Skip to content

ON - Account Categories & Forecasting Data Sources - 2025-05-27

Metadata

  • Date: 2025-05-27
  • Company: ON (On Holding)
  • External Participants: Rodrigo Cabrera (Treasury), Yulia Ershova (Treasury)
  • Palm Participants: Christian Sobkowski
  • Type: Customer Call
  • Domain Areas: Categorization, Cash Forecasting, Variance Analysis
  • Recording: https://tldv.io/app/meetings/68357f429bc592001375bc18/

Summary

Context

Regular customer call with ON treasury team (Rodrigo and Yulia, with Amanda on holiday). Christian introduced the new account-specific categories feature and discussed forecasting data source integration plans.

Key Discussion Points

  • Account-specific categories launched - new feature allows defining category sets per account for better accuracy
  • Currently 5-10% of IC and receivable transactions are miscategorized; LLM gets confused between cash pool vs regular IC transactions
  • Forecasting data sources identified:
  • Hex data lake: AP/AR open items (~500k lines, 10+ weeks horizon)
  • Google Sheets: 3-4 sheets used by different teams (tax, talent/HR, etc.)
  • Salaries are repetitive with same description/timing - good for ML
  • Hex file confusion: Palm was only seeing short-term data (<1 week horizon), but Rodrigo says there should be 10+ weeks - needs fresh export
  • ON's timeline: Q3 for forecasting go-live, June busy with TMS (Kyriba) work
  • Variance analysis gap: Yulia notes no tool in market helps identify forecast vs actual differences - very time-consuming to explain to management

Pain Points

  • IC categorization confusion - even Yulia says "I have the same problem as the LLM. I don't understand what is what" in intercompany transactions
  • Multiple data sources - forecasting data scattered across Hex (AP/AR), multiple Google Sheets per team, ad-hoc email/Slack requests
  • Variance explanation time-consuming - management always asks "why are we off?" but no tool helps identify the drivers
  • No market solution for variance analysis - Yulia asked competitors at Mannheim, they said "we haven't started"

Feature Requests & Needs

  • Account-specific categories for IC transactions (priority)
  • Ingest Hex AP/AR data into forecast
  • Ingest Google Sheets from various teams
  • Better forecast vs actual variance views with drill-down to understand what drove differences
  • Decision support: ML forecast vs known internal data - let user pick

Jobs & Desired Outcomes

Job: Explain forecast variances to management when actuals differ from forecast

Desired Outcomes: - Minimize the time required to identify what drove the variance - Reduce the effort to explain forecast misses to stakeholders - Increase ability to know how to amend the forecast based on variance insights

Job: Combine multiple data sources into a reliable short-term cash forecast

Desired Outcomes: - Minimize manual consolidation of data from Hex, Google Sheets, and other sources - Increase the accuracy of known future items vs ML predictions - Reduce uncertainty about which data source to trust for each category

Domain Insights

  • Data source hierarchy at ON:
  • Hex/data lake: AP/AR open items (structured, high volume)
  • Google Sheets: team-specific (tax, talent/HR, ad-hoc payments without invoices)
  • Salaries: repetitive, same description, same time of month (good ML candidate)
  • Talent payments vs salaries distinction: talent sheet covers everything except actual salaries
  • Teams track non-invoice payments in Google Sheets with: recipient, amount, currency, date, description, completion status
  • Payment patterns are often repetitive ("last year this was done every other quarter")
  • Hex data currently connects to Kyriba

Action Items

  • [ ] Christian to share IC category mapping proposal via Slack (async work)
  • [ ] Rodrigo to upload fresh Hex file to shared Google Drive
  • [ ] Palm to create forecast vs actual variance view (can start with ML forecast before Hex data)
  • [ ] Cancel next sync session (long weekend - May 30 holiday)

Notable Quotes

"I have the same problem as the LLM. I don't understand what is what?" - Yulia Ershova (on IC categorization complexity)

"What is also don't see in the market... there is no tool or solution which can help identify the differences what was forecasted and what was actual... it is very time-consuming to find out what went wrong. This is what management is very much interested in always." - Yulia Ershova

"A lot of these payments are repetitive. So based on the description in the Google sheet, we could see okay, last year, this was done also every other quarter so maybe we can forecast on those descriptions." - Rodrigo Cabrera

Full Transcript

Date: 27/05/2025, 11:00

00:00 Rodrigo Cabrera: So no and everybody speaks English. Also that doesn't help.

