Skip to content

Palm Internal - ON FC Validation Deck Sync - 2026-03-10

Metadata

  • Date: 2026-03-10
  • Company: Palm Internal (discussing ON)
  • Palm Participants: Emma, Giannis
  • Type: Internal Discussion
  • Domain Areas: Categorization, Pulse, Cash Forecasting, Variance Analysis
  • Attachment: ON FC Validation: Palm vs Kyriba

Summary

Context

Internal sync between Emma (PM) and Giannis (CS) to discuss ON's forecast validation deck comparing Palm and Kyriba. The deck was prepared by Giannis with ON's treasury team and covers categorization challenges, variance analysis needs, forecast accuracy observations, and ON's 2026+ ambitions.

Key Discussion Points

  • AI-powered categorization accuracy monitoring — Giannis ran CapEx and bank fee analysis with Amanda using Palm MCP, found miscategorizations, improved prompts, saw results after one month
  • Proactive insights vs reactive chat — Emma pushes for automated scheduled analysis (via Rodel's agent framework) that surfaces findings without users asking
  • Customer process documentation — Giannis creating reusable "customer process" docs in Notion Document Hub, already has 3 (bank fee analysis, account funding, CapEx analysis)
  • CS engagement model shift — Giannis moved from weekly group meetings to 1:1 sessions with individual ON team members to deeply understand their day-to-day processes
  • Variance analysis KPIs — ON wants MAPE, wMAPE, and bias tracking (over/under forecasting) beyond current variance percentage/amount
  • Forecast accuracy anomaly — one isolated case where week-5 forecast was more accurate than week-13 (latest), possibly due to category migration during Dec-Jan test period
  • ON 2026+ ambitions — cash concentration optimization, FX hedging, identifying open FX positions, longer-term forecasts
  • Three pillars for AI strategy — correctness, governance (treasury policy compliance), and hyper-personalization (customer-specific institutional knowledge)
  • Starter templates/skills — generalize customer-specific process docs into reusable templates for onboarding

Pain Points

  • Categorization is continuous — thousands of transactions, can't monitor everything at once, false positives/negatives hard to find
  • Manual categorization auditing — Amanda was manually filtering transactions to check accuracy before MCP workflow
  • Dashboard requests weren't impactful — Giannis felt weekly meetings focused on incremental dashboard changes weren't creating real value
  • Users see AI as "just another chatbot" — don't understand it can be a proactive assistant, need templates to guide them

Feature Requests & Needs

  • Automated categorization accuracy reports (scheduled agent, monthly cadence)
  • Additional variance analysis KPIs (MAPE, wMAPE, forecast bias direction)
  • Proactive AI Digest blocks showing agent analysis results
  • Starter/template skills for common treasury analyses
  • FX exposure tracking and hedging insights (2026+ ambition)

Jobs & Desired Outcomes

Job: Continuously monitor and improve categorization accuracy across thousands of transactions

Desired Outcomes: - Minimize the time required to identify miscategorized transactions (false positives/negatives) - Reduce manual effort in auditing categorization correctness across categories - Increase the speed of categorization prompt optimization feedback loops

Job: Deliver proactive treasury insights without requiring user-initiated queries

Desired Outcomes: - Minimize the delay between an actionable insight existing in data and the user becoming aware of it - Reduce dependency on users knowing the right questions to ask - Increase the frequency of valuable, unsolicited recommendations reaching treasury teams

Domain Insights

  • ON uses Gemini company-wide as their LLM — asked if Palm MCP could connect to it (Giannis clarified Pulse serves the same purpose)
  • CapEx categorization workflow: vendor list → MCP searches transaction descriptions → flags mismatches → prompt refinement → retest after 1 month
  • Bank fee analysis: harder to find miscategorizations, but MCP found Chinese-description entity errors
  • Giannis creating "customer process" docs: problem statement → current process gaps → AI improvement plan
  • CS shifting from group weekly meetings to 1:1 deep-dive sessions with individual users (Julia, Rodrigo, Amanda, Lucius)

Action Items

  • [ ] Giannis to add CapEx analysis to Document Hub customer processes
  • [ ] Explore automating categorization analysis via Rodel's scheduled agents
  • [ ] Surface agent analysis results in AI Digest blocks
  • [ ] Investigate isolated forecast accuracy anomaly if it recurs
  • [ ] Emma to share Treasury AI strategy article with Giannis

Notable Quotes

"The categorization is a continuous process. It's not just a one-off job for them. Because they have thousands of transactions, they cannot monitor everything at once." — Giannis

"We need to be proactive. We need to show the user something worth seeing. Not running it every day. Not spamming the user. Hey, we ran this. We found something. Here you go." — Emma

