The AI Tsunami Is Here. You're Just Standing There | Sören Petsch, Root Insurance

Shamil Malachiyev: Good morning everyone, and welcome to this week's episode of The AI Floor. My guest today is a Senior Director of Procurement at Root Insurance, where he's also taken on leading AI integrations across finance. He's spent 15 years in procurement across software, retail banking and insurance. He's implemented workflow automation platforms multiple times over, built custom AI agents, and even ran his own consulting practice focused on procurement automation using LLMs and multi-agent systems. So please welcome Sören Petsch.

Sören Petsch: Nice to meet you. Great to be here, Shamil.

Shamil Malachiyev: Thank you so much for joining me on the show. Can we start with you telling us a little bit about who Sören is, and what makes you so enthusiastic and experienced with the way the technology is evolving?

Sören Petsch: Well, I think the starting point for me is that I grew up in a country that doesn't exist anymore. I grew up in East Germany. The only reason I bring that up is that I've experienced it before — where countries just disappear, where the status quo disappears. And it feels like we're in a moment with AI where a lot of the things that we hold to be true are changing, where the ground is shifting. The way we will be working with AI in the future will be very different from the way we have worked in the past.

So I take that skill set of not seeing the world as it is today as a given, and I'm projecting towards the future in terms of what the technology enables me to do today and what it looks like potentially in the future. I don't think anybody knows — and whoever claims they do is probably kidding themselves, at a minimum. So I'm trying to be at the forefront, because it feels like there's this big wall of water, this tsunami, coming. I can stand there at the beach and hold my hands up and try to stop it, which is going to be futile. Or I can try to ride the wave. So I'm trying to ride the wave, and the best way I know to do that is by continuously challenging myself to learn more about the technology and figure out how it actually works and how I can use it in a useful way. Some part of those efforts are wasted — and we'll probably talk about that — and some of them yield real results, where I'm just like: this is an indication of what will work and what can't work.

Shamil Malachiyev: I really like that, because part of the reason for this season of The AI Floor podcast is that a lot of people are reading all of the LinkedIn posts, which are heavily influenced by the marketing spend of OpenAI and the leading AI companies, and they force you into seeing a huge tsunami that is crashing on top of you. And the more you research, try things out, see the failures — the adoption failures — and talk to other people in the industry, the more you think: the wave's not that big really, or maybe it's just really far away and not right next to your nose. So it would be really interesting to see how far away that wave is and how large it is — how much time you have to prepare. In a conversation we had before, you told me that you analyzed around 500 job positions in terms of how much automation is really happening within them. The answer you found is quite shocking. Can you tell us about that?

Sören Petsch: Yeah — and my data set is about a year, year and a half old; I was looking at positions a year and a half ago. If you look at LinkedIn, everybody's talking about automation in procurement, and everybody has apparently already implemented it. And then you actually look at the job descriptions themselves, in terms of what companies hiring for leadership roles in procurement are looking for. And the number, as per my very obscure data set, is under 4%. Everybody else — the other 96% — still talks about the tooling you'll be using being Microsoft tooling and Slack, or Microsoft Teams and Microsoft tools, and SAP Ariba, which in my experience is not super automated. So everybody's still looking at spreadsheets, email and Slack messages, but then says "work in a fully automated way." And that doesn't work. Those tools don't work, and they're all very separate from each other. There's only a small subset where companies are looking to truly have leaders automate in a way that actually works.

I find that interesting. You have LinkedIn as an echo chamber where thought leaders are talking to themselves, and then you have the reality of companies just not hiring, or not knowing what kind of skill set to ask for and how automation actually works. And I think there's a little bit of a chicken-and-egg problem there, where you're looking for leaders — let's say I have 15 years of experience, and for over ten years of that I worked in emails and spreadsheets, and now I need to figure out how automation actually works. It's a really challenging pivot that leaders have to make in order to educate and actually drive that thinking. And I think this is where reality meets the dream state of future automation.

Shamil Malachiyev: And when companies actually make a big step into adopting AI, the first step usually comes down to understanding your data, understanding your processes. What is it that a lot of companies get wrong when they're faced with that task from shareholders or C-level executives — guys, you have to implement AI, go and do this, bring us back the results, we expect a lot from you?