00:00 Christian Sobkowski: Of course, not. I'm the same student I haven't picked up, Swedish

00:06 Rodrigo Cabrera: yeah, it's hard, especially well if that's one that one's even worse now because you will only use it in Sweden,

00:12 Christian Sobkowski: Elio. Yeah, yeah, I guess so.

00:15 Rodrigo Cabrera: There might think you can use it in a few other places, but

00:17 Christian Sobkowski: Yeah. You're right, 100%. How with you move from to Switzerland?

00:26 Rodrigo Cabrera: I'm originally from El Salvador, but I was living in in Colombia right before Switzerlandia.

00:34 Christian Sobkowski: Tim, very cool. Very cool. Hey, Julia, how are you doing?

00:39 Yulia Ershova: Hi, good morning, little fine. I I went off as I couldn't find the room, so it might be, I get in some noise, I will be in the majority of the time.

00:47 Christian Sobkowski: Oh good.

00:48 Rodrigo Cabrera: It's okay.

00:49 Christian Sobkowski: That's all good. Cool. Should we kick it off?

00:57 Yulia Ershova: Yeah, Mondays on holidays that week. So yeah.

00:58 Rodrigo Cabrera: Yes.

00:59 Christian Sobkowski: Okay, okay, she's back next week.

01:03 Yulia Ershova: Yes.

01:03 Christian Sobkowski: Amazing. Nice. Hope she has some well deserved time off. Very cool. And hey I think Ed, I want to get It's me, maybe feel free to if have any point I'm saying something where you're saying, Hey, we're not getting value out of this. That let me know if you're free to steer these conversations, right? I want to.

01:33 Christian Sobkowski: communicate one thing, we just launched and how that would How I think that will benefit you and we can we can chat about how that works. And then I want to quickly Figure out the focus for the next few. Next one, two weeks on where you get the most value from from us.

02:00 Christian Sobkowski: Maybe starting with a feature and so one of the things we've done is and We we launched something called account specific categories. So what that means for us is So far what we've done.

02:16 Rodrigo Cabrera: Panic.

02:18 Christian Sobkowski: for clients is, Right? The the whole we were looking to improve categorization in terms of you know, with with Llms hoping that you can get more accuracy, you're getting more flexibility. And ultimately don't have to administer a long catalog of things. M. D, The way this worked so far is that you set basically global categories that work across your business.

02:47 Christian Sobkowski: Which you also have we've aligned on those and But they were count applied across accounts. And, and they They were customized on an account level. now, what we've done now is that you're able to basically if you want define for certain accounts, a set of categories that apply which, Would we've seen and testing is increasing accuracy of categorization, quite a bit.

03:17 Christian Sobkowski: If set up correctly, I'll give you an example for that. So, I know. I'm there, there is a bit of I think there's probably five to ten percent of intercompany and receivable color transactions for you that are miscategorized in the palm platform right now. Especially for inter company. I think one of the things it in LLM gets confused about is things such as differentiating between cash pull and then into a regular company transaction.

03:52 Christian Sobkowski: Where even within the inter company transactions, you have a bit of a hierarchy and a knee sat where you're differentiating even more And we weren't really able to support that properly because Clm gets confused. Now for you it's very obvious that in the cache pool, you won't have into companies transactions.

