#135 - AI with guardrails: how to bring agentic AI safely into commerce

Joubin Rahimi

Wonderful to have you back for a new episode of insights! My name is Joubin, Joubin Rahimi, and joining us today, Nils Breitmann from Intershop. Hi Nils.

Nils Breitmann

Hi Joubin. Glad to be here. I'm Nils from Intershop. At Intershop I'm responsible for AI, so AI as we deploy it in our products and AI as we use it internally. And for anyone who might not know, Intershop provides a commerce platform for agentic B2B commerce.

Joubin Rahimi

And I think it's brilliant that you came all the way from Jena so we could record this. It's always much nicer for everyone, and the conversations are hugely valuable. We already had a talk that touched on AI and co-pilots and how to work with them on the buyer side. There we were more in a search-and-find mode. But it goes much, much further than that. So how can agentic AI also support sales? What's your view, at a really high level, on the topic of agentic AI abolishing salespeople's work or not?

Nils Breitmann

The co-pilots are agents, and of course they'll take over work that used to be done by people. That's clear. But that will also enable companies to do things they couldn't do before, because they simply didn't have the teams for it.

Joubin Rahimi

Do you have an example?

Nils Breitmann

For example, we see very many customers, particularly wholesalers, with poor product data, which has been the case for years, because they simply don't have the capacity now to enrich all these products – in places this runs into millions of products they have – one by one with some content editor. And that's where our Product Content Agent comes into play, for example, which does this in a fully automated way, going out onto the internet, downloading manufacturer data sheets and then creating attractive descriptions and attributes.

Joubin Rahimi

How do you ensure the data is correct there?

Nils Breitmann

That's natural too …

Joubin Rahimi

That's the first question customers always ask: my employees don't make mistakes, but the AI, I don't want it.

Nils Breitmann

Naturally you look at what came out first. At start you always have kind of process where you say, you give a prompt. Nice thing is, you can say, which target group should this product description be tailored for? We've got medical device manufacturer who among other things has range of office supplies. Then you can prompt it, do it in way that's now well suited for a doctor or a medical care centre. And I read it through, there's some kind of paper shredder, and I think: boring product, but suddenly it's phrased how this creates value in this medical care centre, sounds totally different. Then you think: oh yes, that's a paper shredder for me. That's possible too.

Joubin Rahimi

When patient data has to be shredded so no one sees it, and otherwise it always sits on top of the counter until it ends up in the bin.

Nils Breitmann

That means you start with a process where you say you adjust it roughly first, see how it looks. And then we also have the agent itself, the other one, check it again to see whether it's actually correct.

Joubin Rahimi

So you have a second agent there that checks the other one?

Nils Breitmann

Right.

Joubin Rahimi

Okay, yes.

Nils Breitmann

Clever. That also finds flaws again. Of course it's best to then somehow use a different model, maybe a bit … Sometimes it's already enough to say, just run ChatGPT Pro over it now and take a look. You don't have to do much at all. That's the way to go then, because the biggest fear now is naturally: is this data correct? Is there now somehow something wrong in it? We have a customer, an electrical wholesaler, who simply played it completely openly and then in his shop there's a tab with the standard product description, four bullet points or so, and then next to it there's an extra tab with the AI fairy dust stars, and then you see, there's now the AI description attributes.

Joubin Rahimi

Have you been able to run analyses on how that behaves?

Nils Breitmann

I don't have it to hand right now, but I think it definitely looked significantly better, what was in there then.

Joubin Rahimi

So what's definitely 100 percent better is topic-specific SEO. But if, in that context, or then also geo in that context, this content is included, such as with destroying patient records and not just files.

Nils Breitmann

Yes, that's also this GEO topic, Generative Engine Optimization, so it shows up well in ChatGPT and Perplexity, you should really approach it as a problem. You should do the problem description, the solution... Because that's a thing that works well there, because behind it there are always these vector embeddings, this semantic search, and it really recognises which problem is at hand and then finds the right product based on that problem, which the customer may have phrased in ChatGPT.

Joubin Rahimi

I have another question on my own behalf, and for everyone else too. We have quite a lot of locations and keep merging locations together. And especially in Cologne we're merging locations. And then we face the challenge that all sorts of different pieces of furniture end up together. It's all a bit dull. But then we say: okay, we want to set it up again so it feels homely and not like a shared flat where everything's been thrown together randomly. And then we have a floor plan. We know where which tables stand. We know how many people are there, whether it's a hot-desking workspace or not. Could I, purely hypothetically, purely in principle, give such a floor plan to a co-pilot from the office fitter – you have one or two of those too –, who then says: okay, let me just make you a suggestion for which products might fit, also based on budget, based on: what have you already bought from me?, and so on. Is that something worth considering?

Nils Breitmann

Well, theoretically yes of course, because this whole architecture is super extensible. It's this agent-and-tool mechanism. You have the agent, the large language model, and you hand it various tools so it can also do things. And I even had a nice talk about MCP last week, there was an AI bar camp in Jena. So MCP is like USB-C for the agent era, as people like to say. That means various programs or solutions can offer their capabilities as an MCP tool. That's particularly well suited so agents can use it too. And he had this example, there's this Blender 3D tool where you can make models. I don't know if you know it. No, don't know it. You break your fingers with that one too. But he'd connected it, so to speak, via the MCP tool and then did a prompt, wrote in it now: model me the Jentower from Jena here, that big tower in Jena.