Sören Petsch: Well, if you start with "just implement AI" and you don't have an orchestrated process, you're starting at the wrong end. When I think about automation — process automation — it's actually data flow automation. What I mean by that is you have to think about all the data elements that go into, let's say, a procurement request: the spend data, the vendor record, certain approvals and so forth. And you want to think about how, if I only want to enter it once, that carries all the way through to the end. That's not AI. That is workflow process automation. And the technology that's commonly used in procurement to solve that is procurement orchestration — intake-to-procure software. That works really well, and it helps you articulate how data flows through.

The other piece where this software shines is that these modern platforms are implemented as a no-code type of implementation, meaning as a procurement leader you don't have to go to IT to fix a small step or build the process at all. You can do it yourself. There's a no-code set of configuration settings that you build out, and that becomes your process. And that's why you want to start there. I haven't written a procurement policy at the last couple of companies I've worked at. What I've done instead is implemented orchestration software, and the conditional logic and the approval steps, as well as how the data carries through — all that config — that's the procurement policy. So the system becomes the policy, and it spins up different approval steps depending on what kind of approvals you need.

In terms of the way automation works, think of it as data flow all the way through. That's not super difficult, but it's certainly different, because you have to be very mindful in terms of who approves what, how you carry that data piece through, who modifies it, who validates it — if you only want to enter it once. And then you can think about where you can drop AI in to relieve some part of that administrative burden of triple-checking certain things, in order to start with the automation.

Shamil Malachiyev: And how much do you feel implementing AI solutions within procurement workflows is a good fit for what the AI technology actually brings?

Sören Petsch: We're in very, very early days. There are some really useful places to leverage AI. Analyzing legal documents, for example — because I'm not an attorney, but it's useful to have, let's say, a custom GPT or some step where I can give it my playbook, in terms of what I find acceptable: here's what a human reviews, and everything else, legal reviews. At least I want to make sure the AI assists me in that. That's super useful. Also pulling together multiple PDF files and entered data points to make sure the data pieces agree — the technology is very good at that. Just general research on companies — the technology is very good, obviously; it can go to the internet and pull together various resources. We're using it for our InfoSec review: we upload a SOC 2, and our InfoSec team has a built-in list of review pieces — the NIST compliance they're looking for — and it looks on the internet for whether there's been a breach or other notifications. So there are spots where the technology is very, very good, where an AI agent is really, really helpful and does speed things up.

But in most cases it's certainly not end-to-end — although we are building a fully autonomous, AI-enabled sub-process. No procurement person is replaced, and I don't really think that many procurement people will be replaced. But the skill set certainly has to change.

Shamil Malachiyev: Just to make sure we don't go too fast for some of the listeners, let's take a step back and think about how you would advise people to start visualizing what AI is. Right now there are so many ways to think about AI. And when you say AI agent — a lot of people don't know how to envision that within their processes. Some people might think: you have ChatGPT help you with things, and maybe that's already an AI agent. Or maybe you find RAG, and you can have a database. How would you explain it to other procurement managers out there who haven't made that step into researching what AI is? How would you help them differentiate between what AI currently is, where it can be played, and what it definitely is not?

Sören Petsch: The underlying assumption of my answer is going to be that you have orchestration software implemented — procurement orchestration software where the steps are outlined and you know what certain approvals mean. Into that, we have dropped some custom-built AI agents to automate certain steps. Think of it this way: we had a step in one of our processes where a person needed to look at what the vendor had entered as their legal name, address and so forth. The vendor also provided us with a certificate of insurance, and they provided us with their W9 — their registration papers. And we needed to make sure that all the legal-name information aligns. So instead of somebody doing this multiple times a day — looking at the various PDF files and what was entered — we built an AI agent to do that. It's tedious, nobody wants to do it, and it's faster that way.

We also have an AI agent for our InfoSec team, as I mentioned earlier, that we had custom built: when a SOC 2 Type 2 report is uploaded, the agent runs automatically. It looks for external information, and the input gets forwarded as an output automatically, based on our risk profile as a company. Think of that as a custom-built GPT with a few more bells and whistles. It's a RAG database, meaning when it evaluates security information it only looks internally, and then it does some very specific web searches externally, brings it all together, and butts all that information up against our risk profile. And that's really useful, because if I think about where the pain points are — InfoSec review and legal review have a tendency to take a bit longer, and they're a bit of a black box. Here we've built functionality that should speed that process up. Early days, but I would assume that where reviews first took eight to ten days, we're looking at three to four days.

