A small shipping company can use AI to draft courier quotes from its current rate card and answer tracking questions from the latest verified scan, while a dispatcher checks both before sending. Give the assistant anonymised shipment fields rather than customer details. For damage claims, use it to list missing evidence and draft an acknowledgement, not to decide liability.
The first requirement is a trustworthy source for each answer. A model that sees “out for delivery” at 09:00 cannot promise arrival at 11:00 unless your system actually provides that window. A quote based on last month's surcharge can be wrong even when the arithmetic is perfect. Keep the rate version, tracking timestamp and claim policy beside the draft so a person can approve the response quickly.
Separate the three queues at a small courier desk
Most incoming courier messages ask one of three different questions: what a job will cost, where a parcel is, or what happens after damage is reported. Give each queue its own fields and review rule. Quote requests need collection and delivery requirements, dimensions, weight, service level and the applicable rate card. Tracking requests need an internal consignment ID and a timestamped status. Claim enquiries need the package condition, delivery record, photographs and the relevant terms. Mixing them in one free-text prompt encourages the model to borrow facts from the wrong job.
Use an internal reference such as Q104 or P208 in the AI workspace. Keep the actual sender, recipient, address, phone number and account details in the dispatch system, not in a general chat prompt. A person can paste a redacted record or create a limited export with service zone, dimensions, status and event times. If a message must be personalised, add the recipient's name after the checked draft returns to the authorised system. Follow the organisation's data-handling policy even when the assistant has business privacy controls.
A filled-in intake for an illustrative quote might read: “Q104; one carton; 38 × 27 × 24 cm; actual weight 6 kg; standard service; collection window 13:00–16:00; delivery zone 3; no restricted goods; rate card version 7.” That is enough for a first price calculation if your rate card defines chargeable weight, zone and extras. It is not enough to commit to a collection slot until dispatch confirms capacity. The quote template should therefore distinguish calculated price from operational acceptance.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowAn illustrative day of 12 enquiries
Imagine a courier firm with four vans and a dispatcher handling 12 customer enquiries before lunch: five new quotes, four tracking requests and three damage reports. It spends too long rewriting routine replies, but each answer still requires current operational data. The firm creates three short forms and uses AI only after an employee has placed the correct facts in the appropriate form. The assistant produces drafts and a list of missing inputs. No message is sent automatically during the pilot.
For quote Q104, suppose the internal rate card says zone 3 standard service has a $14 base charge for up to 5 kg, $2 for each additional chargeable kilogram, and a $4 booked collection surcharge. These are illustrative business rules, not market prices. The carton has a 6 kg actual weight. If the firm's own volumetric rule (a weight worked out from the parcel's size rather than the scale) produces 5 kg for those dimensions, chargeable weight is the greater value, 6 kg. The draft price is $14 + $2 + $4 = $20 before any separately quoted optional service. A human confirms zone, weight rounding and whether the collection surcharge applies.
One quote request gives only “small box, urgent”. The model should return a question list, not a price: exact dimensions, weight, collection window, delivery zone, service requirement and any handling restriction. Another request says “same as last time”, but the previous job used a different service level. The assistant can show the last quote as context if authorised; it must ask for confirmation rather than silently copy it. These two requests remain in the pending queue until dispatch has enough information.
| Queue item | Checked input | Draft output | Human gate |
|---|---|---|---|
| Q104 quote | Zone 3, 6 kg, rate version 7 | $20 draft price | Capacity and surcharge check |
| P208 tracking | Hub scan at 10:42 | Latest event and next check | Confirm no later event |
| C031 claim | Crushed carton photo, delivery record | Acknowledgement and missing evidence | Claims handler review |
The dispatcher measures the pilot by correction type, not by a vague impression that replies feel faster. If the five quote drafts include one wrong zone, that is a data mapping problem. If the four tracking replies contain two invented arrival times, the prompt and approval rule need work. If a claim draft promises a refund, the model has crossed its boundary. The firm then has three specific fixes rather than a generic verdict about AI.
Make the rate card calculable before drafting quotes
A rate card often has exceptions that live in people's heads: minimum charge, weight rounding, remote-area supplement, weekend collection, fragile handling, account discount and fuel surcharge. Put the active rules in a dated table with a named owner. If the firm cannot explain how a human gets the price, AI will not make the calculation reliable. Ask the assistant to enumerate which rule it applied and to flag any rule it could not decide. Do the final maths in a spreadsheet or dispatch system.
For the Q104 example, the assistant could draft: “Thank you for the details. The current standard-service estimate for one 6 kg carton in zone 3 is $20, including booked collection. This is subject to a final check of the collection window and parcel details. Please confirm the dimensions and preferred collection time.” The phrase “estimate” is appropriate while capacity is open. If the rate card has a validity period, add its actual expiry from your policy. Do not ask the model to invent one.
