AI in Procurement | AI agents: Autonomous processes & efficiency boost in the company!

Fabian Heinrich (00:01)
Hello dear listeners of Procurement Unplugged. I am particularly pleased that you are back with us for another episode with Dr AI, where we are here for the third time to shed light on the topics surrounding AI in procurement. In the last episode, we looked at how I actually get from the general AI of the basic model, the big keyword, which passes the Bavarian Abitur with 2.0, to the use cases that I want to have solved with special expertise.

Then we somehow learnt from Dr AI that the whole thing has to do with Vertical AI, that there is the magic word ‘RAG’, Retrieval for Augmented Generation and that this can then be boosted with vector databases. We've already heard the magic word ‘agents’ a few times in recent weeks and that's what today's episode is all about. I was also in the USA recently, where everyone is talking about AI agents and AI agents are the new home-grown tool that will take a lot of the work off your hands.

Yes, of course you shouldn't confuse that with conventional agents, but I mean Klaus, maybe you can shed some light on it, as always. What are these agents all about? What do they mean and why has there been so much hype about them in the USA in recent weeks and months?

Dr. Klaus Iffländer (01:39)
Yes, of course, definitely. So AI agents are definitely a new development in this area, because people naturally want to go beyond these chat applications in their day-to-day work. We already discussed the use cases that can be realised with them in the last episode. And now, of course, the demands on these software systems are growing. And I think it's also clear because...

If you already have computer systems that are as intelligent as a human being in many areas, then of course you also want them to help you more in your day-to-day work, i.e. to be a real support, to become a digital colleague, so to speak. And that is precisely the approach with these AI agents.

Namely... They are now programmed in such a way that they do not work on this input-output principle. So I don't just ask the model a question and get an answer back, not even an answer with augmented company data, but I give a more complex problem to this agent and expect it to solve it autonomously.

In other words, these AI agents are software entities that use AI, in particular generative AI, to independently solve or implement tasks.

Fabian Heinrich (03:18)
Is that also the next step in development somewhere? In episode 1, we also talked a bit about the fact that perhaps German companies are very process-driven, which is why they are perhaps still a bit, I don't want to say blocking, but expressing restraint regarding the introduction of AI somewhere. So are these agents really the next step, where you say you're now moving away from, I'll call it kind of funny chatting with the AI?

Dr. Klaus Iffländer (03:18)
Exactly.

Fabian Heinrich (03:44)
Too much process automation in the existing systems, in the digital software systems?

Dr. Klaus Iffländer (03:53)
Yes, I would definitely recommend that, because it really makes a big difference in everyday working life. In other words, these agents can really act autonomously and make independent decisions and interact with their environment. So it's not that you just get a chat output, but that they can actually carry out real actions or operate other systems. That's how you have to imagine it.

Fabian Heinrich (04:18)
Yes, and if these, so it sounds like these agents are now the new wonder weapon, I mean where do I get them from, can I then also use them with, let's say, Chat GPT Gemini, the other basic models, so maybe you could somehow organise that again, do I have to program them myself somehow from my in-house IT or is that the expectation that the software I already use provides me with these agents?

Dr. Klaus Iffländer (04:45)
Yes, there are both. There are, of course, ready-made agents. Firstly, they are usually based on these basic models that we have already discussed. Or they use generative AI models or large language models as a basis for understanding the problem in the first place. In other words, being able to understand a problem, think it through, structure it and then break it down into individual subtasks that can then be mastered.

Fabian Heinrich (05:16)
But if I'm a purchasing manager, I can't just say in our corporate version of Chat GPT, ‘Hey, everyone use the market analysis agent for category buying now, because it doesn't exist. It doesn't exist in the Chat GPT App Library either. So how do you have to imagine that?

Dr. Klaus Iffländer (05:42)
As a company, you can either choose innovative providers such as Mercanis, for example, who already provide certain agents that are preconfigured and have all the tools and understanding for certain use cases and can solve tasks. So that's one option. So you buy it externally, so to speak.

Or if you take a very fundamental approach and want to develop your own AI applications in the company, then you can also develop such agents yourself. Because they also use basic models to some extent or models that they have trained themselves. And it's more of a design approach to software development. So it's no longer input output, but the approach is that the agent can act autonomously.

In other words, it doesn't carry out a predefined sequence of tasks, but can also vary this sequence, i.e. the agent determines it itself. That's exactly the difference. So it's clear, isn't it? There are providers for agents or you can develop them yourself with this software.