Shamil Malachiyev: Let's say we've identified all of our processes, we found orchestration software, we wrote down all of the process, and we've identified where things can be automated — where they have folders to access to get the data, analyze it by your instructions and guidelines, and pass it on to the next stage. I guess a lot of people would say: wait, but we have a lot of proprietary data. This is secret stuff — confidential data that we don't want to pass to an LLM. How would you advise people to think about what kind of models to use? Do I use a local LLM that's installed on-prem, something open source? Or do I actually go and trust something like Anthropic's models or OpenAI's models?

Sören Petsch: Well, my perspective is that procurement data isn't super confidential data. It's not health data. It's not really PII. It's confidential, but it's not at the level of somebody's health record. But obviously, if you are concerned that there's some secret sauce here and we definitely don't want it to get out, you can stand up a local instance of an LLM — most likely an open-source model on your own hardware. With one of the LLM providers that we use, we have a zero-data-retention type of setup, meaning the company doesn't retain any data. And obviously we also have it in our contracts that they don't train on our data, that kind of stuff. If you don't trust "we won't train on your data" and you want to make sure it doesn't sit in their instance, you can stand up a ZDR — zero data retention — type of agreement and make sure that's appropriately configured. That's a conversation with InfoSec and with engineering; there are ways to build that. And I think that is a bit of an overblown type of concern, if you want my honest opinion.

Shamil Malachiyev: And what would you say are some of the main challenges? There are a lot of people who have the same task as you out there — thousands of companies, everybody trying to make their processes more efficient. Teams anywhere from one to maybe 20 people who have to orchestrate everything together to make all of this AI implementation work. What do you see as the main challenges these teams face when they have to integrate AI solutions?

Sören Petsch: If you're in an orchestration type of setup and you've articulated your process into certain approval steps, that's a really, really good start, because you're already thinking in a more production-line type of environment. What I mean by that is: most procurement teams, because they work in spreadsheets and email, work like craftsmen. And what we need to do is take that craftsman approach and take a look at how Henry Ford automated the production line. In our white-collar jobs, so to speak, we need to disentangle and be very deliberate in terms of what each step does and what the conditions are that drive it. It's revolutionary when you first do it, and then it's very evolutionary as you continue to tweak the system — no code, you can do that; I do that multiple times a week.

And then, on the question around automation, what has helped me is a slight reframe of the conversation. It's not about jobs. It's about tasks. When you think about tasks — what's this task here? Think of it as a really small step. Is this a task where I add value, or is it really just administrative? If it's highly administrative — like I mentioned earlier, there are multiple PDF documents and I have to look in the system just to validate that the information is correct — that feels like a really good task for AI. I trust my gut there. Then I think about: how many times do I do this task? If I automate this task, is there value — is it repeatable? As I told you earlier about the W9, COI and entered-information work: the person does that about 100 times a week. So if I can save two minutes of tedious work that's deeply unrewarding, I'm possibly saving three to four hours of somebody not having to do that. And by the way, once all the kinks are worked out, it's probably done better. That person's three to four hours can be deployed on something more productive than the tedium of cross-checking different files.

So that's how I think about it: tasks, repeatability, and then what can I do with that person's time. And there's plenty to do. I believe that 80, possibly 90% of procurement work is highly administrative. So there's a rich area of thinking about the various tasks and figuring out how you can automate some part of that with the help of AI.

Shamil Malachiyev: Not long ago I was visiting a legal company's office in Spain, and what shocked me was the amount of actual paper everywhere — stacks of documents all over the place. Right now I hear a lot of companies say the next barrier to AI adoption is going to be the data. Do you see that being a large problem for organizations? And how would you advise companies to look at their data and understand how good it has to be for an AI system to come in and help automate?

