AI in Procurement | Vertical Agents: Specialized AI for more autonomy

Fabian Heinrich (00:01)
Hello dear listeners for another episode of Procurement Unplugged with our special AI with Dr Klaus Iffländer. Klaus, I'm delighted to have you with us again today. Welcome back.

Dr. Klaus Iffländer (00:16)
Hi Fabian, I'm delighted to be here again.

Fabian Heinrich (00:18)
Today it's all about vertical agents in procurement. We have already shed some light on this topic in recent episodes and today we want to delve a little deeper into the subject of vertical agents. What it's all about has been in the media a lot in the last few weeks. Now somehow the Google CEO has said, okay, maybe soon there will be no more software-as-a-service companies. A very controversial quote, which we'll come back to later. But to start with Klaus, what are vertical agents and how would you differentiate them from normal agents?

Dr. Klaus Iffländer (01:02)
Exactly, we've talked about AI agents before. AI agents in general differ from traditional software, I would say, in that they are not so rule-based. So it's no longer the case that if this happens, then that happens, but an AI agent itself has autonomy and the ability to learn and react. In other words, you can think of it much more like a digital employee. It acts much more autonomously, so to speak, and could take on things like supplier research, price negotiations or order processing in a procurement context, for example. In other words, completely independently.

Fabian Heinrich (01:45)
Yes, but let me ask you a stupid question, Klaus, how could I not simply do this with the normal ChatGPT? Well, I could probably enter it there too, maybe build agents there that would then perhaps take over such tasks for me.

Dr. Klaus Iffländer (02:01)
Yes, well, with ChatGPT it's just a chat window. So it can't really take over complete tasks, but you would at least have to copy it back and forth into another system, the questions and answers. And then, as many of our viewers have probably already experienced, ChatGPT is a very generalised system. This means that very large amounts of text are taken from large tech companies that produce this system.

ChatGPT, for example, OpenAI and Microsoft. So there are content deals, for example, where large amounts of very high-quality text are processed to train these models. But of course, this training or the material on which these systems are trained is very general. In other words, they are not very, very specifically tailored to a particular area such as procurement. In other words, you couldn't use ChatGPT in such a generalised way.

Fabian Heinrich (03:06)
So Vertical Agents basically means that I build my own Procurement ChatGPT, B.T., which then, to return to our example from the last episodes, doesn't have the knowledge of a Bavarian school-leaver, but rather the knowledge of a category buyer. So I have the ChatGPT with the brain of an experienced category buyer.

Dr. Klaus Iffländer (03:31)
You can see it that way, yes, very well explained. Yes, you have to imagine it like this... Well, ChatGPT is already very good in many areas, even in specialised areas, you have to admit. Nevertheless, when it comes to proprietary data on the one hand and very specific expertise or extensive background knowledge from a domain such as purchasing on the other, the human still wins.

So, it's not yet at the level of a real digital employee. And that's exactly what a vertical agent is. They integrate themselves vertically into a sector, an industry or a function and bring this extensive background knowledge with them so that they can really get started autonomously.

Fabian Heinrich (04:22)
Yes, very exciting. I mean, now in the corporate environment, could I build it myself or do I have to buy software again or how do I get there, so to speak?

Dr. Klaus Iffländer (04:37)
In the corporate environment, it is ultimately software like any other, but it differs in the way it works and how it is connected. It's not like that...- So of course, sometimes there are ready-made software products that then have to be configured, as with other software systems. And of course there are also platforms that you can build yourself. But that's relatively time-consuming, because then you have to...

Interfaces to the existing systems and also this agent-based approach. This also has to be implemented first. So you also need the right skills to be able to build this software system. These are the main challenges when you are faced with the make or buy decision.

Fabian Heinrich (05:25)
I mean, regardless of whether you build it yourself or buy it in somewhere, the point is that I automate existing things, but above all things that are not yet automated, because then of course I have an additional buyer on demand, so to speak. But if I understand the whole thing correctly, I'm already somehow cannibalising existing software processes with these agents.

Dr. Klaus Iffländer (05:49)
Absolutely, but that's exactly the goal. Because right now, you may already have certain software systems, but they are very rule-based, for example. That means there are still a lot of manual steps in between. You need someone to operate it. You need someone to carry out one step after the other, making certain human decisions in between. And it is precisely these processes that agent-based systems are supposed to replace, because they are supposed to solve these problems independently.

Fabian Heinrich (06:17)
Yes, and I mean every development has opportunities and risks. I think maybe we should look at the risks first. What risks do you see here if you now become completely agentic in the process landscape?

