AI is not a coat of paint: Why Enterprise AI success requires a process-centric approach

https://delivery-p141552-e1488202.adobeaemcloud.com/adobe/assets/urn:aaid:aem:704e95d2-a75d-433a-aaa8-85644120aa0d/original/as/Brown-Pickup-Truck-Unsplash.jpg

Simply layering AI onto existing processes often fails to deliver meaningful value. To unlock better business outcomes with Enterprise AI, organizations must rethink their workflows and free their processes from the confines of the systems on which they run.

In this special Celonis interview, I speak with Vas Moza, CEO of Varick Agents, about his article, Don’t Apply AI. During our conversation, we explore why “applying AI” falls short in the enterprise and how organizations are unlocking greater value with a process-centric approach.

The following is a transcript of the interview edited for readability.

https://delivery-p141552-e1488202.adobeaemcloud.com/adobe/assets/urn:aaid:aem:6a18cb4b-d574-484f-9cfa-b9575fe30e29/original/as/261001_EAI_VasMoza-Interview_YouTubeThumbnail_1280x720px.jpg
https://celonis-videos.wistia.com/medias/t500n8tplp

Why is layering AI over traditional ways of working ineffective?

Bill Detwiler: I want to kick off our conversation with a two-part question. First, why is simply layering AI over traditional ways of working ineffective? And second, how does incorporating process redesign as a core part of an AI transformation, how does that unlock greater value?

Vas Moza: Yeah, I think that really is the billion dollar question. So the first part is why does applying AI broad strokes not really fundamentally change the business? It's because, let's take an example for a Finance department workflow, for example. If you have 15 steps in the process and you use AI to speed up every individual step in that workflow, what you'll still have is the cycle time, the drag between each one of those steps.

So even if you have Claude Code, I have Codex, and a third guy has GitHub Copilot, and AI is doing our work faster, it's not actually accelerating the handoffs between each of us, for example. And this happens hundreds and hundreds of times over in the average enterprise. So applying AI as sort of a coat of paint to an enterprise-wide process is not really going to give you the same results that you want.

To the second question of how does process redesign fit into this? It's a lot of understanding what are the steps at play today and what bucket do they fall into? Is it something that can be solved with deterministic software? If this happens, then this should occur. Is it an AI or LLM call where judgment is required? Or is it human in the loop where it's too risky to have an agent take over the entire task and instead we want to have a stop gap so that if you're sending out a payment or you're approving a payment, for example, you have a human in place to make sure that stopped before it goes through.

Bill Detwiler: Yeah, and I want to get to that. I know we're going to talk about that in a little bit, and I love that analogy, the applying AI as a coat of paint as being ineffective because as a little bit of a car guy, I think about it as applying paint over rust. The rust is still there. You just covered it up or papered it over, but you've got to fix that rust before the new coat of paint really delivers its full benefit.

How does incorporating process redesign unlock greater value?

Bill Detwiler: Before we talk a little bit about how you decide which steps within the workflow need to be redesigned and how, what they need to be, what's the best way to approach them, deterministic, agents, human in the loop, I'd like to get some perspective on your insights from being on the ground from organizations that you've worked with. Can you give us an example of one that recognized this opportunity to reimagine a legacy workflow during their AI journey rather than just, like you said, speeding up individual tasks?

Vas Moza: Yeah, absolutely. The department that I would say we have the most experience in is absolutely Finance departments as well as Operations, Sales, et cetera. So I'll continue to give an example in the finance department. If you have, for example, Accounts Payable process. This was a large public SaaS company that we did this workflow for, so you can imagine they had tens of billions of revenue, thousands of employees, I think over a thousand in the Finance department alone. It's funny, I say AP and everyone expects AP to be solved because of how simple in theory it should be, but the reality is it's not at most enterprises, whether on Dynamics or NetSuite or any of the large ERP players.

So, the Accounts Payable process is very complicated for a variety of reasons. The first is invoices can, for example, come in a variety of different shapes and forms. It can be a PDF, it can be an Excel spreadsheet, it can be an image, it can even be a scribbled down piece of paper. We've seen it all. And the issue is that oftentimes it comes in from a variety of different places. So that might be email accounts that are disparate, spread out. It might be sent to a certain ledger where stuff happens because it's outdated. These companies are also large enough to where they have dozens of entities around the world. So you have one company that is in America or several subsidiaries, one that's in Europe, one that's in Australia, one that's in Argentina, one that's in India. And this means that each one of them has their own process, their own software, their own tooling, their own ways of doing things. If you were to say, okay, let's apply AI everywhere, you're going to be left with a mess. You're just going to be making that pile of trash faster.

