery today. Window: [window]. We'll text if that changes. Reply if you need a person.
Hi [first name], parcel [tracking no.] is on the vehicle for delivery today. We can't give a time yet but will text if anything changes.Delivery attempted
Hi [first name], our driver tried to deliver [tracking no.] at [time] and couldn't hand it over. We can redeliver [next slot]. Reply with your choice.Held at depot
Hi [first name], parcel [tracking no.] is at our [depot name] depot. Collect it [opening hours] with photo ID and this number.Customer says it was not received (this one acknowledges and hands over, it never argues):
Hi [first name], our record shows [tracking no.] was delivered at [time] to [recorded place or name]. [team member] is checking the proof of delivery now.Gap in the scans
Hi [first name], we can see a gap in the scans for [tracking no.]: the last one was [time] at [depot]. We're checking and will update you today.Two details matter more than they look. First, anything in the message that says "will update you" or "is checking" is a promise, so it needs an owner and a deadline inside your business: pick the response time you can honour and write the template to match. Second, chat assistants like curly apostrophes and emoji. SMS providers such as Twilio treat either as a different character set, which cuts a segment from 160 characters to 70 and can turn one text into three billable messages. Tell the AI to use straight apostrophes and no emoji in texts, and look for both in your test.
The instruction block that keeps the AI in its lane
Put the rules below in a ChatGPT Project or Claude Project alongside your templates, or in the instruction field of your helpdesk's AI. Each time, you give it the customer's message followed by a block of tracking data for that one parcel. It returns an action line, so an automation or a human can see at a glance whether to send or to stop.
ROLE: You draft replies to customers of [courier name], a small local courier. You never send anything. A person or an approved automation sends it.
INPUT EACH TIME: the customer's message, then a block headed TRACKING DATA with the parcel's scan history, service level and any dispatcher notes.
RULES
1. Use only TRACKING DATA and the approved templates. If a fact is missing, say we are checking. Never guess.
2. Quote the latest status in plain words. Give no reason for anything unless a dispatcher note supplies it.
3. Never give an arrival time unless a delivery window is in the data.
4. Never promise or imply a refund, compensation, fee waiver or replacement. Never say a claim will be accepted or refused.
5. Never blame the customer or the driver.
6. Reveal status, time and depot only. Never repeat full addresses, recipient names or sender details.
7. If the message mentions damage, loss, theft, missing items, "delivered but not received", a driver complaint, a solicitor, a chargeback, a medical or perishable item, or is the third contact about one parcel, output HANDOFF, a three-line summary for staff, and a holding reply that names who will reply and by when (the fifth template above for "not received", the damage intake message for damage).
8. Text replies: plain text, straight apostrophes, no emoji, under 160 characters. Start "Hi [first name]," and, where it fits, end with how to reach a person.
OUTPUT
Line 1: ACTION: SEND or HANDOFF
Line 2: REASON: one short phrase
Then the reply, or the summary and holding reply.Two illustrative exchanges show the format. The data is invented and the replies are the corrected versions:
CUSTOMER: Where's my parcel?? Was meant to be here yesterday
TRACKING DATA: 12 Oct 16:42 Collected | 13 Oct 05:58 Arrived depot | 13 Oct 14:10 Delivery attempted, no access | 14 Oct 08:05 Out for delivery
ACTION: SEND
REASON: out for delivery after missed attempt
Hi [first name], parcel [tracking no.] is out for delivery today after a missed try at 14:10 yesterday. No time yet; we'll text if anything changes.
CUSTOMER: It says delivered but I haven't got it
TRACKING DATA: 14 Oct 13:12 Delivered, left with neighbour (driver note)
ACTION: HANDOFF
REASON: delivered-not-received
Summary: (1) Customer says parcel shown delivered 13:12 has not arrived. (2) Driver note: left with a neighbour. (3) First contact; customer wants it today.
Holding reply: Hi [first name], our record shows [tracking no.] was delivered at 13:12 to a neighbour. [team member] is checking the proof of delivery now.First drafts rarely look this clean. The usual slips are a reason added to the first reply ("because nobody was in"), which rule 2 now forbids, and the neighbour's house number copied from the driver note into the second, which rule 6 now blocks. Expect two or three rounds of testing before the rules feel tight, and add a rule every time a bad reply slips through.