Now change one input: the remeasured carton is 9 kg actual weight. The same illustrative rate rule gives $14 + four additional kilograms at $2 + $4 collection = $26. A quote of $20 would be $6 short. This before-and-after is why dimensions and weights belong in structured fields and why the quote version should be saved. The dispatcher should be able to say exactly which changed measurement caused a revised quote, without blaming a mysterious AI calculation.
For more detail on rate-card replies, see the courier quote drafting tutorial. This shipping-company workflow keeps quoting alongside tracking and claims, so the important hand-off is between a calculated draft and the person who accepts the job. A quote can be numerically correct but operationally impossible if all vans are committed during the requested collection window.
Quote requests often arrive as photographs or forwarded messages, but a clean form is still worth the effort. A dispatcher can transcribe the required facts into the intake row and attach the original message to the internal job record. If AI extracts dimensions from an image, compare them visually with the original before calculating a price. A printed “18” can be read as “81”, and the resulting volumetric weight can be wildly wrong. The form should also have a field for “measurement unconfirmed” so the team can give a conditional estimate instead of pretending a photograph is a calibrated scale.
Answer tracking questions from the event feed
Tracking language should describe an observed event and an agreed next step. A scan saying “received at hub at 10:42” is not proof of a delivery time. Make each AI input include the latest event, event timestamp, expected service commitment if verified, exception flag and next internal check time. The response should distinguish “last recorded” from “currently”. The dispatcher must refresh the event feed immediately before sending; even a ten-minute-old status can be overtaken by a delivery scan.
An illustrative P208 reply could say: “The last recorded event for your parcel was a scan at our hub at 10:42. We are checking its next movement and will update you after the 14:00 dispatch review.” The 14:00 time must come from the firm's actual process. If a driver has already collected it at 11:30, that draft is stale and should be replaced. The assistant should not say “your parcel is safe” or “it will arrive today” unless authorised records support those statements.
Another case is an exception status such as “address query”. Do not paste the address into the prompt. The dispatcher can supply “delivery detail needs confirmation” and have AI draft a request for the customer to contact the secure support channel. The actual correction belongs in the dispatch system. If the parcel is marked delivered but the customer says it is missing, use the delivery record and an investigation workflow; do not reply with “it has been delivered” as though that resolves the dispute.
A customer might ask about three parcels in one message. AI can help put the three status lines into readable order, but each parcel needs its own latest event and timestamp. If one has a delivery proof, one is at a hub and one has no scan since collection, a single reassuring sentence would hide the third problem. Give the dispatcher a three-row summary with evidence and a suggested response per parcel. The delivery updates tutorial goes deeper into proactive updates when tracking calls keep recurring.
Tracking scans also arrive out of order. Suppose P208 shows a hub scan at 10:42 and a collection scan uploaded at 11:05 with an event time of 08:50. The latest uploaded row is not the latest journey event. Sort by the event timestamp and inspect any delayed upload before composing a reply. If your system distinguishes scan time from upload time, keep both fields. The assistant should cite the event it used; a dispatcher can then notice whether the answer was based on a late-arriving record or a genuinely newer movement.
Decide what to do when events stop. For instance, the firm could set an internal review trigger when a parcel has no expected scan by the next scheduled hand-off. That threshold is a business rule tied to its route and service, not a universal shipping standard. The assistant may flag P208 for a human investigation, but it cannot invent an event to fill the gap. A clear “we are checking” message is better than a precise but unsupported estimated arrival.
Use claim intake to collect facts, not assign blame
Damage reports arrive with different levels of evidence. Create a claim intake form that captures the consignment ID in the internal system, reported damage, packaging condition, delivery date, photographs, contents description and any supporting invoice or value record required by the firm's own terms. Ask AI to compare the redacted form against that checklist and draft an acknowledgement. The claims handler reviews deadlines, coverage, exclusions and any decision. Those are policy and contract matters, not facts a chatbot should guess.
For illustrative claim C031, the customer reports a crushed outer carton and attaches one photograph of the box. The delivery record shows a completed delivery but the record available to the AI says nothing about the contents. A suitable acknowledgement thanks the customer, confirms the claim reference and asks for photographs of the contents and inner packaging through the approved channel. It should not say that the courier caused the damage or that compensation has been approved. The next action is evidence collection.
A second damage example has complete photographs but no proof of value. The assistant can identify that gap if the firm's policy requires the document, then use approved wording to request it. If a customer cannot provide it, the claim handler decides the next step under the actual terms. Do not send an AI-generated list of arbitrary documents. Equally, do not request unnecessary personal data because a model suggested “more evidence”. Ask only for fields the team genuinely uses to assess the claim.