Joubin Rahimi

Our tower?

Nils Breitmann

Yes, that was once the case too. We've now moved. And then the large language model does this, first checks its knowledge base, or perhaps it can also, let's say, use search engines in the background, then knows, right, a big cylinder is good, with a bit of, how, and then found the tool, this Blender tool, and modelled it in there, and in this Blender you could really see the 3D model of this tower. And that actually also shows how powerful this is and how well it can be extended. And, let's say, your furniture topic could also be worked in there. Though I've of course also often done it myself, because you take ChatGPT, take a photo and put the wallpaper in there or something. My wife always likes doing that.

Joubin Rahimi

My expectation was also that you'd say: Sure, that'll work out somehow. We also have a client who built it together with you in that environment. We'll be in touch. I think it's really cool when you can say: Okay, this is my challenge. I throw it in and don't have to worry about it, whether it's the buyer or a seller who then handles it, who might not even need to read it out anymore.

Nils Breitmann

But with these tools … We have a customer, Laminatdepot, and when I buy laminate flooring, it really matters that I calculate it so that all the pieces fit, so that there isn't too much left over somehow. And I'd say a language model isn't inherently that good at calculations, but if I now connect an extra tool for laminate calculation there, then I can factor that in very well too.

Joubin Rahimi

They have tools like that too, right? Yes, exactly. And with all the LLMs, if I have an image and want to measure lengths? So they say: okay, the room is now four metres by six metres. Do they manage that well, or?

Nils Breitmann

You know what? I haven't seen that yet, I don't know that either.

Joubin Rahimi

I'd bet there's a tool for that too.

Nils Breitmann

Probably. That's not special either.

Joubin Rahimi

How do customers arrive at such ideas, and how do you support customers with ideas like these?

Nils Breitmann

What we always do is, of course, first tell the story with lots of examples. Those often don't fit directly. Perhaps we have some example from the office environment or this and that, but they then manage the transfer relatively quickly. Once they see what's then possible. I have a product here, I can ask a question about it directly, and Copilot then finds the data sheet, reads through the data sheet, answers the question right there, and then they see: that would help us too. They come up with ideas really quickly then. Ideas we'd never come up with ourselves.

Joubin Rahimi

The combination of domain knowledge and what the product can do

Nils Breitmann

We have a food wholesaler in Ireland. We rolled out Copilot for them too, and for them it was important to find the nearest branch where you could then collect it. So they built in geolocation, so it first knows what the nearest branch is. And then they did something really neat, giving it a tool so it could also query the current time and then say: Oh, that branch closes in an hour. Better hurry up, it could then say that in speech.

Joubin Rahimi

We touched on that last time too, with a fuse box, where I take a photo for the tradesperson and send it back. Are there cases already? I know with us …

Nils Breitmann

We now have one with a lift manufacturer who has exactly this topic. We haven't finished implementing it yet, but I'm already looking forward to it. But you hear this a lot too, really this: hey, especially in B2B, I've got some machine in use here and now need a spare part for it.

Joubin Rahimi

We worked, I believe, seven, eight years ago with Bosch Cognitive Service to recognise spare parts, and also did that within group, but then not with Bosch, for flowers, flower hinges, to recognise these hinges. But that's not so far off anymore and that's only part of whole thing.

Nils Breitmann

Did she then need lots of photos of the hinges from lots of different angles, or what was the approach back then?

Joubin Rahimi

That's how it was, and Bosch Cognitive Service then provided a container. You could place the product on it and it was photographed 360 degrees, weighed. A 3D model, a CAD model, also came out of it, and then they fed their own neural system, their network, with it.

Nils Breitmann

Right, that's the difficulty, actually having this data in the first place. What we see, sure, the GPT model already has images built in and can already recognise things. We have a Microsoft Surface Hub, that's always our example. In the end it already recognises that superbly, but of course any special hinges probably not so much.

Joubin Rahimi

Yes, and sometimes there are small differences, like saying, okay, left or right. It's really important that you order a left one and not the right one. And LLMs are still too imprecise for that. Yes. Okay, but that would ultimately be a follow-

Nils Breitmann

Although that's not done by the LLM either, you use a vector search in the background again to find the right images.

Joubin Rahimi

But ultimately those have to be given up somehow in the end.

Nils Breitmann

Yes, all these images need to be indexed at some point.

Joubin Rahimi

All this work otherwise usually flows to the salesperson. So I have the switch cabinet, and if I don't know myself as a tradesperson, I send it to the wholesaler and say: which one is this? And then the inside sales team or sales is still sitting there – do we still need them in that form, or do we need a different kind of organisation for this?

Nils Breitmann

So if that works really well, I think it's also better for the customer if they get the answer straight away, without having to get someone on the phone.

