qAI agent for late shipments
AI agent for late shipments: what the planner approves
An AI agent for late shipments works a delay note through to the customer update. The planner approves the new date, and an unapproved carrier is refused.
3 min readRunning it, Across organisations
by Grisha Pertaia
- session
- credential
- write
- notify
- notify
- file
- write
- score
- notify
A carrier sends a note: the load will be late. From that note to a corrected promise, someone reads the note, finds the order and the customer, works out the new date, writes to the customer and records the freight charge. Each step is in a different system.
An AI agent for late shipments can take those steps. This post answers one question: when a load is late, which steps should the agent take, which should wait for the planner, and which should it be refused?
The workflow is designed. It has not run in the sandbox yet. The steps below are how we build it in a pilot, and the logistics page shows it beside the other logistics workflows, with the state of each.
The promise date is the expensive part
Most of the work on a late load is reading and writing. The step that costs money is the promise to the customer. A date moved without the planner can clash with other orders on the same capacity. A date the plan cannot meet means a second late notice. So the agent drafts the new date, and the planner approves it before the customer sees it.
Two other steps cost money in quieter ways:
- A rebooking with a carrier you have not approved. The load moves, but under terms nobody agreed, with a party nobody checked.
- An instruction inside the carrier's note. A note can carry a line that asks for the delivery address to be changed. A person would query it. An agent that acts on text must not act on it.
The workflow, from the note to the customer
- Step 1
A credential for one shipment
The agent takes the carrier's delay note and receives a credential for this shipment only.
- Step 2
The note read at the door
The note is read before the agent sees it. A line asking for the delivery address to be changed is struck out.
- Step 3
Order and promise found
The agent finds the order, the customer and the promise date in the ERP.
- Step 4
Unapproved carrier refused
The agent tries to rebook the load with a carrier that is not on your approved list. The step is refused.
- Step 5
The planner approves the date
The new promise date waits until the planner approves it.
- Step 6
Customer update drafted
The agent drafts the update to the customer with the approved date.
- Step 7
Freight recorded
The expected freight charge is recorded against the order.
Why the rebooking is refused
The permission for this shipment lists the carriers you approved, and nothing else. A rebooking with any other carrier fails at that list, even when the agent picked the carrier for a good reason, such as an earlier slot. You set the list, and you can read it.
The refusal goes on the record with its reason. If a rebooking is the right answer, a person makes it with a carrier on the list.
Which step stays with the planner
- redacted
Strikes out a hidden instruction in the carrier's note
- allowed
Finds the order, the customer and the promise date
- session
Drafts the new date and the customer update
- allowed
Records the expected freight charge
- refused
Rebooking with a carrier outside your approved list
- human review
Approves or changes the new promise date
- human review
Decides whether to rebook, and with whom
- human review
Answers for the promise to the customer
The planner gets one decision with its context: the order, the old date, the proposed date and the reason in the carrier's note. Approve and reject both go on the record, with the planner's name.
The same rule at the shipping desk
The order to shipment workflow works the outbound side, and it uses the same two controls. A split shipment the customer has not agreed to waits for the planner. A customer email that asks for delivery to a new address does not change the ship-to address. The label is printed only to the address on file, and the request goes to the order desk.
In both workflows an instruction in an email is information for a person, not an order for the agent.
When the work crosses to another company
A late load involves at least two companies: yours and the carrier. Often there is a third, the customer. When an agent's work crosses to a supplier or a customer, the permission narrows at the company boundary, and each side keeps a record of what crossed. The sandbox run between a buyer and a supplier shows this on a quote, and the page on work across companies explains it.
What you can check
You or your auditor can check the record with the open verifier, without us. The guarantees page shows how the permission for one job is set and where it stops: it is only as good as the rules you approve.
What it does not do
- It does not move a customer promise, issue a credit or book freight outside your approved list without a person.
- It does not replace your transport, warehouse or ERP systems. It works between them.
- Inspection covers stated patterns and languages. An instruction outside them can reach the agent, and the permission still limits what it can do.
- It has not run in the sandbox. The closest run today is supplier follow-up: a late inbound part and a date change the buyer approves.
Related notes
- illustration
- session
- credential
- notify
- write
- sign
- draft · refused
5 min read
Narrowing across organisations: why bilateral agent integrations do not scale
Every cross-company agent integration today is a bespoke agreement: a shared key, a contract, a hope. What scales is a permission that narrows at the company boundary and a record both sides hold.
- illustration
- session
- credential
- write
- intake
- route · human review
- write
3 min read
AI agent for utility work orders: who approves each step
An AI agent for utility work orders carries an inspection finding to a closed job. The planner approves spend and the crew lead signs the field ticket.