Hand-off triggers: the messages a person must pick up
Write the triggers as two locks. The AI's instructions carry the first, and a simple keyword filter in your automation or helpdesk carries the second, so a message the model misreads still gets caught. Every trigger names a person and a response time you will honour, and the holding reply promises exactly that and nothing else.
| Trigger | Why a person | Who picks it up |
|---|---|---|
| Damaged, broken, crushed, leaking, opened, items missing, stolen | Liability and evidence; the AI must never admit fault | Claims owner, same working day |
| Status "delivered" plus "not received" | Needs the proof of delivery checked and the driver asked | Dispatcher, within the hour |
| Third message about one parcel in 48 hours | The process is failing, not the wording | Supervisor |
| Declared value over your threshold, fragile, medical or perishable goods | Higher cost of a wrong answer | Supervisor, before any reply |
| Solicitor, chargeback, review, press, social media | Relationship and reputation risk | Owner |
| Driver behaviour, vehicle safety, near misses | Conduct and safety, never templated | Manager |
| Illness, age, mobility or "I have to be in for it" | Duty of care; a phone call beats a text | Dispatcher |
| Redirect to a new address after dispatch | Fraud risk; identity check needed | Dispatcher |
| "Can I speak to someone" or "is this a bot?" | Always honoured at once | Next free team member |
The proactive route also reduces what lands in the queue. If customers get a text when a parcel is collected and again when it is out for delivery, fewer of them ask. That is the subject of cutting parcel-tracking calls with AI delivery updates, and the two approaches work well together.
Damage claims: what the AI collects and what it never says
The AI's job at a damage report is intake, not judgement. It should gather what a claims person needs, tell the customer a person will review it and by when, and say nothing about fault, cover or payment. Phrases to ban outright: "we're sorry we damaged it", "this is covered", "we'll replace it", and "you should have insured it". A softer line works instead: "I'm sorry the parcel arrived in this condition. A colleague will review it and reply by [time]."
Hi [first name], I'm sorry the parcel arrived in this condition. To help [team member] review it, please reply with:
1. Three photos: the outside of the box, the delivery label, and the damaged item.
2. What the item is and roughly what it is worth.
3. How it was packed, if you know (box type, padding).
4. When you received it and whether you signed.
We'll reply by [time]. This message does not confirm or refuse a claim.When the customer answers, the AI prepares a hand-off note, so the claims person starts with the facts instead of a scroll of messages:
CLAIM INTAKE (AI-prepared, needs review)
Parcel: [tracking no.] | Service: next-day | Delivered: 14 Oct 13:12, left with neighbour per driver note
Customer says: outer box crushed at one corner, ceramic item inside broken. 3 photos received (outside, label, item).
Packing as described: single-wall box, padding on two sides. Declared value: not stated.
Missing: sender's packing details, evidence of value, whether anyone signed.
Customer expects: replacement or refund. NOT promised.
Suggested next step: check proof of delivery and driver notes; contact the sender if a shop sent it.How a claim is then triaged, matched to proof of delivery and answered is covered in AI for damage claims and disputes in courier services. Keep the boundary clear: the AI collects and summarises, and a person decides.
Missed deliveries and the parcel marked delivered but not received
These are two different problems. A failed attempt is routine, and the AI can handle most of it: offer the next slots from a list you maintain, confirm the delivery address, apply a safe-place option only if your policy allows it, and quote a redelivery fee only if your rate card has one. If the customer disputes that an attempt happened, it stops and hands over.
"Delivered but not received" is never a template job. The AI's only contribution is speed and tidiness. A plausible sequence (illustrative): a customer writes at 15:20 that nothing arrived. The holding reply goes out within a minute. The coordinator's queue gets the scan history and the driver's note in one message. The proof-of-delivery photo shows a door with the number 12, but the address is number 14. A call to the driver and a knock next door find the parcel by 16:00. Without the summary, the coordinator spends the first ten minutes finding the scan history before making a single call.