Fabian Heinrich (06:53)
Yes, so that means that, historically, if you look at it now, it's like, let's say, 20 years ago, where I then looked at on-premise or somehow cloud or 10 years ago and I then had a make-or-buy decision, so to speak. Similarly, it now also has an influence on which software I choose.

So if I now choose a software for purchasing that doesn't somehow already include or offer the agents, I automatically have a make decision somehow, because then I have to need the agents somehow with my own IT. Is that right?

Dr. Klaus Iffländer (07:29)
Exactly. And as with software selection, you also have to put together certain systems. And I don't think it's typically the case that you buy a single agent, you buy the entire platform. So, for example, if I'm a call centre operator, I don't buy an agent and place him somewhere to take calls, but I buy a ...

platform that is able to accept, allocate and handle customer service calls. And this can of course be supported by autonomous agents. But these are then typically integrated into this platform. And that's exactly how it works at Mercanis, for example.

In other words, as a manufacturer of a software suite for the entire purchasing process, where certain agents combine certain skillsets and then function autonomously as agents. But within the sites.

Fabian Heinrich (08:27)
Which means that no matter which system I introduce in my role, I have to look at whether or not they have agents with them when selecting software these days. And is it now Salesforce with Agentforce or somehow HubSpot with the Breeze Copilot or you mentioned GitHub or then in the purchasing system like Mercanis with the Mercanis Co-Pilots, where these agents are already included. So that changes for me as a purchasing manager, that I now also have to look at the software output: Does the system have agents?

Dr. Klaus Iffländer (09:07)
Yes, I would say that if I am a traditional company that is not active in software development or in the field of AI, then it would be unusual to develop agents myself, but rather to buy them in with a software platform that is correspondingly innovative.

Fabian Heinrich (09:27)
Yes, I mean, as a result, I would now say that we should look at this in a little more depth, in the sense that we have such a great abbreviation as APA, which stands for Argentic Cross Automation, what it's all about and what the fields of application ultimately are. If you read up on it a bit, you hear more and more about something at Google: 25 per cent of the code is written by agents, which, given the size of Google, equates to the engineering output of 19,600 software developers.

These are already gigantic figures that such agents can bring in in terms of efficiency, which can then really lift companies 10x in efficiency. Perhaps that's another interesting question: what are these vertically integrated AI agents and what is the APA?

Dr. Klaus Iffländer (10:27)
Yes, APA, the term is somewhat analogous to RPA, i.e. Robotic Process Automation. It actually comes from the fact that people have tried to automate existing workflows or at least partially automate them with this robotic approach. Hence robotic process automation. The only thing about RPA is that strictly predefined processes are automated. For example, in Excel, you can imagine taking cell A1, copying it to cell B1 and calculating the sum of the two in cell C1.

So it's always exactly predetermined. And of course it can sometimes... make certain decisions itself, but these are very deterministic calculations. For example, if I summarise the figures for the last week, then it's just the figures for today minus or from today to today minus seven days. So exactly one week ago, all these days are then calculated, which days they are, then the numbers are added. Robotic Process Automation can do things like this.

And in contrast, Agentic Process Automation, or APA, goes beyond that. It uses existing workflows in the same way, but is much more far-reaching. So it has a much deeper understanding and can then act agentically, i.e. it can make decisions itself, can consider which figures make sense, where do I get them from, which systems do I have to operate, then put these steps in sequence, then execute each individual step, actually address the systems and then formulate an answer and then perhaps process it further.

So it's more this agentic approach where the systems work autonomously and are not given precise instructions for every single sub-step. That's the big difference. And that's where the term RPA comes from.

Fabian Heinrich (12:37)
So on the one hand, it sounds like the RPA is not always so easy to implement because I had a lot of work to set it up originally, as you mentioned, with the matching of the fields and the cells. Now the APA sounds from many variants as if I have even more advantages from the RPA, plus it's easy to set up.

So if I look at it now, I could use the APA in some very complex processes where there is unstructured data and I first have to understand the interrelationships, but on the other hand I don't have to set anything up beforehand because I sometimes or perhaps always work with ready-made agents that the software provider already provides me with. Do I see that correctly?

Dr. Klaus Iffländer (13:26)
Yes, of course you can't do it completely without setting it up. You have to have these interfaces. It has to be able to operate meaningful systems and things like that. Or, of course, you have to test whether the understanding is really there so that you can let the agent act autonomously. So you have to make sure of that.