Dr. Klaus Iffländer (06:35)
One thing is the integration of data from different systems. For example, the ERP system or CRM system or market data, depending on what the agent is to be used for. This is always a problem, but also with other software systems. Then there are also things like training data. Of course, before you unleash this on reality, you have to analyse the training data...

And also test that it really works the way it was planned. Otherwise you have the ‘garbage in, garbage out’ problem again. So sensible training data is essential in order to get high-quality results. And then

Fabian Heinrich (07:22)
The training date is a multifaceted word. Isn't it the case that I just have to feed my brain, which already has such a strong generalist basic level? Or do I have to train it first, as I did in the past with machine learning algorithms using huge amounts of data? Or is it enough if I feed it various documents?

Dr. Klaus Iffländer (07:44)
Yes, you're right. It's not like training as with traditional machine learning, but you do feed in the documents first. But of course, the agent can only provide the knowledge or give good answers in the areas that it knows. In other words, you need to have the exact training data so that it can react accordingly. Of course, it understands it better than a machine learning system.

But here too, logical thinking, for example, has not yet fully arrived in the LLMs. They are not yet able to develop new insights for themselves, but only know the documents they have seen. And in those areas they can only react and give good answers. In other areas that they don't know, they are just as clueless as GPT.

Fabian Heinrich (08:38)
And I guess then of course there are the legal risks as always, right?

Dr. Klaus Iffländer (08:45)
Yes, legal risks. do you mean in the sense of who bears the responsibility? That is precisely the question. Is it the software manufacturer or the person who operates the system? Yes, I think it can be decided on a case-by-case basis, if at all. Otherwise it's a huge problem. I don't think there are any good legal regulations on this yet.

Fabian Heinrich (08:48)
Completely. Right, yes and of course, that's why we're talking about the opportunities and advantages, which of course outweigh them massively. Perhaps we can go into this again briefly, because of course I have enormous advantages here. So if I see the whole thing in a future state, then I can somehow scale my team of category buyer agents up and down for the various tasks.

Don't actually pay a monthly software fee anymore, but then pay sort of like an on-demand fee for how I use these virtual agents as a category wire.

Dr. Klaus Iffländer (09:51)
Yes, or a salary for part-time employees.

Fabian Heinrich (09:56)
And I mean, how could I then... So what would be things that go beyond the scope, where we now have tasks where you say, okay, these are really tasks that a category buyer does today, which the agents could then also do quite solidly.

Dr. Klaus Iffländer (10:15)
Yes, what I find fascinating would be to analyse use cases such as supply chains. So if an agent manages to bring this background knowledge to the table and, above all, to analyse data from different systems or different data sources, to recognise such complex relationships, then it has great potential because, for example, you can't simply trigger an order now if the stock is too low in the warehouse, but because such an agent could now also take into account where there are currently supply bottlenecks, how prices are likely to develop in the future and can take all these factors into account in order to place significantly optimised orders.

So I find things like this fascinating, or even to be aware early on if there are supply bottlenecks somewhere or if stocks need to be increased.

Fabian Heinrich (11:16)
I mean, risk assessment is a big topic in general. So every procurement SRM system, let's say, has to evolve in this respect somewhere. At Gardner and SpendMatters, with the analysts, this is also becoming increasingly important. Are there ways in which these AI agents can then provide real-time updates on risk factors, risk assessments, perhaps also based on global events, and thus of course make the SRM system more resilient?

Dr. Klaus Iffländer (11:48)
Yes, definitely. That's exactly the kind of thing you can do. For such an agent-based system, it's another data source where you might have to decide how it should be analysed. But in principle, it is absolutely possible. And this is of course something that is very difficult for humans. Even if you have a team of such category managers, it is difficult for them to constantly monitor all possible geopolitical events and their implications and then derive good sourcing decisions from them. But for a computer system, it is of course completely possible to keep an eye on all data sources at all times and then derive the right decisions from them.

Do you mean such risks?

Fabian Heinrich (12:37)
Yes, I think that's a super, super use case. So we can see that it can be used in many different ways. What I'm still wondering is, when you're in purchasing, you always want to keep everything on a budget somehow for years to come. Of course, with the software licences, I think that was quite predictable over three, four or five years. Of course, if I now switch the agent on and off at any time and then pay the agent according to consumption, then on the one hand it's like having an army of intelligent temporary workers on call, but of course it makes it very difficult for me to plan and quantify the entire budget.

Dr. Klaus Iffländer (13:24)
Yes, that's true. But you already have the same problem with cloud offerings ... ... and software-as-a-service offerings ... where you pay on demand. So I think there will also be ... predictable ... in the future. pricing models will crystallise in the future. Or ... ... it is ... a win-win in any case, because the savings are so great. Or because the use is also so ... inexpensive compared to an employee that it's a no-brainer for companies. So these things will happen, I think.