04:17 Christian Sobkowski: You will have Cash pulling transactions. and something that would be one of the obvious things where we could We could roll that out for you and if you map that out and I think the other one is things such as wholesale, retail ecom collections for you, where I think you, you only have some accounts where you're for example, and settling Econ into.

04:47 Christian Sobkowski: And if we give that that type of context, you should see in pretty big increase in Accur. On those data sets. And that's life in the app right now, it's gonna get live in the look and feel for you in the coming weeks. So you would be able to self-serve that What's life right now is that we can already set it in in the backend for you.

05:19 Christian Sobkowski: So if we started mapping it out, we could already test it out or really, actually map it out properly across your account. and I think the, the question just where, where you would most benefit from it,

05:41 Yulia Ershova: Payment from which product, right? Like in the company or anything else.

05:44 Christian Sobkowski: I think so. Yeah.

05:47 Yulia Ershova: I mean, so far from my perspective, I would say actually enter companies quite good because or quite important because I have to say is in u to the company. I

05:52 Christian Sobkowski: Okay.

05:59 Yulia Ershova: have the same problem as a limb. I don't understand. What is, what is what?

06:04 Christian Sobkowski: Yeah.

06:06 Yulia Ershova: So, from my perspective, this is something that would be good. That the system does because I don't think every system can do. Kind of as precise. So, any improvement there, it's absolute desirable.

06:21 Christian Sobkowski: On into company, amazing. Cool. I know him maybe help me out them so and let's do, let's fix this for into company. Who should I? who should I best work with in the team on this so specifically, right, I think I would love to share a few in Excel sheet with a few things.

06:47 Christian Sobkowski: That For my understanding of your business we should be doing and but we'll probably need a quick back and forth with someone team. To set this life.

07:01 Yulia Ershova: I would say it should be Amanda at the same time, I would like, also to join the

07:03 Christian Sobkowski: Okay.

07:06 Yulia Ershova: calls just to see what the system can do. And also it is a kind of like a deep

07:08 Christian Sobkowski: Of course.

07:11 Yulia Ershova: dive for me, also in our room for company.

07:13 Christian Sobkowski: Amazing.

07:14 Yulia Ershova: I don't know Rodrigo, if you also like If there is a benefit for you to join, would you like? Or you think about you think that I'm sure the main content?

07:24 Rodrigo Cabrera: yeah, I was going to Put myself in. But yeah, that also worked it. If I'm Anna, can do it detail. So

07:32 Christian Sobkowski: Got it.

07:35 Rodrigo Cabrera: it's also good. I I think we all know a little bit. Oh yeah, there already. If we want to start this soon, I can I can join. And once a month that comes back, we can mostly transition to her. So, but if this is ready now and something that we can start working more, collaborating together, we I can, I can do it and then We translate into her.

08:02 Rodrigo Cabrera: Yeah.

08:02 Christian Sobkowski: Okay.

08:04 Yulia Ershova: Just not to so that we don't do the double work, right? Like in case if there is a one presentation so that she's up to speed. I don't know if there's an emergency from work from one week. I also will leave then it on YouTube to decide 30. Though is the public holiday? I don't know.

08:21 Yulia Ershova: Christian, do have the

08:20 Rodrigo Cabrera: Yeah.

08:22 Yulia Ershova: same or

08:23 Christian Sobkowski: We have the same.

08:24 Yulia Ershova: Okay.

08:24 Christian Sobkowski: We'll see. Let's and what we can do is with this one. I I don't think this needs to happen in a call. I think this is work that happens. Much better on on slack. So um let me let me share some thing this week and and in this in the open Slack Channel, Rodrigo, I'll continue and then super happy to fill in Amanda as well.

08:53 Christian Sobkowski: Next week, is that fair? And then the searching session I will cancel given. It's it's long weekend and let's do the work async on slack.

09:05 Yulia Ershova: Yeah, sounds good.