A six-driver courier's Monday, before and after
A local courier with six drivers handling about 300 parcels a day for shops and small businesses makes a useful test case. All figures here are illustrative; run your own count before you rely on any of them. On a typical weekday it gets around 60 inbound messages. After a week of sorting, the buckets look like this.
| Bucket | Messages a day | Route |
|---|---|---|
| In transit and out for delivery | 33 | Send |
| Held at depot or collection questions | 6 | Send |
| Delivery attempted or missed | 12 | Draft |
| Address or safe-place changes | 3 | Draft |
| Delivered but not received | 3 | Stop |
| Damage or missing contents | 2 | Stop |
| Driver complaint | 1 | Stop |
| Total | 60 | 39 send, 15 draft, 6 stop |
Before, the dispatcher spent about three minutes on each message (find the parcel, read the scans, type a reply): 60 x 3 = 180 minutes a day. In the first week, with every draft approved by a person, the sums change to 54 drafts at about 45 seconds each (about 40 minutes) plus 6 hand-offs at about 6 minutes each with the AI's summary (36 minutes), roughly 76 minutes in total. Once the "send" routes run on their own, the day looks like this:
- 39 messages sent automatically, plus a 10-minute spot-check of a sample.
- 15 drafts approved at about a minute each: 15 minutes.
- 6 hand-offs at about 6 minutes each: 36 minutes.
That totals 61 minutes, against 180, so roughly two hours a day come back to the dispatcher. Over 22 working days that is about 44 hours. The more important number is not time. It is that the six hard conversations now start with the scan history, the photos and the customer's account laid out in one place, which is where a courier keeps or loses a regular account.
Copy-paste, helpdesk AI or automation: what each set-up costs
A Project and copy-paste
The cheapest start: a ChatGPT or Claude individual plan (about $20 a month at list price; vendors change plans often, so check the pricing page) holds the instructions and templates. A person pastes in the message and the scan history, and copies the output out. It suits a few dozen queries a day. Business plans such as ChatGPT Business and Claude Team need at least two seats, so a one-person operation is better on an individual plan. Do not paste data you would not want sitting in a consumer account: switch off the model-training setting, or use a business plan, and keep addresses out of prompts where you can.
A helpdesk with an AI add-on
If messages already land in a shared helpdesk, the AI sits inside it. As of October 2026, Help Scout lists Standard from $25 per user a month and prices its AI add-on at $0.75 per resolution (its FAQ calls it AI Answers, its pricing table calls it Envoy). It counts a resolution only if the customer does not escalate, search the help docs, ask again or say they need more help, and it offers a monthly cap on AI spend. Its pricing page also describes three months of free AI resolutions for new accounts, so check the terms on the page. Intercom's Fin is priced at $0.99 per resolved outcome, counts an "assumed resolution" when a customer goes quiet for 24 hours after Fin's last answer, and can also count a completed workflow, hand-offs included, as an outcome. Quick sum: 300 AI-resolved conversations a month cost $225 on Help Scout's rate or $297 on Fin's, plus seats.
Read the definition of "resolved" before you sign. A customer who gets an unhelpful answer and gives up is silent, not satisfied. A courier set-up hands off more conversations than most (every damage report and every "delivered but not received"), so ask how a conversation the AI passes to your dispatcher is billed. Also ask the vendor to show a live tracking lookup using one of your own parcel numbers, because most helpdesk AI answers from help articles and will only quote scans if someone has connected your data.
An automation that reads your tracking data
A workflow in Zapier or Make can catch each new message, look the parcel up in your tracking export or system, ask an AI step to draft the reply, and post the draft into your helpdesk or a team channel for approval. Zapier's Free plan allows 100 tasks a month and only two-step Zaps, and Professional starts at $19.99 a month on annual billing for 750 tasks; an AI step can use more than one task per run, and Make's AI steps can use more than one credit. The comparison of tool types and their costs in customer support automation tools: seven types and what each costs helps you decide how far to go.
If customers message you on WhatsApp, two more points apply. From 1 October 2026 Meta charges for free-form service replies sent through the WhatsApp Business Platform beyond the first 1,000 service messages per business number each month; the free WhatsApp Business app is separate. And Meta's business messaging policy expects a clear route to a person when you automate, which the hand-off triggers above already give you. Confirm the details with your provider.