But you no longer have the insane effort of first creating perfect data quality and then programming this robotic process so that it always works reliably and robustly. So this effort is no longer necessary, but you have to think of it more like a colleague in principle. Like a digital colleague who also has certain skills and simply understands certain things and can then act on them.

Fabian Heinrich (13:58)
Hmm. Okay, so if I trade with agents that are already integrated in the platform, I don't have any integration effort, I suppose?

Dr. Klaus Iffländer (14:31)
Exactly, because that is precisely the advantage of this integration into the platform, that these interfaces are already there.

Fabian Heinrich (14:38)
And then again I have the advantage that I don't have to pre-structure and set up anything, but on the other hand I can interact like with a colleague, as I can also outsource tasks or processes with high risk, high complexity, which will hopefully also bring me a high benefit, to these agents.

Dr. Klaus Iffländer (14:59)
Absolutely. So this understanding and the data and the interfaces are all brought along. Exactly. And then the interaction with the agent is also organised by the platform. Yes, people are often afraid that it will go so far that everything will be automated straight away and they will no longer have any say.

But that's not typically the case; the platform manufacturers are also aware that the agents don't work perfectly in every unforeseen situation, so there are typically always people involved in the process at the useful points.

Fabian Heinrich (15:38)
And do these agents also learn with them or do they also get better or how should we understand that? If I now somehow have my procurement category manager, who is constantly evolving, does the agent then stand still or does it also evolve?

Dr. Klaus Iffländer (15:52)
It depends on how this is implemented, but typically it is already done, precisely this interaction with the users or with the human colleagues is used to learn. If you already have this integration with the platform and have utilised this agentic approach to software design, then you also have precisely this interaction in use.

A human does this, an agent does that, a human says it was okay or not okay. And it is precisely this human feedback, whether it was sensible or useful or the right decision, that is worth recording because it is then already available in exactly the form that is needed to further train such an AI system. So precisely these subtleties, what was right or not right in the situation, can be used very well to improve the agent so that it doesn't make the same mistakes in future.

That's why it's typically integrated in this way, because the data is already collected in this form anyway. Because in other scenarios, if you haven't implemented an agentic process automation system, but simply an LLM that answers queries, then you typically don't have this feedback integrated in this form. And with the agentic approach, you typically have it available. That's why you should also use it as training data.

Fabian Heinrich (17:30)
I mean, it all sounds extremely cool and maybe too good to be true, but what about data security? Because let's assume it works the way you've just explained it. Then I somehow have, let's say, an additional category manager who I have next to me as an agent.

And then I can train it somehow, or it gets better and better, prefabricated, it can already... these five processes, can do increasingly complex processes, perhaps at some point my entire strategic category management for me. How can I make sure that it doesn't spit out this data for another customer of the platform, but that I train it for myself and my company or my categories accordingly and that it doesn't necessarily spit out the knowledge to my competitors? That is certainly a legitimate concern.

Dr. Klaus Iffländer (18:25)
So in this case, if the agent works on the platform, then it is ensured by the platform. An agent in itself is actually just a design principle for designing software. But data security, i.e. that it is kept separate, must of course be ensured by the software manufacturer. In this case, the platform on which the category manager works.

Fabian Heinrich (18:57)
You would say that if you have the usual providers, then that is actually always guaranteed.

Dr. Klaus Iffländer (19:03)
Yes, I assume so. I know it's the case with Mercanis, but I think it's implemented in a similar way with the other competitors.

Fabian Heinrich (19:12)
I mean, there is always a lot of interest in the use cases. Maybe we can talk a bit more about that. I mean, if I were to start tomorrow, for example, which cases could I put agents on?

Dr. Klaus Iffländer (19:33)
Yes, are there various use cases that are relevant in the purchasing sector? Market analyses, for example. You often have that or... It's also a challenge for companies to ensure permanent monitoring. But if you are active in certain areas and are constantly procuring important materials in procurement, then it is important to keep an eye on the market, to anticipate or be aware of developments and to be able to utilise them for yourself as a company. For example, prices or indices or industry news that you are constantly aware of and you could definitely appoint an agent for things like this, so that you can...

Fabian Heinrich (20:21)
So if I break it down again, if I'm in the old world now, then I have my category manager, so I'm perhaps already quite well positioned digitally, the software tracker, various indices, monitors, various news, perhaps making a report or even having the software report, how would the whole thing now or could then of course also see which articles are... are influenced. Of course, he always has to spend time on this.

What would the whole thing look like with such an agent, which would then take over the entire job, perhaps even replacing this risk software?