This company, we are fortunate enough, was very willing to have us come in and redesign the entire process from the ground up.This is what we do in our "audit phase" where we have forward deployed engineers come in, map out how things work at every single different region, every single subsidiary entity, and try to consolidate not on top of a single system of record because that's oftentimes impossible, but at least in terms of the same process. So when an invoice comes in, these steps will take place. Then you have to go into line item parsing. You can use OCR, you can use a visual language model to parse out items, then you can use GL coding to make sure that every single line item is appropriately matched to the right book, for example. And that step can be done agentically.

Then it passes on to this single group of people who understand the process. It's all around the world. They go to this single pod and from there it flows out as needed. So all of a sudden you have a 15-step process in one region, 12-step process in another region, now becoming the same six-step process. Several items were deleted, several items are now agentic or deterministic, some human in the loop where needed for approvals. And this company was very willing to let us come in and redesign that. Does that make sense?

Bill Detwiler: Yeah, no, absolutely. And I'd love to drill down on either that example or another one because what I want to do is...that's a really great example of the how, how you get there. So I'd like to focus on the outcome, what successful execution really looks like in practice.

Example of an organization that reimagined a legacy workflow during its AI journey

Bill Detwiler: Could you share an example of a company that achieved those meaningful business outcomes, whether you were talking about Finance, AP or AR. How did they achieve what their goal was? What did that look like on the back end by doing what you just said, combining that process redesign directly with AI solution deployment?

Vas Moza: Yeah, absolutely. So interestingly enough, data is incredibly unknown to these companies. It's invisible to these companies. So for example, this company thought that their invoice, the cost of handling each invoice was $5, and we actually came in and showed them the real cost was $31 to handle a single invoice. And you can imagine there's thousands of these invoices per week or per month or whatever that is. The result of our coming in, redesigning the process, building and deploying agents was bringing that cost down to $6. Now, the funny part is obviously if you went with the number they had, we increased the number from $5 to $6, but the reality is we brought it down from $31 to $6. And there's a bunch of other metrics that we helped show them. There's 17 steps in the average AP process. Now it became seven steps. There's 24 days of cycle time and we turn that into 6 days. The straight through rate went from 18% to 87%.

And this is super powerful because companies have very little visibility, especially C-suite, CFOs for example, have very little visibility into how their organization is actually running. Is it humming? And we deliver this through process mining, remapping of the workflows, baselining KPIs, then showing them, okay, this is what the agents are delivering in production. So those are some examples of the outcomes that we've delivered for finance workflows, for example.

Achieving meaningful business outcomes by combining process redesign with AI solution deployment

Bill Detwiler: And let's talk a little bit about that operational intelligence. You and I were talking before our interview, and let's talk about the importance of understanding what you just said, how work actually flows through an organization for Enterprise AI success. So as part of an AI transformation, how do you evaluate a workflow to determine which steps should be handle—we were talking about this earlier—deterministically if this, then that, which should belong to agents which can handle it a little more ambiguity. We were talking about unstructured data, agents can handle a little more decision and then where human decision making really provides the highest value. How does your team go about doing that?

Vas Moza: Yeah, so it's funny, I think Palantir popularized the idea of an ontology, which is basically creating a mapping of your entire company. And now we see that that is a very clear prerequisite to AI implementation. So what our team will do is we'll come in with forward deployed engineers on the ground for anywhere from four to six weeks for an entire department and we'll conduct interviews, yes, because tribal knowledge always exists in people's heads. We get this with every single company that we talk to. Yes, we do process mining agents on top of their systems of record. Yes, we do screen reporting agents as needed to map out their desktops, their computers to understand what the employees are doing. We also do document ingestion for their SharePoint, their Notion, their Google Drive, the process mining on their Salesforce, their NetSuite, their Dynamics, et cetera. But without also combining that with human interviews, you are always going to be missing the full picture.