Sören Petsch: Yes, there's a lot of it. You have 50-page contracts with certain contract language that sits someplace in a shared file folder, and it's not necessarily clear that whatever you agreed to in the terms and conditions is actually being monitored — let's say uptime or certain deliverables, SLAs and so forth. That's not necessarily aligned or connected to the billing: what do you get invoiced at — the rate that's in the contract versus the rate that's on the invoice? So again, as I mentioned before, when you think about process automation, it's data flow automation. How do I make sure I get this piece of data — right now it's on a piece of paper or a PDF file — digitize it in a meaningful way, and have it available at the right time when I make a decision, like approving an invoice? There has to be a highly circular, so to speak, environment where the data pieces live, potentially in different systems, connected through APIs, and get enriched as the process progresses down the approval flow or the review flow that procurement supports. Again — very early days in how to think about it.

Intake-to-procure, which is the beginning of orchestration, is really meaningful just because you start digitizing the process. Somebody's request that usually comes through an email or a Slack message is now digitized: dates, dollar amounts, vendor records and so forth. That's a really good starting point. Then you add a P2P module to that: now you're issuing a PO for the approved spend, and the approved spend is already categorized in terms of which GL account and which entity it hits, fully approved. So when the invoice comes in, it can be matched up — by, let's say, an AI looking at the invoice and looking at the PO — it's pre-approved, and it goes through the payment process very easily. Then, if you have a CLM — contract lifecycle management — they usually have AI functionality, and having the information that's in the contract feed back into your orchestration system is very useful. As you think through the process, you start with the bare bones — spend gets approved — and then you start stacking on top of that, and there are many pockets of value: what's in the contract, what gets approved by InfoSec in terms of what flavor of integration we're OK with, how often we review those integrations, and so forth. These things stack on top of each other.

And I have to re-emphasize: we're in very, very early days. I don't think anybody has figured this out to the point where, of the 80% of administrative tasks I mentioned earlier, they have automated more than half away. I don't think anybody has truly automated more than 10 to 20% of that stack away. So there's a lot of runway for improvement. The point is you've got to try and fail in order to figure out where the ragged edge of AI is today, where it benefits you, and where you can find value. That may be slightly different for different people and different processes, but you have to spend a lot of time chasing the bright shiny things to discover: here's actual real value. You try different tools out. That didn't work. And then you come back sometimes and — now this works, actually; it's predictable. Ultimately it starts with orchestration, and then you can drop AI in in a really meaningful way. And you've got to keep trying, in terms of where the tooling is today and how it supports the process.

Shamil Malachiyev: We talk a lot about tools, and I was thinking it might be good for listeners to see: what are these tools, and what is good and bad about them? Obviously there's no one perfect tool for everyone — every single one has its pros and its cons. Can you talk about some of them, outlining the strong points of each?

Sören Petsch: Yeah. I want to start by saying that P2P — procure-to-pay — tools are not procurement tools. It's a finance tool that's being used by procurement, because procurement usually reports up to the CFO, the finance organization. And it's useful for them, because now they have spend approval as part of the P2P process — Coupa, Ariba and so forth are the tools there. But it's not a procurement tool; I just want to be very, very clear. Now, a lot of companies are integrated with P2P providers, because they are useful for invoice reconciliation and so forth. But it's not — at least not my version of — automation. These companies will tell you they have an intake-to-procure module bolted on the front of it. It's not best-in-class, in the sense that, at least from what I've seen and what I've heard, I don't think those tools are no-code.

Shamil Malachiyev: And with intake-to-procure, are we talking about the actual invoices coming through emails being read through with OCR, or —

Sören Petsch: No — intake-to-procure is basically: the request comes through that tool. Somebody says, hey, I want to buy this software, or I need to get this service stood up. Historically somebody sends an email — can you help me with this? Or: I'm ready to sign this contract, how do we get it executed? Intake-to-procure handles that whole process upfront: financial approval, legal review, InfoSec review, what actually gets approved. It's sort of a new software category — like a ticketing system for procurement. Zip was one of the first companies — I think they were founded in 2020 — that leaned in very heavily and created a custom-made-for-procurement type of intake-to-procure software. The industry woke up to them being valued at — I think they became a unicorn within a year and a half of founding. And all of a sudden everybody's like, oh, me too, because they realized: hey, there's something here. Then there are a lot of other companies in that intake-to-procure space that are pure-plays, and they're very, very good. I've looked at Opstream, Focal Point, Pivot, Omnea — they all compete in that pure-play intake-to-procure arena, and a lot of them are extending into the P2P arena, because you want to do the intake, what gets approved, and also the invoice reconciliation as well.