Fabian Heinrich (13:59)
You don't see this cost trap, which cannot be planned, because you say, okay, the value that is donated is always higher than the costs that you would otherwise have.

Dr. Klaus Iffländer (14:10)
Exactly, either that or you make it plannable with appropriate pricing models. They may then be more expensive in the block, but they can be planned and then you can decide as a company whether it is worth it or not.

Fabian Heinrich (14:23)
And if I were to get started now, regardless of whether it's make or buy, but what would the technical implementation look like and what challenges would I face?

Dr. Klaus Iffländer (14:36)
Of course, we need the data as a basis, as we have already discussed. In other words, high-quality data with which it is to be trained. You should also have thought about the use cases, i.e. exactly how you want to use it. And then it's about interfaces and, as with other software systems, adapting change management processes. So I see things like that more as a preliminary reason. The purely technical implementation.

Fabian Heinrich (14:59)
Yes.

Dr. Klaus Iffländer (15:03)
Again, it's a question of how it fits into the existing infrastructure. Do you have the skills for it in the company? Do you have the developer capacities? Or are there perhaps ready-made products that are more favourable? So these are the main framework data for any software hijacking.

Fabian Heinrich (15:20)
But if a software provider offers me this now, let's say out of the box, Vertical Procurement AI in Mercanis, such as Mercu AI Copilot, then I don't have any technical effort for the time being and can get started tomorrow, so to speak.

Dr. Klaus Iffländer (15:37)
Exactly, software as a service with all its advantages - plug and play, so to speak.

Fabian Heinrich (15:42)
Good, now we have the keyword Software as a Service here again. We have already mentioned Satya Nodala's quote that this will disappear. If you look at LinkedIn, you can see more and more that Software as a Service is dead. Yes, now, of course, is this perhaps a new or different type of software as a service or where will it go from here? If you look at purchasing now, people have been getting used to software as a service for a few years now. We've moved away from the old on-premise systems and now we're suddenly being told that Software as a Service is dead again. How do you see the whole thing?

Dr. Klaus Iffländer (16:28)
Yes, so Nadella, I saw that interview too. He explains it like this, it's actually like another form of software as a service. The way he sees it, software as a service is basically the user interface. And then there is a layer for database interaction. Create, update, delete operation. Imagine you create suppliers. Then it's either Create or you.

You change everything you have saved for the supplier, then it's an update operation, or you take it out of your portfolio again, then it's just Delete. And Nadella believes that all these things will disappear because they will simply no longer be needed, as the software agents will be able to interact directly with the databases in future. This means that a large part of this software stack will no longer be needed because the LLMs will simply be able to do it themselves. They speak the language of every database. This is completely interchangeable and doesn't matter.

Whichever data marker you use, the agent will be compatible with it because it can handle it.

Fabian Heinrich (17:32)
Although I think it's similar here to sales, of course you need a kind of system of record, where I can somehow keep all the data...

Because if I don't have the data anywhere, of course - you mentioned garbage in, garbage out or I don't have any data at all but only analogue - then of course it becomes difficult. So I also believe that SRM is now becoming irreplaceable in purchasing, because that's my data basis somewhere, where I have all the difference master data, where I have all the difference evaluations, all the onboarding qualifications and, based on that, I don't actually need the whole other software stack anywhere.

Dr. Klaus Iffländer (18:16)
Exactly, but this supplier database can also maintain your agent.

Fabian Heinrich (18:23)
Right, exactly, I just need a system that gives me the database, so to speak.

Dr. Klaus Iffländer (18:30)
Exactly. And what's left is just a user interface. And Satya Nadella also says that this is reduced to the interaction with the agent. So we as software users will only talk to the agent. In other words, what's left but a small chat window in which I tell the agent what to do. And the agent even goes so far as to use its flagship solution, Microsoft Excel

Fabian Heinrich (18:41)
Yes.

Dr. Klaus Iffländer (18:57)
And to say that this will largely disappear, because why do we still need these Excel spreadsheets, in the end they are only there so that we as viewers, as software users, can understand what the agent is doing so that it can visualise it for us. Because it doesn't matter to the agent if we say they should do budget planning, they can do it in a database or somewhere else or internally. He doesn't necessarily need to visualise it, he can output it as text or in any other way and then take it.

Microsoft Excel only as a form of visualisation for me, so that I can understand and visually comprehend what he has calculated. But as a user, I don't actually have to work in Microsoft Excel any more. I can just tell the agent what to do, what to calculate and it does it all for me. So I no longer have to do it manually myself.