And I think that's the beauty of what AI implementation, AI transformation should be is getting that full view, those three angles. From there, what we'll do is actually a lot of co-creation with the company, with the stakeholders. So a lot of the reason that AI implementation, AI transformation has failed in the past is because stakeholders are not brought along that journey. The head of accounts payable, for example, will be very against the solution that these "consultants" or forward deployed engineers propose. If you're basically ripping apart their baby, which they've been running for 10 years and redesigning it, they should have a say in the matter, especially if they're the ones who are expecting to use this agent system in the future. So a lot of it is co-creation with them, spinning up workflows, UIs for them to respond to, and also making sure that we build inside the systems of record.

So we're not here to make a new surface for you to interact with. That's the feedback that we've gotten from a lot of our clients. They would rather us build the agents inside their NetSuite, inside their Salesforce, inside their dynamics, and then when they need to get pinged, it's in their Slack or in their teams. And then finally, to your point of how we decide what should be deterministic, agentic or human in the loop or flat out deleted, oftentimes if it is strictly cycle time, meaning you're going from one team to the next, we view that as a step that should be deleted. AI should be the reason that you consolidate the amount of people involved in a workflow, not the reason that you keep one person for GL coding, one person for invoice approval, one person for line item parsing. That's how you're left with the same length and cycle time of the process.

So those steps are very frequently deleted. Now for deterministic code, it's in the name, it's where it's a very clearly defined set of steps. So if you have an invoice come in, we need to send it to this group of people in Slack, for example, that's deterministic. There's no need for an agent. There's only one group of people, there's only one type of invoice, send it forward. And this is very common in, for example, invoice parsing where that can just be a simple call to OCR. You don't actually need to call an LLM here. Now there's benefits and drawbacks, but that's a conversation for another time. Agentic calling is where you have enough historical data and slight judgment involved to where you can make a reasonable assumption as to what the outcome should be. GL coding is my favorite example here. If you have a thousand past examples of, hey, staples or iPhone or desktops or monitors that are bought for the office, you know what category, what book those should be applied to.

And then finally, human in the loop is where there's either not enough past historical data or there is too much risk involved in using this step, so you're going to have a human to sign off on it before anything happens. I know that was a mouthful, but that's kind of the rundown.

Bill Detwiler: Yeah, no, and I think that last point about risk is really important. But you talked a little bit about the example that you gave before about a large enterprise with multiple systems spread across different geographies I think is so true. So many leaders assume they need to consolidate on a single ERP or CRM before starting an AI initiative. In my experience, before I was in the media and at Celonis, I was in IT myself, and so this is neither possible nor really practical, especially today. And AI makes this, or is helping make this, unnecessary.

How do you determine which process steps should be handled deterministically, by AI agents and by people?

Bill Detwiler: So why is establishing a unified process definition, you talked about that, across these businesses and using agents across existing systems, why is that a more effective path to value? So even if we say, "Hey, look, you're going to throw a billion dollar ERP transformation out there over the next five years." It's going to take five years. And that may be necessary. You may have to do that, but you don't have to always do a massive rip and replace. So how is establishing that unified process—a better process like you talked about where you're deleting things, where you're applying AI appropriately, where you're applying humans in the loop appropriately—how is that more effective than trying to consolidate on one platform before you start?

Vas Moza: Yeah, I mean, your point is so true. It's not even just about the cost, it's about the time it takes. We have a client that spent five years migrating to their new ERP, and that was probably on the lower end, if we're being honest. So you can't afford to wait five to 10 years or whatever, however long it takes to move to a new ERP. And the reality is everyone hates their ERP, but you're going to move to the next one and you're going to hate that one too, so what's the point? But jokes aside, process re-engineering is super crucial here because, or I guess process consolidation as well, because one, it's strictly unrealistic to expect every single entity to be on the same ERP. If you have one entity on NetSuite, one entity on Dynamics, good luck, it's over. You can throw money at the problem, you can throw years at the problem.

At the end of the day, you're going to be several years behind your peers who said, "Ignore that. Let's just implement AI where we are today." And the reason why it's so powerful is because the features that you expect from your ERP are going to be generic by nature. If this is a software that's sold to everybody, there's going to be a bit of customization that you're going to want no matter what. And that's why everyone hates their software, it's because it doesn't match their exact needs. So instead of trying to move to one that does, because you'll never find it, just fill in the gaps and just have AI, to your point, help these systems of record talk to each other. One of the reasons why AI is so powerful is because it's the first time that software can enter what I call a fluid state, where it doesn't need to be rigorously fitted in and outputted out.