Then you have the Coupas and Aribas of the world, the P2P providers, because they have thousands of companies fully integrated. And there are other companies out there, like Oro Labs, that create a bolt-on to those systems in order to have a no-code experience. I think that's a really valid type of positioning, and those systems are awesome. They modernize and give procurement leadership the tools where they can actually own the configuration, and you don't need a ton of IT support to stand them up. Actually, in my last implementation of an orchestration tool, we had no IT support, but we did API integrations — call it four integrations, half an hour each. Done. Integrated to the ERP, integrated to our Docusign, and so on and so forth. The rest of the time we spent just configuring different workflows, approvals and logic conditions — basically expressing our procurement process.

And because it's no-code, you become sort of the systems admin. Because you have that ownership, you get better at configuring your systems, and you develop the language of what an efficient step is: additional steps, removing steps, changing conditions and so forth. These little tweaks — each is worth a penny, but you stack up enough pennies and all of a sudden you have a dollar. Those improvements add up to a process that's truly customized and continues to evolve with your own evolution as a company, in terms of what you need approved and who wants to see what. That is extremely powerful, and that technology is very, very good. And then once you have that done, you drop in AI at the task level — and all of a sudden you see some magic happen, where you see a glimpse of the future of what it can be and will be.

Shamil Malachiyev: If we look at the standard consulting models of change management, they always say: you come into a company, there are people who are champions of the change and people who are resistors, and usually you have 30% resistors, 20% champions, and the other 50% just waiting and going with the flow. If we're talking about automating processes within departments, I can imagine there's a guy called Billy whose main job is to read the document, compare it, put a stamp on it, take the paper from Sarah's desk to Mark's desk and put it there. And now: Billy, we need you to map out your process and start automating it. And Billy is sitting there thinking: wait a minute — if this is automated, what the hell do I do? It creates — I wouldn't say a rebellion, but a lack of motivation across departments to commit to this switch to a more efficient workflow automation. How do you suggest companies go about this, reasoning with people and explaining that it's not that black and white?

Sören Petsch: Well, change is always difficult, 100 percent — there's no two ways about it. The way I see it: there's a lot of tedium in our day jobs, and if I can reduce some of the tedium — like, you need to download this contract from email, save it over here on the shared drive, then upload it over there — that's tedious and not really interesting. I don't think anybody enjoys doing that type of work. You'd much rather use your brain: instead of doing all these tedious jobs, maybe I can stack some time together, eliminate some, and actually spend more time on analyzing data, making decisions, driving analysis. I think everybody would tell you that that is better. And somebody who says no — then maybe they shouldn't be on your team.

Where the change difficulty comes in is: how do you envision it? If you're a craftsman — your work is a black box, stuff comes in and stuff goes out, but nobody can articulate what happens in the middle — that's difficult to envision. What will that actually look like? For me, I went from managing 28 people at one company to managing one in my next job, in order to have the time to figure out how this automation thing actually works. That's where I implemented orchestration software for the first time. We actually started off by using a ticketing system and bastardizing it for the procurement process — which, essentially, is what these orchestration systems are. And then when I saw Zip — we implemented Zip, for disclosure, back then — it was clearly better. Instead of me trying to draw like Picasso, there's Picasso painting the picture, so to speak. It resonated very quickly. But I took the time to try to build an automation myself, to teach myself how to work differently. I read books on how IT and engineering operate — Scrum is one of the books I read, A World Without Email by Cal Newport, and a handful of other books around different processes and how to think differently — just to understand how to articulate the Henry Ford way of working in a white-collar job. And then we built it. I gained firsthand knowledge, and I work very, very differently today than I did five, six years ago. But I had to teach myself.