Dr. Klaus Iffländer (21:01)
Yes, you would have to structure it a bit in advance. You would have to say exactly what kind of information this agent should collect and roughly how it should be processed and what conclusions should be drawn from it or how it should be interpreted. You can certainly orientate yourself on the results of human work if a human has actually done it before ... a colleague, a category manager for example, then you can use these reports as a template, so to speak, to teach the model or the agent what kind of report is expected here and where the information comes from.

But after that, it's largely automatic and a lot of the colleague's labour is freed up. Because, to be honest, it's a very repetitive task. He searches the same industry news again and again, gets new information again and again, but on the same pages, in the same formats, the same price information or industry indices. And this data is compiled again and again.

Manually, as you can imagine, it is of course an incredibly manual process that is not particularly exciting and does not provide employees with much training or challenge them at all. That's why it's a very good idea to automate things like this using the agent. And they could do exactly that. Once it has learnt or is connected to the data sources, it can constantly retrieve this information and generate reports from it.

Fabian Heinrich (22:43)
Keyword data sources and reports, I mean topics such as business intelligence jobs, dashboards, analyses, I could probably handle that quite simply via such an agent anyway.

Dr. Klaus Iffländer (22:54)
Exactly, exactly. Typically, performance data is available in companies. It is stored in databases or in other reports or in production systems, production systems, trading systems, purchasing systems. Everywhere there is data that depicts the actions of this company.

And as soon as they are connected to a business intelligence system, an agent can of course create reports or analyses and interpret this data. And an agent is actually perfect for this because it is similar to the skillset of a business analyst, who also looks at and interprets the data and draws conclusions from it. And a digital agent would do the same.

Fabian Heinrich (23:45)
So basically, that means I’m making my team significantly faster and more efficient, freeing up resources somewhere, and on top of that, I’m also tackling the skills shortage with the agents—because, in a way, I’m developing my own specialist workforce through the whole LPA topic.

Dr. Klaus Iffländer (24:05)
That’s one way to look at it, and I’d even say that people are less bored with these tasks. Because we know that in every job, even highly qualified ones, there are always tasks that are very repetitive and require a lot of manual copying and transferring of information. And those are exactly the kind of tasks that these agents take over.

Fabian Heinrich (24:28)
We’ve already touched on this in previous episodes—the search for suppliers is always a recurring topic. Finding new suppliers often falls to the strategic buyer or the category buyer. That would probably be a perfect case for such an agent.

Dr. Klaus Iffländer (24:42)
Absolutely, the initial search in particular. Of course, conversations and further development are very people-centric tasks. But the pure act of finding—researching on the internet, gathering and preparing information—is highly suitable for agents.

Fabian Heinrich (25:02)
Exactly! If I say, "Hey, I have my preferred supplier, and based on that, I’d like a longlist and a shortlist," the agent could handle all of that for me.

Dr. Klaus Iffländer (25:13)
Exactly! The agent would abstract from that, analyze the information about the preferred supplier, identify the key characteristics, and determine what to look for. It would then carry out this research automatically—faster and often more comprehensively than a human.

Fabian Heinrich (25:34)
Exactly! From a procurement manager’s perspective, I could take it even further and say, "I’m freeing up resources, tackling the skills shortage, and—when I think about it—I might even save on software costs." Especially in times when companies are looking to consolidate and make savings, I could argue that, depending on the use case, I might not need separate software for risk analysis or supplier scouting anymore.

I no longer necessarily have to go with a niche provider; agents can handle that for me now. Is that perhaps too casually put, or could I actually say that, over time, these agents will replace one or the other software provider and that I will more or less have my agent for such things?

Dr. Klaus Iffländer (26:23)
Agents are also just software. But you can definitely see it that way because, essentially, you're replacing older software that can only handle very specific tasks and still requires manual operation with a much more autonomous agent. So, while it’s still software, it’s a much more modern, autonomous one that covers a broader range of functions.

So it’s possible that there will be some cannibalization if there are existing industry solutions that are very niche or operate in a traditional way. In that case, it could well happen that several of them are replaced by an agent-based system.

Fabian Heinrich (27:05)
Yes, Klaus, thanks again! This was—clearly for many—a completely new topic. Agents, Agentic Process Automation, and once again, we’ve learned a lot. From understanding what an agent is and how to use it, to diving into Agentic Process Automation, and of course, as always, exploring real-world applications. And an exciting thought to wrap it up—agents potentially cannibalizing traditional software. I think that really gives us something to think about.

We look forward to seeing you again next week. Until then, best regards, and thanks!