It can take differently structured data or unstructured data and make it work for itself. And that's the real unlock for any system of record, any ERP, for example, is having AI do intake and output and allowing it to be a single pane of glass across your entire software stack. We had a client that had over 120 pieces of software in just their sales and marketing departments, and AI was the single pane of glass that allowed them to not have to go into every single one of them or worry about how they all talk to each other, and that's how AI was most beneficial to them.

Bill Detwiler: It's so true because what you said is having been in IT for decades, vendor lock-in was my enemy then, and it's not just vendor lock-in, but it's system lock-in with the processes. You really want to free your processes from the confines of your systems. Your systems are there to serve the processes, but so often it's the other way around. You have conformed the process to the system that's being used to run it as opposed to saying, "Hey, here's the most effective," like you were talking about, "The most effective, the happy path. Here's the most effective way to get this done as opposed to the one that just conforms to the system that you're using to do it." So I think what you said is so true.

Who are the right executive champions for process-centric AI transformation?

Bill Detwiler: Now, one of the things I'd also love to ask you about as we in the final minutes here is you touched on something that I think is really important, and that is about, you talked about your forward deployed engineers, Celonis, we call them our value engineers, that really go out and work with companies to understand their needs, to get that knowledge, that tribal knowledge that's locked away into people's minds so that they can really understand how work flows.

So who is it, in your experience, as you work with customers, who within the executive leadership team really needs to be a champion for AI or for integrating AI transformation so that teams really feel empowered to optimize their workflows, especially when it spans multiple departments? Where does that sit? Obviously the CEO, you need top level executive buy-in, but below that, you talked about CFO, we talk about the CIOs, who are those key people there? And then even a level or two down, who are those key people that really need to be those internal champions to make this work?

Vas Moza: Yeah, it's a great question. It's like you said, we need senior buy-in because one, the workflows span across departments. If you have an HR workflow, it spans IT, Finance, HR, so all of a sudden you're touching a lot of people's territories. But we've also been brought in by people like directors, VPs, MDs who then help or we help them work their way up the chain and get approval from who we need to. But the short answer is the more senior you go, the better because this is a really intense initiative. And oftentimes we'll land with say the CIO and they'll bring in their head of AP or a head of a process and we'll start there, but we need to know we have their buy-in because it's an involved process.

How do you engage front-line operators during an AI transformation?

Bill Detwiler: Yeah, absolutely. And then how do you engage with those frontline folks during this transformation so they actively participate in redesigning their workflows? Like you said, they're the ones that have to use the tools every day. They're the ones that are going to need to use this new process that you've set up, even if it's with tools that they're already familiar with. How do your people, your forward deployed engineers, when they go in as part of this audit during the transformation, how do they actively involve those frontline operators?

Vas Moza: Yeah, this is where I would say I'm the most prideful. I would say our forward deployed engineers are the best of the best, so forgive me if I come across a bit biased. It really is a lot about communication, right? We meet them where they are. We meet them face-to-face in their office, and we walk them through, "Hey, AI is not here to take your job. It's here to take away the things that you don't want to do," which very often resonates with them. There's always something that an operator just wishes they could hand off to somebody else. It's just a drag. It's the same thing over and over again, and they'd truthfully much rather use their intellect for higher leverage, higher skilled tasks. And we try to show them the light of this is what our plan is to help you do so. Finally, it's also co-creation with them saying, "Hey, look, this is what an error is costing your organization, your team, your department, your company, and here's how often this happens.

Here's the volume, and here's the revenue uplift that we can generate based on agent implementation, and here's the cost savings that we can save, give your department back some budget." And this also really resonates with them. So we co-create the agents in front of them, very quick feedback loop saying, "Hey, would you use this?" And they say, "No, actually I want to change this." And they feel like the owner because at the end of the day, they're the ones designing what this looks like in conjunction with our FDEs. So that's how we co-create and make sure everyone's on the same page so that by the time deployment happens, there's no surprises. They actually are looking forward to it.

Bill Detwiler: Yeah. Ownership and agency is so powerful. Well, Vas, this has been a really great conversation. I don't think there's any doubt, at least for me, that a process-centric approach is critical for any AI transformation, and it enables you to go from all the AI hype and uncertainty to meaningful value that compounds.

Editor’s note: Header Image Credit - Photo by Nikolay Loubet on Upsplash