Logistics doesn't need an AI TMS

There's a version of the AI logistics story that every large vendor is currently telling.

It goes: we have the system of record, the system of record has the data, therefore we'll put agents on top of the system of record. The AI TMS. The AI forwarding platform. The intelligent supply chain suite.

It's a reasonable story. It's aimed at the wrong problem.

Look at where the hours go

Walk a forwarding operation and count what people actually spend the day doing.

Reading emails and deciding what they are. Opening attachments and retyping what's in them into a form. Chasing a carrier for a status that hasn't updated since Tuesday. Answering "where is my shipment" for the fourth time this week, for the same shipment. Working out why two systems disagree about the same consignment, and which of them to believe. Deciding whether a commodity description on an invoice line is enough to classify on, and if not, emailing the shipper to ask.

Almost none of that happens inside the TMS.

The TMS is where the answer gets recorded once somebody has worked it out. The working-out happens in Outlook, in PDFs, on WhatsApp, on the phone, and in a spreadsheet that one person maintains privately because the system can't hold the thing they need to keep track of.

An agent layer inside the system of record automates the recording. The recording was never the expensive part. Data entry into a well-designed form is maybe fifteen percent of the effort on a difficult shipment. The other eighty-five percent is finding out what to type.

There's a decent proxy for this if you want to measure it in your own operation. Count how many of your ops team's working hours are spent in the TMS versus in a mail client. In every operation I've looked at, the mail client wins comfortably, and nobody has ever been surprised by that when I've said it out loud.

Why incumbents build it that way anyway

Not stupidity. Structure.

If your product is the system of record, everything you build has to strengthen the case for the system of record, because the system of record is the renewal. An AI capability that worked equally well for a customer running a competitor's TMS would be commercially incoherent to build, however useful it might be. You'd be funding your own displacement.

So the agents get built where the moat is rather than where the pain is. Entirely rational as a strategy, and it leaves the actual mess untouched.

There's a second structural reason, which is data access. The messy work involves the mailbox, the attachments, the carrier portals and the vendor's own reference data. A platform vendor generally has none of those. Their agent can reason beautifully over the fields already in the database, which are precisely the fields somebody already keyed in, which is the point in the process where the work was already finished.

The other shape

Start from a different assumption: the system of record stays. It isn't going anywhere, replacing it is a multi-year project nobody has the appetite for, and the good ones do their job well.

Then ask what happens if the messy layer around it gets handled.

Inbound mail classified and routed before a human opens it, so a rate request goes to the desk that quotes and a customs query goes to the desk that clears. Attachments read into structured fields with the source document retained next to the extraction, so the checking is a glance rather than a re-read. Carrier events normalised into one vocabulary regardless of how eleven different carriers describe the same physical thing. Commodity descriptions classified with the rule and the ruling attached, so the reasoning survives the departure of the person who did it.

None of that requires ripping anything out. All of it is work that currently has no home in any system, which is why it's done by people, in an inbox, at whatever quality that day allows.

That's an overlay, not a replacement. A less exciting category to write about. Considerably easier to buy, because it doesn't require anyone to bet their operation on a migration, and because you can put it next to the existing system and compare.

It's also easier to stop using, which is a feature from the buyer's side and the reason vendors don't like building it.

The test we'd apply

If somebody is selling you AI for a logistics operation, ask one question. Does this reduce the number of times a human retypes something that already existed in a document?

If the answer is yes, it's touching the real cost, and you can size the benefit yourself by counting.

If the honest answer is "it makes our existing screens smarter" or "you can ask questions of your data in natural language," it's touching the part that was already automated. Structured data in a database was solved in the 1990s. It wasn't the bottleneck then and it isn't now. A better query interface over data that was already clean and already keyed in is a nicer version of a solved problem.

A useful follow-up: where does this system get its input from? If the answer is "your TMS," you're automating the output of the work rather than the work.

What we're not claiming

Not that the system of record is unimportant, and not that incumbent platforms are bad software. They're not, and the good ones are very good at the thing they were built for, which is holding an accurate, auditable record of a complicated commercial transaction across many parties. That is a genuinely hard problem and it's been solved to a high standard.

Only that the remaining hard part of logistics software isn't inside the platform.

It's in the inbox.