MENU

We built a supplier portal instead of an AI agent

A brand came to us last year running 300 to 500 samples a month. The brief was PLM.

When we got in and looked at how the team actually worked, every sample request was an email. Every round of fit feedback was an email. Every approval, every chase, every revised date.

There was no system underneath any of it. There was an inbox, and there were people who were good at reading it.

The obvious move in 2026 is to point an agent at that inbox. Pull out the style, the round, the comment, the date. Write it into a record. We could have built that in a fortnight and it would have worked.

We didn't build it, and it took me a while to be able to say why without sounding like I was talking myself out of revenue.

"You can spend the next 5 years extracting the data, or 6 weeks changing where it enters."

Oliver Rhodes, CEO, NOLO

The parser never leaves

An agent reading supplier email is a translation layer, and translation layers are permanent.

Every field it produces is an inference about what somebody meant. Which of the 3 dates in the paragraph is the ex-factory date. Whether "shipping Friday" means leaving the factory or landing at the 3PL. Whether the quantity in the body overrides the quantity in the attachment.

It gets those right most of the time. The trouble is the other times, and the fact that you can't see which is which without going back to the email.

Then the factory hires someone new who writes differently, and accuracy drops with nothing in the system to tell you it has.

The email is still the source of truth in that setup. Your record is a copy of somebody's prose, and you've taken on a dependency that needs maintaining for as long as you trade with that supplier.

You already have an extraction layer

We ran 3 discovery workshops at a fashion brand in January, 4 years into working with them.

One was with the person who handles inbound logistics. He processes carrier email notifications by hand, keeps an Excel shipment log past 5,000 lines, and types pre-advice documents into the 3PL portal one at a time.

Somebody asked what happens if he's off sick. The room went quiet.

He's the parser. He's been doing the job an agent would do, for years, accurately, and nobody on the outside can see what's on the water without asking him.

The only real question is who does it, and what the business does when they leave.

What changes when the supplier types it

3 things, and the obvious one matters least.

The retyping goes. That's hours back for the ops team, and it's the line that ends up in the business case.

The data arrives structured, because the supplier filled a field instead of writing a sentence. Nobody has to work out what was meant.

And the supplier's version and your version are the same row. There was never a second copy, so there's no reconciliation and no afternoon spent deciding which one is right.

That third one is the whole reason to do it.

The portal that writes nowhere

There's a version of this that makes things worse, and plenty of brands have built it.

A login, some forms, and a separate database nothing downstream can read. The supplier dutifully enters data into a system that doesn't feed the product master, so ops export it and retype it anyway.

That's a silo with better manners. It also makes the AI problem harder, because now the mess has a user interface and looks like it's been solved.

A portal only counts if it's a view of the same record everything else reads from. Same product, same PO, same permissions, same audit trail. The supplier sees a narrow slice and can write to a narrower one.

What we actually build

The portals we've put live do 3 jobs.

A factory logs in and handles its own samples: requests, rounds, comments, approvals, against the style record rather than beside it.

It turns a purchase order into an inbound shipment, with real dates and real quantities against the PO it's actually shipping.

And it uploads its invoice against that same PO, so finance is matching against a record rather than a PDF and a memory.

All 3 write to the product master. None of them create a second copy.

The cost is not where people expect it

A supplier seat on a system you own runs at single digits a month. A supplier licence on an enterprise PLM does not, which is why most brands buy the PLM and then never roll it out past their own team.

The build is a known quantity too. It's a defined piece of work with a number against it.

Set that against the ops salary currently spent on retyping, and the key-person risk sitting underneath it, and it stops being a technology decision.

Then the AI conversation gets interesting

Once the data is arriving structured and at source, an agent has something worth reading.

Supplier performance across a season. Which factories slip and by how much. Where the critical path is going to break 3 weeks before it does. We're running that kind of analysis on client data now, and it works because the data underneath it was captured properly rather than reconstructed afterwards.

If you're looking at your supply chain and wondering where AI fits, the honest first question is where your data comes from. If the answer is an inbox, that's the thing to fix. The rest gets much easier afterwards.

What to take from this

01
An agent reading supplier emails is a permanent translation layer. Every field it produces is an inference you can't audit without going back to the email.
02
You already have an extraction layer. It's usually one person, and the business finds out how much it depended on them when they leave.
03
A supplier portal only counts if it writes to the same record everything else reads. Anything else is a silo with a login.