Joubin Rahimi

Especially younger people, they don't even want to talk to others anymore.

Nils Breitmann

And then, what we're also seeing of course, is that there simply aren't as many salespeople with this deep expertise any more, and we need to make sure this knowledge goes into the systems instead.

Joubin Rahimi

And you're not the transformation unit for sales, but perhaps you have some insight into this from real life, so to speak: how can sales teams change? If sales really have time to no longer make such enquiries, should they retire? Or are there other valuable tasks that an AI can't do, which would also help the company?

Nils Breitmann

Absolutely. You could also have an agent doing churn prevention here, that identifies, this is a customer who perhaps hasn't bought anything for a long time. And there it's again a good task for the sales rep to just call and ask what's going on, to find that out.

Joubin Rahimi

Exciting topic. There's an anecdote coming up on this too: should the salesperson do it or should an AI call then? That's exciting.

Nils Breitmann

I think a salesperson might still have a bit more sensitivity than an AI that just sends out a questionnaire or asks some standard questions.

Joubin Rahimi

I find the question genuinely fascinating.

Nils Breitmann

Let's go into that briefly. Perhaps the customer is of course more honest with the AI, saying what the real reason actually is.

Joubin Rahimi

Yes, speaking for myself, I've had calls like that before, and the first time I hung up straight away. Simply on the grounds that anyone who can't be bothered to call me personally isn't worth a cent. I'm not interested. And it goes further: if you get loads of calls like that which you don't even want, then the phone channel eventually becomes dead. You just can't use it any more once that starts happening.

Nils Breitmann

Yes, that's phone calls anyway, that's difficult as it is. Someone calls you. Why should what they want from you be more important right now than what you're currently doing? That's always the question.

Joubin Rahimi

I have an idea. We need an agent up front that first checks, is this human or not? So a virtual secretariat on my phone number.

Nils Breitmann

Does the iPhone already have that, where if an unknown number calls, it first has to speak on the answering machine what it's about, and you just get a message about what it concerns? Exactly. I think that already exists.

Joubin Rahimi

Yes, exactly. Perhaps thinking about it differently, but that's interesting. I actually tried that once too. I asked the agent back then. So the other way round: who did you call before?

Nils Breitmann

And? Yes. Did it work?

Joubin Rahimi

Told me the name.

Nils Breitmann

That's naturally also an important topic with these agents, establishing quality and security overall. There are even dedicated agencies whose sole job is to test agents, try to hack them in some way and check that they do the right thing. In our case it's things like: does it perhaps offer competitors' products and claim they're better than our own? Absolutely not allowed. Or there was also this issue of overstepping competence. Does it suddenly start giving safety advice for a construction site because it thinks it recognised that, when of course it's not allowed to, and should instead recommend at most one product.

Joubin Rahimi

Yes, that was funny though. He could then … I knew what the topic of conversation was.

Nils Breitmann

And did you get in touch there? Quite bluntly.

Joubin Rahimi

I wrote an email about that. Exactly. Yes. Exactly. But that was more, first time I hung up and afterwards thought: I should ask different questions there. Okay, we came to the agent and salesperson should... So my tendency is, if the salesperson calls there, it's simply even more valuable, and if they even have, maybe a saleswoman, a personal relationship, then I'm more likely to pick up. But we'll see how it develops from here into future.

Nils Breitmann

Yes, otherwise the way things go is that certain activities … Last time I met someone at a hotel, a travelling salesman as they say, who sold fire extinguishers of some sort and thought: "Oh, that's also the kind of thing that gets replaced by e-commerce, which has long since happened."

Joubin Rahimi

Yes, that's right.

Nils Breitmann

We don't need to worry about that then, somehow.

Joubin Rahimi

And actually this affects us too. So how do we develop software? How do we also develop the client's solution? I'd ask, if you still have time, shall we make a separate episode on that?

Nils Breitmann

Sure, gladly. Yes, great.

Joubin Rahimi

We've settled into it now too, I notice. Because today, what I'll take away directly first, AI can massively support the salesperson – I'd say support – and present their activities differently. And that's possible already today. You can start implementing that today already. And in the preliminary discussion we also said: in the past it was a few hundred days quickly done, and today it's considerably smaller, to build in such quick wins too. And that, I believe, is a huge characteristic and change you have to have. And I'm pleased you're ahead on that front. Right, and the trade outfitter, we'll be in touch. Alright. Do you have anything to add? I've only said two things now. Often I have a third point too, where you say, this is simply something else worth sharing.

Nils Breitmann

Of course you also have the seller side, but of course also the buyer side. That will ultimately also decide what prevails. Convenience, simplicity, will always win out in the end.

Joubin Rahimi

Completely. Exactly. There's no limit to laziness, my father-in-law says. Right? That's probably exactly it.

Nils Breitmann

And of course tip: start as early as possible, so you can already learn from it and already get data. You then see what customers actually want, especially in such a chat. You've got all the logs there too.

Joubin Rahimi

And I'll underline that. Anyone who wants to discuss further, drop a comment below or message Nils or me. Thanks for listening and watching. And thank you. Thanks.