09:06 Christian Sobkowski: Cool. Okay. Then I have the first action point. and the, the second action point is and from, from our side, the It was tease a bit of slack and Rodrigo, I think you, he reacted to forecasting already in A in a positive way and so we We're building a fair, few use cases at the same time right now around.

09:37 Christian Sobkowski: Sort of we we're doing a bit of dual tracking as a platform on the one hand, we're We're building more use cases for. Some of the Treasury specific, but non-cash things. So that's everything from with X investments. I'll class into company on there. It's it's cash, but it's very specific.

10:03 Christian Sobkowski: And that's a lot of views, and intelligence, and analytics, and those things. And the other thing that we have a big focus on for the rest of the year is forecasting and and you know, right now the platform does generate machine learning forecasts but really that's not good enough to use that in production on a day-to-day.

10:32 Christian Sobkowski: level for you and and So, short term, there's three things. We're doing a we're Building functionality for you to pick up existing data from the organization. and that you already have in terms of known future things that can be and what do you often see is a Google sheet where for example the tax team plans, future payments And that can also be things such as known future receivables, out of an ERP system or out of the data lake.

11:13 Christian Sobkowski: and then crucially sort of building functionality, where you basically still have machine learning, is based you get to pick Whether you have better internal data. Then the machine learning model would give you. And so, giving you giving you a bit of decision, support in the platform for that. And then better forecast, which is actual views on both your existing data.

11:44 Christian Sobkowski: So The appear receivables, for example, or the machine learning model. And with the idea to to create much much stronger. Certain week forecast for you. I would love to to start working through that with you and the team. And does that align with with yourself plans for the next? Three plus months.

00:00 :

12:14 Yulia Ershova: Yeah, absolutely.

12:16 Rodrigo Cabrera: Yeah. Yeah.

12:17 Yulia Ershova: I think from our side it might be the June is a little bit more more bad because our like TMS we need to push certain things to go live at to bonus. Forecast is

12:26 Christian Sobkowski: Yeah.

12:29 Yulia Ershova: is anyway in our timeline for the Q3 to basically to go live Q3. So we are absolutely happy to work on that as well.

12:39 Christian Sobkowski: Amazing. so then I think me, maybe The way I would I would put this up, is we did talk about this back. I think in in April received the hex download from you on known, Future, I think especially apar transactions. Yeah, right. is so, Is that some of the Maybe zooming zooming out when when you're thinking about your forecast, what are some of the data sources you're looking to use in a certainly cash forecast where you already have? Good day, really?

13:34 Rodrigo Cabrera: Okay. I think that one will be quite important because it's all of the open. AR and Open Safety. A

13:42 Christian Sobkowski: Yeah.

13:43 Rodrigo Cabrera: that it's created on our system but then we have a couple thinner, three or four, Google sheets that are used by different teams so the talent team they have one Google Sheet attack team. They have another to Google s. So depending on on the team, they have their own Google s that whenever we don't have invoices.

00:00 :

14:02 Christian Sobkowski: Yeah.

14:07 Rodrigo Cabrera: But there are some payments of the AP has to make. They input that in the in the Google Sheet and it's like a trucker. Okay, we have this payment. It's you by this date and of course, the AP team. It's it's a marking whenever something is completed or not and you have all of the details.

14:30 Rodrigo Cabrera: Who is it too? Amount, currency date, the description, everything that you need to know. And of course, that it will be helped to build the The forecast and of it. And I think it also help to build the logic for future scenarios. I think a lot of these payments are repetitive.

14:52 Rodrigo Cabrera: So based on the description that it's in the Google sheet, we could see. Okay, last year, this was done also every other quarter so maybe we can forecast so open on those descriptions through. A. And I think that that will be, it will be Google Sheets and and the Hex Data Tower or the the Data Lake that we have for for their pork, for the apir forecast.

15:21 Rodrigo Cabrera: Maybe there are some other things that come by email, or by slack or whatever requesting payment. Yeah, but we can also be a little bit more organized ourself and, and also track it in, in a Google should write a. So we have that organized, I think the only one that it's it's tricky.

