Last week I sat in a customer review meeting and watched a grown man get giddy. His word, not mine. 😄
We had just shown him something that took our developer Jacob a few weeks to build and takes about ninety seconds to demonstrate: a normal, off-the-shelf Claude chat window, asking a PIM questions and getting real product data back. "What fire-rated downlighters do we have?" A list. "Tell me more about the second one." The full attribute set, with a link to the size diagram it had found on the way.
Nothing exploded. No one had to export a spreadsheet. And the customer immediately started inventing uses for it that we had not thought of. That is usually the sign you have built something worthwhile, so I thought I would write up what it is, why we built it this way, and why I think every PIM is going to need one.
The thing we deliberately did not build 🚫
The request that started all this was simple enough. A customer of ours, a UK gaming and bedroom furniture brand, uses an AI customer-service platform to answer questions on their website. They wanted it to know things like the weight limit on a particular bunk bed without someone having to keep feeding it spreadsheets. Could we hook it up to their PIM?
The obvious answer is yes, we'll build an integration to that bot. The better answer, which took a day or two to see, is no, we won't integrate with that bot. We'll make the PIM something every bot can plug into.
That is what an MCP server is. MCP stands for Model Context Protocol, and if you want a two-word explanation it is "USB-C for AI". Any AI that speaks the protocol, and all the big ones now do, can plug into a system that exposes one, discover what tools are available, and use them. Claude, ChatGPT, Copilot, a customer-service bot, a workflow tool nobody has invented yet. One doorway, many visitors.
The alternative is the one the industry has been pursuing for two years: every software vendor builds their own little assistant, trains it on their own little slice of the world, and asks you to open yet another chat window. I have written before that the portal is becoming a conversation. This is the next step: the conversation should happen wherever the customer already is. They have a ChatGPT business account. They are investing in Copilot. Their support bot runs on Claude underneath. Why would I want them to come to mine?
Bring your own brain. We'll bring the data. 🤝
How it actually works (the unglamorous bit)
Here is the part that matters if you are responsible for commercially sensitive product data, which is everyone reading this.
The MCP connection is a user. It sits in a team, and the team has permissions, exactly like a human member of staff. There is a new section in the permissions screen for AI connections, and you tick what the connection is allowed to do: search products, get product details, list channels, list families, list attributes, and so on.
For the customer-service bot, our customer will tick two boxes: search and get details. It does not need to know what families exist. It does not need to see the attribute list. If someone tries to get cheeky and asks the bot what the cost price is, the bot cannot answer, because the connection cannot see it.
For the product manager who wants to use ChatGPT to interrogate the whole catalogue, you tick everything.
Then there is the channel. A PIM already has the concept of a channel: a curated view of which products, and which attributes of those products, are visible for a given purpose. The MCP connection is pointed at one or more channels and simply inherits them. Our customer already has a "portal users" channel for external partners; they can point the bot at that and they are done. No new data model, no new permission system, no new place for a mistake to hide.
Finally, the plumbing. An automated system like a bot gets an API key tied to that team, sent as a header. A person connecting from their own Claude or ChatGPT gets a browser prompt and logs in as themselves. Same doorway, different keys. 🔑
I am labouring this because it is the bit that vendors skip in the demo and customers rightly worry about. "Give the AI access to the PIM" sounds terrifying until you realise it means "create a user with the right permissions on the right channel", which is a thing your team has done a hundred times.
What happened when the customer started thinking 💡
This is where it got interesting, and where I want to give the customer the credit.
Within five minutes of seeing a search result, their product manager had a use case I had not considered. They sell through twenty-odd retailers and marketplaces. The SKU exists on all of them. But the richness of the listing, the dimensions, the weight, whether the right image got uploaded, varies wildly, and some of the errors are introduced by the retailer's own operations team after the hideous Excel template has been submitted. Checking this is a famous chore. Nobody does it often enough.
"Could I say: find this product in the PIM and compare it against what's on Argos?"
Yes. Because the AI doing the asking is Claude or ChatGPT, not a PIM-shaped chatbot, it can do both halves of the job: pull the authoritative data through the MCP server, then go and look at the retailer's page, then write you a report of what disagrees. Next obvious step, which he got to in about four seconds: a weekly agent that emails the buyers with what needs fixing, so the product team never has to.
Then came the one that I think is the real story. He said, half joking, that he is the company's PIM bot. He is the person everyone asks. His colleague, he said, is their ChatGPT. What he wants is to clone himself: to give the whole company the ability to ask product questions at any hour and get the answer he would have given.
That is not a feature request. That is a description of what a PIM was always supposed to be, finally reachable without a login and a training session. 🎯
"Won't it just make things up?"
Fair question, and he asked it. He had built a local agent over their customer-service data the week before, and his honest worry was that he could not verify its answers because he had not personally handled every ticket.
With PIM data the situation is reversed. He does know the products. He can run the first twenty queries, check every answer against what he knows, and only then let it loose on colleagues. And the more fundamental point: models hallucinate when they lack data. Give one the best-quality product data in the company and it has nothing to invent. The models themselves have also got noticeably better at saying "I don't know" over the past year. Good data plus a model that admits ignorance is a very different proposition from a chatbot winging it over a PDF.
The question I was waiting for
"Pat, aren't you worried this makes all your import and export work redundant?"
Not at all, and this is the design principle I care most about. When the product manager asks Claude for a weekly report, Claude does not go and run a report. It creates an export task, or writes a Python agent task, inside the PIM, using the same mechanism a human would. It appears in the task list. He opens it, sees what it is going to do, and decides whether to run it. The decision stays with the human, and every run is logged the same way it would be if he had built it by hand. 📋
What you do not want is an AI that can change ten thousand records in six seconds because somebody phrased a prompt badly. The way to prevent that is not to limit the AI's intelligence; it is to make sure it only ever acts through the doors humans already use, where everything is recorded. No privileged path. (That principle deserves a post of its own, and it will get one.)
Wow, three times a day
I tell my staff that I say "wow" at least three times a day, and I am immersed in this stuff. They look at me like I need a holiday. But it is true: the limiting factor has stopped being the technology and become our imagination. What should I ask it to do? What should I ask it to build?
So here is my prediction. Within a year, "does it have an MCP server?" will be on every PIM evaluation checklist, right next to "does it integrate with Shopify". The vendors who built their own walled-garden assistants will find customers politely ignoring them in favour of the AI they already pay for. And the PIM, the system that was always meant to be the single source of truth but was usually the single source of truth that only three people knew how to open, finally becomes the thing every brain in the business can talk to.
Ours goes live for that customer this week. I will report back on what they ask it. I suspect it will be things I have not thought of. 🚀
Pat Violaris Managing Director, OneTimePIM
Pat has been building AI systems since his PhD in expert systems in 1988, and still says "wow" three times a day.
Want to connect your own Claude, ChatGPT or customer-service platform to your product data? Book a 15-minute intro call.