Tender intake, carrier file hygiene, status updates and invoice matching are the four places the arithmetic works. Everything else is a demo.
Key takeaways
AI genuinely speeds up the parts of a brokerage that are reading, matching and repeating: pulling structured fields out of an inbound tender, checking that a carrier’s file is still current, chasing a status update, and reconciling an invoice against a proof of delivery. Those four consume most of a coordinator’s day and none of them requires a judgment call with legal consequences.
It does not speed up the things brokers usually ask about first — pricing a lane, deciding whether a carrier is trustworthy, or committing to a shipper. Those are decisions, and a decision needs a person who can be held to it.
Most brokerages receive load information as unstructured text: an email, a PDF rate sheet, a spreadsheet attachment, a portal notification. The clerical task is the same every time — find the origin, destination, commodity, weight, dates, references and accessorials, and put them in the fields your system expects.
Extraction handles this well because the answer is on the page. What it cannot do is notice that a shipper has quietly changed a term. Build the rule that flags a difference from the last tender by the same customer, and you convert a reading task into an exception queue.
The expensive failures in brokerage almost always trace to a stale file. A certificate that expired, an authority that lapsed, a safety rating that moved. This is the highest-return automation in the business because it is pure date arithmetic against documents you already hold.
In Ontario, a carrier’s safety record is a public artefact: the province publishes how to obtain a commercial vehicle operator’s registration abstract or carrier record. A rule that re-pulls on a schedule, rather than at onboarding only, is worth more than any scoring model.
It also matters because a carrier’s insurance obligation is real and enforceable. In British Columbia a for-hire carrier must secure and maintain cargo insurance and produce proof of it on request. The certificate in your file is only a snapshot of that.
Check calls are the most automatable work in a brokerage and the most regulated. Two regimes bite.
If you send commercial electronic messages — a marketing email to a shipper, a promotional SMS — Canada’s anti-spam legislation applies, and the CRTC is one of the bodies that enforces it. Whether you are relying on implied or express consent changes what you must be able to prove.
If you use a pre-recorded or synthesised voice to place calls, you are operating what the CRTC calls an automatic dialing-announcing device. The definition is explicit: an ADAD is equipment capable of storing or producing telephone numbers that is used to convey a pre-recorded or synthesized voice message. The Unsolicited Telecommunications Rules then impose conditions even where there is no attempt to solicit — such calls are restricted to 9:00 a.m. to 9:30 p.m. on weekdays and 10:00 a.m. to 6:00 p.m. on weekends, in the recipient’s local time, must open with a message identifying who is calling and why, with a valid email or postal address and a local or toll-free number, must display an originating number, and the equipment must disconnect within ten seconds of the recipient hanging up.
An AI voice agent calling a receiver to confirm an appointment is squarely inside that. It is entirely doable — it is just not unregulated.
Invoice, rate confirmation, proof of delivery, accessorial claim. Four documents, one number that has to agree. Automating the match and surfacing only the disagreements is the clearest saving in the back office, and it compounds because disputes get raised while the shipper still remembers the load.
Cargo crime in Canada has shifted from breaking into trailers to defeating paperwork. Équité Association, which maintains the national cargo theft database, describes three patterns brokers meet directly: a fictitious pick-up, where criminals learn a legitimate carrier’s pick-up details and collect the load first using forged paperwork and replicated logos; double brokering, where a fictitious or stolen carrier identity is used to win the load and then re-broker it to an unwitting legitimate carrier; and alteration of the bill of lading, where part of a load is removed and the paperwork is recreated so the receiver signs for the reduced quantity, with the shortage surfacing weeks later during payment reconciliation.
Every one of those is a document that looks right. That is precisely what an extraction model is bad at catching and a cross-reference rule is good at catching: same carrier name, different banking details; same MC identity, new phone number; delivered quantity below tendered quantity. Automate the comparison, not the conclusion.
The line is not subtle. AI drafts, extracts, schedules and summarises; a person decides and signs. That means a human approves a new carrier, a human commits a rate, a human confirms a payment instruction change, and a human signs anything that becomes evidence. The Office of the Privacy Commissioner’s principles for responsible, trustworthy and privacy-protective generative AI make the same point from the privacy side: accountability sits with the organisation deploying the tool.
Driver names, licence details, abstracts and phone numbers are personal information. If your brokerage handles it in the course of commercial activity, PIPEDA sets out the obligations in brief and the question of whether it applies to your business applies. Anything you automate should collect less, not more.
On retention, the baseline for the commercial documents is the Income Tax Act: every person carrying on business must keep records and books of account at their place of business or residence in Canada. If your automation is the only place a rate confirmation lives, that system is now a records system.
Worked example: a two-desk brokerage in Mississauga
Two people, roughly 45 loads a week, all domestic. Before: each load takes about eleven touches — read the tender, key it, find a carrier, send the confirmation, three check calls, collect the POD, match the invoice, chase the accessorial.
They automate three of the eleven. Tender emails are parsed into the system with a flag when a field differs from that shipper’s last load. Carrier certificates and safety records are re-checked on a schedule, and a load cannot be covered by a carrier whose file is expired. Invoices are matched to POD and rate confirmation, and only mismatches reach a person.
They deliberately do not automate two things: the first call with a new carrier, and any change to payment details. Both are the exact points the fraud patterns target.
The measurable change is not a headcount saving. It is that the exception list is now short enough to work through before end of day, which is what actually stops a shortage being discovered a month later.
It can rank and it can filter on hard rules — expired file, wrong equipment, outside the lane. It should not make the commitment. The consequences of a bad carrier choice are contractual and, in a fraud case, evidential.
A transactional message about a load the driver has already accepted is different from marketing. The distinction that matters is whether the message is a commercial electronic message, which is why the CRTC’s anti-spam material is worth reading before you build the messaging flow.
Yes, subject to the ADAD conditions above — hours, identification message, number display, ten-second disconnect, and no sequential dialling. Those are design requirements you can build once.
Certificate and safety-record expiry. It is deterministic, it needs no model, and it removes the failure mode that costs the most.
A 30-minute call is enough to tell you whether AI pays for itself in your back office.