Test it on last month's messages before a customer sees a reply
Pull 30 old messages across every bucket, including eight awkward ones. For each, rebuild the scan history as it stood at the time, run the AI, and score the result against the checks below. Do this before any live channel is connected.
| Check | Pass mark |
|---|---|
| Invented facts (times, reasons, causes) | None in 30 |
| Trigger messages sent to a person, not answered | All of them |
| Promises or admissions about money or fault | None |
| Text replies under 160 characters, plain characters | All text replies |
| Tone you would be happy to sign | 27 of 30 or better |
Then go live in steps. Week one: the AI drafts everything and a person approves every reply. Week two: let the in-transit, out-for-delivery and held-at-depot routes send on their own, and keep attempted-delivery replies on approval for another fortnight. After that, take a weekly sample of 20 sent replies and score them again, using the routine in spot-checking AI support replies. Track four numbers: time to first reply, the share of messages handed to a person, repeat contacts about the same parcel within 48 hours, and the count of wrong answers found in the sample. A rising repeat-contact figure is an early sign that replies sound helpful but are not.
Four mistakes that surface in the first month
These four are illustrative composites rather than reports from named firms, but each follows from a rule that is easy to leave out.
- The "today" promise. The status said "out for delivery", so the AI told customers their parcels were arriving today, but one van was off the road and its route never ran. The fix was a "route delayed" flag from dispatch passed into the tracking data, plus rule 3.
- The apology that read as an admission. A customer wrote "you damaged my parcel" and the AI replied "sorry we damaged it". Acknowledge the condition, never the cause. Rule 4 and a ready-made phrase solved it.
- The text that cost three times. One curly apostrophe turned a 150-character reply into a 70-character-per-segment message and tripled the bill for it. Straight apostrophes only.
- Silence billed as success. A helpdesk counted conversations that went quiet as resolved. Sampling the "resolved" list each week showed customers who had simply given up.
Most of these come down to the same habit: when a reply goes wrong, find which rule was missing and add it, rather than only correcting that one message.
Questions couriers ask before switching AI replies on
Should replies say they come from an AI?
Yes. Sign off as your company's automated assistant and always offer a way to reach a person. If you have customers in the EU, the AI Act's chatbot transparency rule has applied since 2 August 2026, so plain disclosure is the safe default for every customer. It also lowers anger when the assistant gets something wrong.
Can the AI look up a tracking number by itself?
Only if you connect it to the data. A chat assistant such as ChatGPT or Claude cannot see your dispatch system, so the scan history has to be pasted in, passed by an automation, or pulled by a helpdesk integration. Ask any vendor to demonstrate a live lookup on one of your own parcel numbers before you buy.
What if the customer gives a name but no tracking number?
Do not let the AI guess which parcel they mean. Ask for the tracking number or order reference, and give status only once it matches. If they cannot supply either, hand the message to a person, because sharing delivery details with someone who is not verified is a privacy risk.
How do I keep angry customers away from the AI?
Add stop conditions: words such as complaint, unacceptable or solicitor, a third message about the same parcel within 48 hours, and any request for a person. Each one routes the conversation to a named team member, with the AI's three-line summary attached, rather than producing another templated reply.
Further reads
- AI Courier Quotes: Reply to Delivery Enquiries From Your Rate Card — Turn price enquiries into approved quote replies from your rate card.
- How Can AI Help Courier Services? Quotes, Dispatch and Proof — See where AI fits across quotes, dispatch and proof of delivery.
- Best AI Customer Support Software for Small Teams in 2026 — Compare helpdesk tools before choosing a set-up for your inbox.
- Customer Support AI Tools for Phone Repair Shops: Repair Status Updates — Another status-update business, with its own reply and text-message limits.
- Can a Small Courier Firm Use AI to Plan Delivery Routes? — Fewer late parcels means fewer where-is-it messages to answer.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Help Scout pricing page, Intercom Fin pricing, Zapier plans page, Meta's WhatsApp Business Platform pricing and policy pages, Twilio SMS encoding documentation, all checked October 2026.