15:43 Rodrigo Cabrera: It's the because that the talent or they we call talented human resources, related payments that are in the in the gold s***. It's It's not actually actually salaries but everything else that covers talent payment so the salaries are the tricky ones but they for salaries, there's always the same description, always the same time of the month.

16:07 Rodrigo Cabrera: So that's something that we could actually use the the machine learning, right? But for the rest I think either the data source or the or Google should be more than enough.

16:24 Christian Sobkowski: Incredible. So basically have have two two categories we have hex and we have

16:24 Rodrigo Cabrera: Yeah.

16:31 Christian Sobkowski: Google sheets. And then the Google sheets, have a bit of complexity, in some of them, where there may be some mapping required or there may be some Some data that we shouldn't live in the cash forecast in there.

16:46 Rodrigo Cabrera: yeah, and A Now, right now, the hex data it's connecting to quiliva in. So I don't know if you will pull that from Cabrera or or

16:59 Christian Sobkowski: I,

17:00 Rodrigo Cabrera: We can think how we can do it, but yeah.

17:03 Christian Sobkowski: Had a question on that so we can. And so I think so the work that I would love

17:06 Rodrigo Cabrera: There.

17:10 Christian Sobkowski: to do is you is a understand those two sources properly

17:11 Rodrigo Cabrera: That's right.

17:17 Christian Sobkowski: and then, As as soon as we have the base functionality life actually show that in the platform. Even if it's not fully customizable as yet. And but maybe if you have five minutes now we can we can talk about the hex data that we receive from you already. We ended up having a few questions on it.

00:00 :

17:43 Rodrigo Cabrera: Yeah.

17:46 Christian Sobkowski: and so, so maybe quickly the the thing we we saw on Us as we were running through. That data set, was it seemed like a very short-term forecast horizon of Less than a week for most accounts and categories. Does that match your expectation or

18:11 Rodrigo Cabrera: A.

18:13 Christian Sobkowski: Does that sound like we might have been working with the wrong file?

18:17 Rodrigo Cabrera: Yeah, maybe I think we have a couple of months, at least if I open it. Now, I think I can see, June July a pretty packed with data, maybe in all of us. It

18:29 Christian Sobkowski: Okay.

18:29 Rodrigo Cabrera: starts decreasing a little bit but still it's I don't know, 10 weeks or something like that. And what we do have it's a lot of historical balances that were never paid or they haven't been paid.

18:42 Christian Sobkowski: Yeah.

18:43 Rodrigo Cabrera: A. But but the rest I think we should have a good.

18:49 Christian Sobkowski: Good.

18:49 Rodrigo Cabrera: Couple of months, at least of data, I can share it again if you didn't.

18:53 Christian Sobkowski: And then we definitely had the wrong file. and because we were, we were seeing marks Max week ahead from the date you shared.

19:07 Rodrigo Cabrera: I hope not, we have. Half a million lines. So I don't think it's just a couple week.

19:13 Christian Sobkowski: We have half a million lines but they were the majority were in the past.

19:18 Rodrigo Cabrera: Let me check. Yeah, I'll run it again and I'll I'll check and I'll share it.

19:23 Christian Sobkowski: Okay, cool, cool. Because it's so, maybe, maybe let's start with that then. So and let's, let's take two action points away from this sink and a into company coding. I can't love a categories. b and Rigo, let's talk about the hex file and Look at what data is there.

19:46 Christian Sobkowski: I would love to. So basically, would we have created to view around? Back then actuals versus a pump forecast, versus the data you had in the ERP. and would love to refresh that view with sort of a new Part of that file, and then discuss that and Yeah. It start start doing a bit of the sinking around.

20:13 Christian Sobkowski: How do you want to combine these things? In in a platform.

20:20 Rodrigo Cabrera: Okay. Yeah, makes sense.