In a third case, a fragile item was reported damaged after the parcel was left in a place the customer later disputed. That is both a damage issue and a delivery-record issue. Route it to a human investigation rather than selecting a canned compensation response. The assistant can assemble a chronology: collection scan, hub scan, delivery event, customer report and submitted photographs. Missing times should be marked unknown. A clear chronology reduces repeated requests to the customer, while the claims handler retains the decision.
The team should distinguish a damage acknowledgement from a claim decision in its templates. An acknowledgement confirms receipt of information, the reference and the next evidence request. A decision cites the actual terms and the reviewed record. If AI drafts both from the same short intake, it may collapse that distinction and imply acceptance of responsibility. Give the assistant only the acknowledgement task at first. Once the claims process is stable, a handler may use it to organise an internal decision memo, but the handler must verify every quoted condition and approve the final communication.
The courier damage-claims tutorial covers the separate evidence and dispute workflow in more detail. In the combined shipping desk, the operational question is whether the team has a complete, accurate intake and a polite next response. A fast acknowledgement is useful only if it neither promises an outcome nor loses the evidence trail.
Check the drafts against the systems of record
Run a small pilot on stored, redacted enquiries before using live drafts. Take ten past quote requests with known final prices, ten tracking queries with their event history and five damage reports with the team's recorded next actions. Hide the final response from the assistant. Ask it for a draft and a list of source fields used, then compare the output with the record. Count wrong figures, unsupported promises, missing questions and privacy slips separately. A single wrong price or invented delivery time should trigger correction before the workflow expands.
Keep a simple approval sheet: item ID, source version, AI draft time, reviewer, corrections and sent time. For a quote, verify zone, chargeable weight, surcharge, discount and capacity. For tracking, verify latest scan, timestamp and commitment. For claims, verify that the request matches current terms and that the draft does not decide liability. If the assistant regularly misses one field, change the intake form or rule table rather than asking reviewers to remember the same correction forever.
Measure the effort as well as the errors. For the illustrative 12 enquiries, record how many minutes staff spend preparing, checking and sending each response under the old routine and under the pilot. Do not count only typing time; a draft that takes 30 seconds to generate but four minutes to untangle is no gain. Separate straightforward cases from exceptions so a shift with three unusual claims does not distort the result. If review time drops while correction counts remain low, the workflow may deserve a wider test. If reviewers become less attentive, narrow the use again.
Recheck after a rate-card change. Suppose version 8 increases the zone 3 base charge from the illustrative $14 to $16. Q104 should now calculate to $22 under otherwise identical rules. If the assistant still returns $20, it is using the old reference. Retire the old file from the working set, label the new version clearly and run the saved Q104 test again. The same principle applies when a carrier changes a status label or the firm's claim terms are revised: a draft should show which version it relied upon.
Finally, give dispatchers a safe escape route. If the requested collection is unusual, the tracking feed is missing, or a claim involves contested facts, the assistant should output “human review required” and the specific reason. It is acceptable for the team to answer fewer cases with AI if the remaining answers are reliable. The shipping company gains a consistent first draft and clearer work queues; its own records and people remain the source of price, parcel status and claim decisions.
Further reads
- Can a Small Courier Firm Use AI to Plan Delivery Routes? — Assess route planning after the customer desk is stable.
- How to Market a Courier Business: B2B Outreach Written With AI — Write careful outreach to potential business customers.
- How Can AI Help Courier Services? Quotes, Dispatch and Proof — See the wider link between quoting and dispatch.
- How Does AI Help With Transport? Dispatch and Delay Alerts for 10 Vans — Plan how dispatchers respond when routes change.
- Using AI Services for Customer Support: Courier Parcel-Query Replies — Reply templates for each tracking status, the triggers that pass a message to a person, and a worked example for a six-driver courier.
- Fuel Surcharge Calculator: Set a Weekly Rate for Haulage Quotes — A weekly fuel surcharge policy for haulage and courier quotes: base price, fuel share, step table, spreadsheet formulas and customer clause wording.
- Advance Shipping Notice: Prepare an ASN From Pick Lists With AI — Build a draft ASN from your pick list and carton data, run the mismatch checks against the PO, and send only after the customer's guide is satisfied.
- Delivery Notification Template: Late and Missed Drop Texts With AI — Ready-to-adapt late and missed delivery texts, plus the AI prompt that drafts them from route data without promising times you cannot keep.
- Demurrage and Detention: AI Tracks Free Days and Drafts Disputes — A free-day tracker, an evidence checklist and AI-drafted dispute letters for forwarders who are tired of paying demurrage and detention bills unchallenged.
- Free Shipping Threshold: Set It From Your Average Order Value — A worked order-export example: find your average order value, price the shipping you give away, and test two thresholds with ChatGPT or Claude.
- Shipper's Letter of Instruction: Fill It From the Order With AI — Field-by-field guide to the shipper's letter of instruction, with a prompt that fills it from an order and a forwarder checklist.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.