Logistics & Supply Chain
By Sagar Shankaran, Founder of CallSphere
Release authority, signature rules, medical runs: how grounded AI answers with document citations stop courier dispatchers guessing account rules at 7 a.m.
Key takeaways
It is 7:12 on a Tuesday. Your dispatcher has a driver holding on line two from a dock behind a strip-mall dental office, the warehouse coordinator for your biggest retail account on line one, and forty-one routes still to release in Onfleet. The driver's question sounds simple: nobody is answering, there is a covered side porch, can he drop it, shoot the photo and move on?
The honest answer is that it depends entirely on whose parcel it is. The regional wine club account requires an adult signature with an ID check, and a driver release there is a violation of the shipper's own agreement with the state alcohol board. The orthopedic supply house allows driver release but forbids garage placement. The hospital lab work has no release option at all, ever, and the specimen bag on the seat next to him is a UN3373 Category B shipment that has been out of the 2-8°C cooler for nineteen minutes. Your dispatcher has three seconds and fifty-eight account profiles to hold in her head.
So she guesses. Most days she guesses right, because she has been doing this for six years. The days she guesses wrong show up at 6 p.m. as a service-failure email, a concession charge, and a phone call from an account manager who wants to know why a signature-required parcel is sitting on a porch in a Ring camera clip.
Every courier operation has the same artifact: a folder of account SOPs. It came out of onboarding. Each shipper sent over a delivery instructions matrix — a PDF, or a spreadsheet somebody exported to PDF, or a two-page addendum bolted onto the service agreement. Release authority. Signature requirements. Attempt policy. Redelivery window. What to do with a refused parcel. Which cross streets are out of the covered zone. Whether the driver photographs the label or the door.
Nobody reads them after week two. The rules live in the heads of the dispatcher and the on-road supervisor, and they leak. A dispatcher takes a Friday off, the backup runs the desk, and the wine club account gets four driver releases in one afternoon. That is the whole failure mode. It is not laziness and it is not a training problem you can fix with another meeting; it is fifty-eight documents and a decision that has to be made in under ten seconds while the driver's engine is running.
Grounded answering means the assistant is allowed to answer only from the documents you approved — your account SOPs, your driver handbook, your tariff — and it shows you the file and the page it pulled the answer from, so a wrong answer is catchable instead of merely plausible. That last clause is the entire point. A 2024 chatbot would happily invent a signature rule that sounded exactly like a real signature rule.
Hear it before you finish reading
Talk to a live CallSphere AI voice agent for logistics in your browser — 60 seconds, no signup.
By early 2026 the retrieval-and-citation features in the general business assistants — Claude Cowork since January, ChatGPT Work since 9 July — stopped being an add-on and became the default way these products answer over a folder of your own files. You point the assistant at the account SOP folder. It answers from those documents. When it answers, it names the document and the section, and you can click through and read the sentence it used.
The second half matters more than the first: when the answer is not in your documents, a properly configured grounded assistant says it is not in your documents. It does not fill the gap with a reasonable-sounding invention. For a courier, "the SOP does not cover side-porch placement for this account" is a genuinely useful answer, because it tells the route manager to call the shipper and get it in writing before the next attempt.
flowchart TD
A["Driver at the door, 10:40am, no one home"] --> B["Dispatcher asks: can I release for account 4412?"]
B --> C["Assistant reads only the approved account SOP folder"]
C --> D{"Does that account's SOP allow driver release?"}
D -->|"No, adult signature 21+"| E["Answer plus link to SOP page 3"]
D -->|"Not covered in any SOP"| F["Assistant says it does not know"]
F --> G["Route manager calls the shipper, gets it in writing"]
G --> C
E --> H["Driver reattempts or returns it, noted on the manifest"]
Here is the same moment with grounding in place. The driver messages dispatch through the driver app. The dispatcher types one line: "4412, no answer, covered side porch, can he release?" The assistant comes back in a couple of seconds: no driver release; adult signature 21 or older; ID scan required in the driver app; if no signer, one reattempt next business day, then return to shipper — and underneath, the file name and the page.
The dispatcher reads the answer, not the binder. She tells the driver to reattempt tomorrow and codes the stop as attempted-no-signer, which is the code the shipper's own report keys on. The parcel goes back on the van, the stop lands on tomorrow's manifest, and there is no 6 p.m. email.
Same tool, harder question, later in the day: an account's coordinator asks whether their refrigerated specimen boxes can ride with ambient parcels on a shared route. The assistant finds the answer in the lab's courier addendum — segregated compartment, temperature log per leg, chain-of-custody form signed at both ends — and cites it. That is not a question anyone should be answering from memory at 4 p.m.