Now that I've implemented it multiple times and been in that workflow-orchestration-supported system, it's easier for me to articulate: here's how your job will change. It will always impact the AP team, the legal team, the InfoSec team. But having rolled out orchestration systems to thousands of people at this point, I've never once had somebody refuse to use the system, full stop. The change to an orchestration system is so intuitive to people, and they realize it's actually less time — less time spent searching through email and Slack messages, because the internal conversation all happens in one place — that they embrace it very quickly. But you want to be very mindful in terms of how you work with your stakeholders — InfoSec, legal and finance. You want to be very clear: this is your process, your approval process; let me build it in a way that works for you. You ask them what they need: who needs to look at this? What do you mean by this step, by this approval? That way you build it accordingly — is this the expression of what you want me to build? They see it, they try it a couple of times, and it's like a baby walking for the first time: they fall down, then come back up — try again. And all of a sudden, before you know it, they're walking with the system.

And here, just to illustrate the point — I just looked at this yesterday. Since we implemented our current system at the end of June, we've had over 42,000 approvals, 3,500 requests closed, call it a mid-nine-figure spend approved, we've routed our PO issuances, and we've had about 145 people in the system approving stuff. So it's a big impact. We're not a huge company — we're not Walmart — but it's a big impact on over 100 people. And all I did was hold learning sessions when we rolled it out, and then I created seven to eight mini videos, two to three minutes each: here's what the system is, here's what this approval means, here's how you read your request, here's how you can speed things along. Mini how-to videos, posted on a canvas in Slack, in our procurement channel. That was it. Obviously, if somebody has a question — how do I get started — we support them really quickly, but it's usually just a quick Slack message: click this button and follow the instructions. So I don't think this change is as scary as it could be. But you've got to get started, and that first step is always the hardest.

Shamil Malachiyev: And what I feel — maybe it's also on the personal level — is that whenever you do a process manually, you're in full control. You know exactly what came in, you've touched it, you saw it, you made a decision, and you hold the accountability to pass it on. When a lot of it is routed through a system — and especially if you have Slack workflows working there, Zapier workflows there, Zip workflows here, Salesforce flows there — you end up having so many workflows happening across five, six systems. How do you keep the overall control and visualization of what is going on with each of the instances? Which ones do we have to look at closer? Is there a solution out there that helps show orchestration visibility and maybe remediation?

Sören Petsch: Yeah. I think what you're alluding to is: there's information in different places — how do you pull it all together? That certainly is a bit of a challenge. But what the orchestration solution providers do is integrate your Slack channel or Teams channel with the solution. And then the key point is that, ideally, through APIs, you can connect with downstream systems like your ERP system or your Docusign system, so that the information lives in both places and is accessible in both places — whether you must look in different systems or you just want to be in one system. From a communication perspective, we did a fair amount of teaching people: hey, you want to use the comments section in the system. You'll get an alert in Slack, and you can click on it — but once you do, stay in the system. That way it's very clear: here's the conversation with legal, here's what was decided, and here's how that impacts the contracting. However many strands of information you can tie together, you want to do that. And even if you can't, you're still better off with a procurement orchestration system — even if you didn't integrate it with your ERP. It's not great, but you'd be better off, just because, again, you know what you're approving, and you have all that in one place.

But there is an opportunity for somebody to create a dashboard — AI-enabled — that looks at your email, various Slack channels that are not related to a request, your orchestration system, maybe other systems, where you teach it through prioritization, markdown files and certain harnesses to come to you — where you can just hit refresh and it tells you: of the 15 things you're handling today, here are the three most important; I've pre-drafted that email for you; take a look — do you want me to send it? That doesn't exist today yet, but I believe that is something we will be building here this year.

Shamil Malachiyev: And if we remove people from the loop — the automated process is only as good as the number of stops where somebody has to come in and have a look, and the process that runs automatically will be much faster. But that also raises the question: who do we hold accountable? Let's say a PO went through that in reality shouldn't have. Who are we going to hold accountable — the Zip developer, somebody who built the workflow, or a person who was supposed to do a manual check at some point? How do we solve that problem?

Sören Petsch: Yeah. There are different flavors of AI integration: there's human in the loop, there's human on the loop, and there's full autonomy, just simplistically. Ultimately, a human will be responsible. Has to be. And this is why there is a place for humans to do work — there are certain points where you have to be in the system in order to review, make decisions and approve certain recommendations. How that looks exactly in different parts of the process, I don't know. I think it has to do with risk profile: how risky is it if this goes wrong? You want to start with a low-risk type of profile.