Fabian Heinrich (19:48)
Yes, I think Microsoft is doing this very intelligently, in the sense that they are disrupting themselves to a certain extent. I think it will be the same in procurement. When I look at how many solutions there are, somehow 200, 300 solutions, that's too many. I can already cover many solutions with vertical agents these days. And I think another issue is that as a procurement suite, like Mercanis, you probably have to consider, okay, either ...

A vertical-agent-based solution, disrupt me in two, years, or we start disrupting ourselves by ...

I think that's also the essence of Satyana Dalla's quote, that the way it exists today, that you somehow have x different software applications, will no longer exist. So if you look at two or three years from now, I think there will also be some kind of system of record in purchasing, perhaps a kind of extended database where you have all the topics. I would say a kind of basic SRM system and based on that

Of course, I then have all these vertical agents who are my category buyers, or who can completely take over all my tasks and duties, such as sourcing, price negotiations, risk planning and even scouting. I mean just these 5 or 6 tasks that I've just mentioned, they can do that.

6, 7 software players that I can replace with it, which I probably currently have in use.

Dr. Klaus Iffländer (21:27)
That's exactly how I see it. And it's happening in other areas too. Software development is also moving in this direction and many software solutions will have to reinvent themselves in some way. And in procurement, I don't think the development will bypass the sector either.

Fabian Heinrich (21:44)
So you can already see in the CRM area, the very good CRM solutions that think ahead, they already offer agents. So if you look at how many people Salesforce is hiring in the Salesforce Cor operation and in Agent Force, i.e. the new, worthy, agent world at Salesforce, then I think they are hiring three times as many people in Agent Force than in ...

Salesforce and I think, if you look at it, Procurement is always a bit behind Salesforce in terms of development and digitalisation. So I am firmly convinced that the next big source-to-pay suite will be a simple, user-friendly SRM system of record with, let's say in inverted commas, Agent Force on top, i.e. with the network of all the different vertical agents.

Dr. Klaus Iffländer (22:43)
Exactly, which also interact with each other and solve complex tasks together.

Fabian Heinrich (22:49)
I mean, that's also the keyword again, multi-agents, so if you light that up somewhere. Because I think the outlook was quite exciting in terms of where it can go and you then somehow have a multi-agent system. Perhaps we could take a closer look at the technical side of things.

Dr. Klaus Iffländer (23:11)
Yes, so a multi-agent system, exactly, as the name suggests, is simply a system that consists of several different agents and can be imagined in the same way if you have a team in the company. They work together and solve problems together. Technically, of course, the interfaces and the exchange are always a challenge, but with the LLMs it has also become much easier because they can communicate with each other in natural language.

Just like people do. In other words, very, very similar to how people work together as a team. And that is of course also a design model or a vision in software development or how software will be designed in the future, that you combine several of these agents and then enable them to communicate with each other and solve larger tasks together.

Fabian Heinrich (24:09)
Yes, so I think that was another very exciting conversation, also that we dived deeper into it, these vertical agents, and I also think that...

Controversial statement, software is dead. I don't think it will be dead, but it will change dramatically. I think by 2025, 2026, we will be at a crossroads where many software companies will either decide to cannibalise or cannibalise themselves. And I think that's what makes it so exciting for procurement managers or for people who want to digitise. Which horse are they backing? Are they betting on a solution that might

It's the market leader now, but it doesn't have the chance to cannibalise itself, it doesn't have the chance to go along with it and perhaps in two or three years it will somehow no longer be usable or I'll be behind it, because it doesn't put itself in this Argentic world and go along with it, so to speak, or it relies on a solution that already says, okay, we're disrupting ourselves and we're not going to be able to use it.

Wir kreieren tagtäglich neue Agenten, die verschiedene anderen Nischen-Software-Player dann auch disrupten.

Dr. Klaus Iffländer (25:26)
Genau.

Fabian Heinrich (25:28)
Super, Klaus hat mir wieder sehr viel Spaß gemacht. Es war ein kurzweiliges Gespräch mit dir und ich glaube, war sehr spannend, dass man da nochmal quasi in die Stufe tiefer eingetaucht ist und diese ganze Welt der Vertical Agents. Ich habe mich sehr gefreut, vielen Dank und ich freue mich aufs nächste Mal mit dir. Da wollen wir dann in dieses ganze Thema synthetische Daten, vielleicht für viele ein Fremdwort, aber wollen dann nochmal in dieses Thema eintauchen.

Dr. Klaus Iffländer (25:58)
Alles klar. Danke Fabian, dann bis zum nächsten Mal.