Numbers below are an illustration for a mid-size independent courier, not a survey result. Substitute your own from last quarter's exception report.
| Rules questions reaching the dispatch desk each week | 120 |
| Share answered from memory rather than the SOP | 35% = 42 |
| Of those, answered wrong | 8% = 3.4 |
| Share that turns into a paid failure (redelivery, concession, claim) | 1 in 3 = 1.1 per week |
| Average cost of one paid failure | $240 |
| Weekly cost of guessing | $264, about $13,700 a year |
| Dispatcher time on those 120 questions at 4 minutes each | 8 hours a week |
Grounded lookup does not take the 120 questions away. It takes away most of the guessing: the dispatcher stops answering from memory because looking it up is now faster than remembering. If wrong answers drop from 3.4 a week to under one, that is roughly $9,000 a year, plus three or four hours a week back on the desk during the morning release window — which is the only hour that actually constrains your day.
How you would prove it: pull the last two quarters of concessions and claims, tag each one by root cause, and count the ones where the root cause was "we applied the wrong account rule." That count is your baseline. Re-run it ninety days after you turn grounding on.
Still reading? Stop comparing — try CallSphere live.
See the logistics AI agent handle a real call — complete, industry-specific, and live in your browser. No signup.
Grounding only makes the assistant faithful to your documents. It does not make your documents right. If the wine club SOP in your folder is the 2023 version and the shipper tightened the ID rule last spring, the assistant will cite the wrong rule confidently and show you exactly which stale page it came from. Somebody — usually the account manager — has to own document freshness, with a review date on every SOP.
Second, do not let it answer to the driver directly on anything involving controlled substances, temperature-sensitive medical shipments, or a refusal at the door. Those go to a person. A grounded answer that says "no release" is fine; a grounded answer that decides whether a specimen is still viable is not, and no citation makes that safe.
Third, judgment calls stay with the on-road supervisor: the consignee who is standing there angry, the address that is clearly wrong, the apartment complex that changed its package policy without telling anyone. The SOP folder has no page for those, and the correct behavior is for the assistant to say so.
No, and you should not try. Point it at what you have, in whatever shape it is in — PDFs, the addendum pages, the spreadsheet. Grounding will expose the gaps for you within a week, because the answer "not covered in the documents" keeps landing on the same three accounts. Fix those three.
Nothing automatic, which is why the citation is the safety feature and not a nicety. Your dispatcher's job becomes a two-second glance: does the cited document name match the account on the stop? If it does not, the answer is wrong and she opens the right file. Train that glance on day one.
Eventually, and some operations already do it as a read-only lookup for release authority and attempt policy. Start with dispatch only. Drivers asking questions at the door creates pressure to accept whatever comes back, and you want a person in the loop while you are still finding stale documents.
It does not. Keep the hard rules — signature required, no release, time window — enforced in DispatchTrack, Track-POD, Onfleet or whatever you run, because those block the driver app from completing the stop the wrong way. Grounding covers the long tail of written rules that never fit into a checkbox.
Pick your five highest-volume accounts. Put their current SOPs, addenda and any email threads that amended them into one folder, name the files after the accounts, and give a grounded assistant read access to that folder and nothing else. Tell your dispatcher to ask it every rules question for one week and to forward every "not in the documents" answer to the account manager. You will end the week with five clean SOPs and a list of holes you did not know you had.
The other half of this problem arrives by phone: the consignee who calls at 6:40 p.m. to ask whether the driver can leave it, and gets voicemail. CallSphere builds AI voice and chat agents that answer the line after hours, handle the delivery-status and reschedule calls, and pass the rest to your dispatcher with the account and stop already identified — working from the rules you approved rather than from whatever sounds reasonable.
Written by
Sagar Shankaran· Founder, CallSphere
LinkedInSagar Shankaran is the founder of CallSphere, where he builds production AI voice and chat agents deployed across healthcare, hospitality, real estate, and home services. He writes about agentic AI, LLM engineering, and shipping voice agents that handle real calls in production.
See how AI voice agents work for your industry. Live demo available -- no signup required.
Sort courier calls before answering: cheap model for status and reschedules, strong model for damage claims and missed medical runs. Costs, rules and limits.
Carrier shortlists are built by software before dawn. What makes a 38-truck fleet invisible: stale MCS-150, no lane page, and a phone that stops at 6pm.
Doxycycline before IPL, fish oil before filler, blanching after lip work. In 2026 an assistant answers only from your approved sheets and names the source.
How general contractors keep an AI assistant inside the conformed set, ASIs and approved submittals, with citations a super can check before inspection.
AI that answers only from a firm's approved memos and the client's file, showing the page — so a wrong travel answer is caught before it costs a decade.
Live load, dry run, final pull, #11 OCC, VSQG. Why general AI misbooks roll-off orders, what trade-tuned models fix, and the truck time a misread order burns.
© 2026 CallSphere Inc. All rights reserved.
Made within San Francisco
Watch how CallSphere handles real customer calls, schedules appointments, and processes payments — live.
Try Live DemoBook a DemoCalculate Your ROI