So we are in the process of building a fully autonomous workflow, with the help of our provider, to collect certificates of insurance — COIs. I've been in procurement for 15 years, as I mentioned, and I've yet to be at a company where COI collection was done consistently, for all the partners we need to collect COIs from. The issue is that a certificate of insurance expires after one year, so you need to collect a new one — especially if you are, let's say, named insured; you've got to make sure that coverage is still there. And it gets done in a very ad hoc way. You may have some people at some companies collecting COIs for the high-risk vendors, but not for everybody you started off with a COI when you onboarded them.

So we're building a fully autonomous AI process: from the initiation — this COI is due to be collected — to reaching out to the vendor through email, to the steps for how to upload the COI documents; then it compares the two documents, old COI and new COI. The AI can take a look: the dates are different — that's expected. Is the coverage the same? That's expected. Close it out, and schedule it for next year. And I would deem that: if I go from zero collection to, let's call it, 100% collection, and it's fully AI-automated, I actually am better off — even if the AI were to make issues, like something not correct, I can still audit. But that's a place where I want to learn how good the systems are at automating. And once you do that, you can think about other parts of the process. We didn't just land here because we came up with this idea — we actually tested some of the functionality that we're intending to use for the COI build in a couple of other workflows, a couple of AI agents we had custom built, and we figured out it is able to do that consistently. So now, if I take that functionality and get creative about it, I can actually spin up this whole fully autonomous workflow. So we're excited about it. Again, it's one of those things where I don't think anybody knows what the future holds and what's going to be automatable, or how we will automate some parts of the process. But I feel like this is a process we have to support — and we currently don't, not fully, not consistently. So we're better off doing it. And I certainly wouldn't want to be the one — or the three, four, five people — who would have to collect thousands of COIs and chase companies down for them. So again, I think it's a very positive implication of how the technology can help us.

Shamil Malachiyev: Let's look at professionals like yourself, who are deep into analyzing and building — the architects of this new world of autonomous, or partially autonomous, departments. If somebody has fully automated one of the processes, it doesn't require all of your time to be focused on that specific department — if everything's running, you make sure there are no holes in the hull of the ship, no leakages, everything's working correctly. Could the future of these architects look fractional — where they take five, ten corporations and become like a command center for the procurement divisions in a particular vertical? Would that be something on the horizon for people who go very actively in this direction?

Sören Petsch: Yeah. If I were to say no jobs will be lost in procurement due to AI, that would be silly. I think certain tedious aspects of jobs will just collapse, and I don't think anybody will miss them, per se. The skill set of the people will obviously change. So now I am not looking for somebody who is very good at executing diligent, tedious work really well — I'm looking for people who are actually more decision-makers, who are comfortable querying AI and interpreting outputs and so forth. So there's going to be — I think large procurement teams will get smaller, due to automation and AI enablement.

But it also will create more jobs. Currently — and this is my own perspective — I believe a company has to be about 100 to 150 million in revenue before they even hire a procurement person, a single one. I believe that dollar amount will drop pretty significantly, where smaller companies will be able to drop in a certain AI, a certain orchestration software that connects to their ERP, pre-configured, and then you have a fractional CPO — chief procurement officer — that helps in certain aspects of the negotiation, while everything else — PO issuance, certain approvals — happens in a much more automated way. I believe that is going to be the future of the evolution. And when you think about it, there are so many companies that are small that don't have procurement at all. I think there's going to be a blossoming of openings in that single-person, lone-wolf type of procurement support environment. And if you continue that thought process, then you would also say there should be fractional CPOs — chief procurement officers who support smaller companies with a very high degree of automation, much more than larger companies may be comfortable with. And the outcomes are going to be better, because you have clarity around what gets approved, you have cross-functional approval, you have AI-assisted legal review, and the CFO just looks at contracts and says: this looks good enough. So I do think that there's going to be a shift in terms of the capabilities of the procurement professional, but also a much broader aspect in terms of how many companies will be looking for procurement professionals to support them.