20:25 Christian Sobkowski: A fantastic Rodrigo. Are you if are you able to just upload a new file for us and then we'll rerun it and then we can also do that asynchron slack.

20:36 Rodrigo Cabrera: Hey, yeah. Should I upload it just in slack or anywhere in particular?

20:40 Christian Sobkowski: Google Drive is fine. I think we have a whichever you feel more comfortable with. I think Google Drive is probably the better option.

20:48 Rodrigo Cabrera: I didn't know we had a Google Drive, a, we have a Google Drive together, a

20:52 Christian Sobkowski: I'll send you like I'll send you like, I'll send you the shelter drive.

20:54 Rodrigo Cabrera: shared, right? Okay. Yeah. Okay.

21:00 Christian Sobkowski: Fantastic.

21:03 Yulia Ershova: christian, I have one one question or maybe something for for

21:05 Christian Sobkowski: Please.

21:10 Yulia Ershova: Further thinking, and what what's developing? I was thinking in regarding the

21:11 Christian Sobkowski: Yeah.

21:14 Yulia Ershova: forecast. And so far, what is also don't see in the market and I know that what it is very time-consuming for all the treasures is, whenever we have a forecast, whatever, it is 13, weeks, or further. There is no to or solution which can help identify the differences what was forecasted and what was actual I mean, I've been for instance now in Mannheim as well, I know that there are There is another famous which is also doing forecast and planning.

21:42 Yulia Ershova: I also ask them if they have anything like this and they say Okay we haven't started maybe maybe and whatever. But in reality, I know that it is very time-consuming to find out what's went wrong. This is what management is very much interested in always is asking like why we're off, right? Especially if it's okay, something small, maybe they will disregard if something bigger, Of course, they want to know what it is and how to How does Treasury amend the forecast or so.

22:12 Yulia Ershova: So in case if you are planning also you're further developments. I know it's it's first the forecast should be tackled, but then something that can be very useful.

22:25 Christian Sobkowski: Spot spot on Julia, matches the ambition, 100%. I'll briefly show you one thing and Logging into your your instance one. One second. So, what we have live right now is right? You you have this global view and we need to click in your able to to compare forecasters actuals.

22:52 Christian Sobkowski: across category, and then actually also, Dig down into. If it loads, let me quickly. I'll fix that for you. But right, so, so you're seeing this view? This is ultimately, I think it's not the the best view to answer. All the questions you are having So looking at the slack message, the point three there around better views for forecasters actuals.

23:25 Christian Sobkowski: I think that's exactly what you're referring to to understand. You know, what happened in a given week. What drove it? And then I think crucially where it needs a couple is, how does that mean I need to amend the forecast? Is that what you're after?

23:51 Yulia Ershova: Exactly, perfect.

23:52 Christian Sobkowski: Amazing. So, I'm hoping that and actually even before we ingest the The new new data like from Hex. We can already create such a view with, just a machine learning forecast for you and you can start feeding back on that on that view. And if that's what you're looking for and then Hopefully in parallel can make the forecast better for you and actually get get the right data in.

00:00 :

24:25 Yulia Ershova: That's great. Thank you.

24:27 Christian Sobkowski: Does that sound fair?

24:28 Yulia Ershova: Yeah, absolutely.

24:30 Christian Sobkowski: Amazing. Then I know where to turn to with whistles views and I know where to turn to with the data.

24:38 Yulia Ershova: Thank you very much.

24:39 Christian Sobkowski: Fantastic. Then that's all from my side. And thank you so much.

24:47 Rodrigo Cabrera: Perfect. Thank you. Christian.

24:48 Yulia Ershova: Thank you.

24:50 Christian Sobkowski: Awesome.

24:50 Rodrigo Cabrera: Will.

24:52 Yulia Ershova: See.

24:52 Rodrigo Cabrera: Yeah.

24:53 Christian Sobkowski: We'll connect Zeros. Bye.