Shamil Malachiyev: We've covered the professional tools people should be looking into. But in order for them to get comfortable with the vision of all of those tools within their working environment, I personally believe that people first have to get accustomed to using AI on a personal level. What would you say are the current tools that actually help you become more effective, structure your thoughts more easily, and think through different things? What kind of tools do you use to accelerate yourself?

Sören Petsch: I mean, it's amazing what's out there. So I use all the AI tools — that sounds silly, but the ChatGPTs, Claude, Grok, Gemini — I have a subscription to all of them and I try them all out. I oftentimes run the same thing with them, just to see how different they are, because there's different tools for different purposes. There's a lot of hype around Claude Code, and it's truly amazing — especially once you have figured out the power of markdown files, and you build your own subset of your folder structure, and you have your AI integrated with that, and then the AI saves its progress updates there and it learns from what it did before. Very powerful. And fun — I mean, really fun. I am dabbling a little bit with the whole OpenClaw kind of setup. I am not enough of a technologist to really build that safely. But in the meantime, OpenAI — with, I think, ChatGPT 5.5, which just came out yesterday or so — has a lot more agentic functionality, and they hired the guy who created OpenClaw. So everybody in the AI space is on that bandwagon of creating agents: how do you make sure that the agents have the right context window, how directive are you in terms of what you want, how specific are you, how do you break jobs or tasks down?

And there's a plethora of really interesting research papers that I stumble upon on Substack and X, which I use NotebookLM for, in order to break them down for dummies like me — to take that technical talk and translate it in terms of where the technology is actually going. That is a fast-emerging field, and I just use personal tasks in order to learn a little bit more about agentic AI capabilities. Early days — but I want to learn about it, be able to articulate how it works for me, and ideally keep stacking the pennies to two dollars and more.

Shamil Malachiyev: And for the listeners — it's important to know not to install OpenClaw onto your workstation. Your working computer is probably not the best place currently; maybe in the future it's going to get more secure. And let's say somebody's listening, and they're in procurement, operations or legal, and they've listened to this conversation and they're like: OK, yes, we saw the tsunami coming, we want to keep our place in the industry, in this vertical. For those who haven't yet started moving in these directions — what's the first thing to do on Monday?

Sören Petsch: Start doing something. I challenge myself to try something new with AI every day. Teensy tiny. Just try it and see where it leads. It's kind of like Hansel and Gretel following the breadcrumbs into the forest, to the house of the witch or whatever. Nobody knows the endpoint — we'll probably never know the endpoint. So it's a journey, and then follow the breadcrumbs on that journey. And I don't think you'll lose anything, and I think you gain a lot by just playing around with the technology and figuring out: it can do this. What's a RAG database? What does NotebookLM do for me? And then all of a sudden — NotebookLM: I've used it as a learning tool, just as an insight, for probably a year and a half, and I'm going to use NotebookLM as the training tool for my team as I onboard new people. Because I can build what I want to communicate, and then there are different flavors — videos, podcasts, quizzes, flashcards and so forth — that you can use to supercharge your own learning. Instead of me giving you a stack of documents, I give you the choice of how you learn, yourself and by yourself.

And then ultimately, I think maybe the big insight is: I believe the usage of AI requires more initiative. Don't wait for somebody to tell you how to use AI. Ask that question yourself, and you'll be surprised at what comes back — and how many more questions there are.

Shamil Malachiyev: Awesome. Well, I want to thank you so much for coming on the show. I think it's going to be so beneficial for people to feel inspired by seeing somebody take so much action — to know what is possible, that you just have to start playing with it and be curious about all of the AI news and not be scared, because what's coming is real, and it's OK to start slow and get to the point. Because as you've mentioned a lot of times, we're still very early in the adoption process. And even though everyone I talk to feels like "I'm behind" — everybody's behind, nobody's behind. We're all getting to the start point of the same race, and I think it's up to us to do our best to make sure we're prepared for when the race starts.

Sören Petsch: Yep, 100%. In the spirit of my early example about automation: 100% of people are talking about it. 4% of the companies are actually doing something about it. So start playing and start learning. And you'll be surprised at the ownership you can take — and then ride the wave, because the wave is coming.

Shamil Malachiyev: Perfect. Well, thanks so much. It was a great pleasure, Sören.

Sören Petsch: My pleasure, Shamil. I appreciate the opportunity to speak with you.