"I had a very transparent and honest conversation with Jen that I was feeling for the past weeks... that we were just attending to new dashboard requests... I felt for some time that it was not that helpful or impactful." — Giannis

"SaaS has never been a subject for hyper personalization or hyper customization. Now we finally live in an era where you can build hyper customized SaaS." — Emma

Full Transcript

Emma: I'm good. Sorry. Where's my camera? Turn on video. Well, maybe it's working, maybe it's not. We'll see. I don't know. It's a black screen, but I'm here.
Giannis: Yeah. I don't know what happened with my camera as well.
Emma: Oh, well, that's ignored a camera.
Giannis: Okay? What are the questions around Slide 22?
Emma: So for slide 22, I get the first one. Categorizations. I think we can probably help them easier spot potential issues or in the categorization accuracy is that. So they just have challenges. Categorization errors, plus limitations on forecasting of intercompany flows.
Giannis: Good flag. How Amanda phrase did is that the categorization is a continuous process. It's not just one of, let's say, job for them. Because they have thousands of transactions, they cannot monitor everything at once. And of course, every day, every quarter that passes, more transactions come in. So this has to be a continuous flow. I think the pain point here would be that identifying false positives or false negatives in the categorization takes quite a bit of time. And if there is a way, let's say, to simplify that process. Not simplify, but make the process more efficient. We actually did an example with Capex Categorization and also the bank fee analysis, categorization, accuracy. Together with Amanda and with the Palm cp. And it worked quite well because Amanda, for example, for Capex, sent us the list of vendors. And I feed at least an Excel basically of vendors to the Palm MCP server and it tried to search in the descriptions, those names, those vendor names, and see if they were categorized as CapEx or not as CapEx. And then we created result basically on whether the was correct. There were many issues. So in most part was correct. But there were misfires, let's say. And we worked with Amanda to improve the categorization prompt. And if you see Also the slide 28, the last slide, it worked. It worked quite amazingly as well. So that was one use case regarding the bank fee analysis. Now we follow the very similar workflow they sent me, basically. Sorry. For bank fee analysis, we couldn't do miscategorization analysis. Because it was very difficult. At some part we picked up some errors regarding a Chinese entity. The descriptions were in Chinese and where miscategorized but in most part the bank fee categorization was correct. But it helped a lot with MCP server to find these details. Otherwise Amanda was going through the transactions. And searching basically with filters. If the categorization is correct. So this is the workflow we forward.
Emma: Yeah. That makes total sense. It's kind of what I meant by do we need to find better ways to help them see?
Giannis: Yeah, absolutely.
Emma: That sounds like a great insight that we can actually use AI for this.
Giannis: Yeah.
Emma: Could maybe have.
Giannis: It's good. Quite nicely. It's not only about, let's say, finding whether or not transactions work correctly categorized. But why the AI didn't pick those categories and allocate them to the correct category and help you optimize the categorization instruction prompts. With AI, let's say it works quite nicely. So I asked. In the end, after we got the Capex categorization analysis report, I asked to optimize the categorization instructions. Whatever the cloud spit out, I put in a sense to the categorization instruction. And then one month later we tested the result. And, yeah, we saw improvements.
Emma: This is fantastic. We need to save these things, like as reusable little bits, molecules and atoms that we can reuse across our systems, I think. Does that not make sense? Like, hey, these are Genesis insights and best practices.
Giannis: Yeah, absolutely.
Emma: On recategorizing and helping users find, for example, CapEx related. Okay, do we have a list of this? Cool. Maybe we can just run that for them automatically in the background every other week, once a month. Show the results, suggest the action. Hey, refine the prompt like this. Can we automate this for people?
Giannis: Well, you know what Rodel is working on right now, right? With the agents and running on schedules. So essentially, what we need to do is create a similar scalemd or a process document.
Emma: Y.
Giannis: We feed it to that agent, it runs every. I don't think it matters. I don't think it will show a lot of potential if we run it weekly. I think that should be on a monthly basis. Maybe. I don't know. We can discuss this later.
Emma: This. My point is, can we collect these insights somewhere so we can have like a gold mine of. Right. Okay, so for Capex, for, you know, and we can have. The example is on. Okay, maybe it won't apply for all, but yeah, Capex maybe do like this. If we have these, then, you know, we can do this and you say maybe once a month is good enough. You know, like. And then for another, for intercompany, maybe we come up with something clever as well, for a certain type of intercompany flows. Or maybe we come up and I think. It's just interesting to really share these and save them somewhere, because it needs to be, I think, super explicit and super clear. Because then I think engineering can, in general, will also be able to, like, oh, wow, okay, that's actually true. But how can we capture more of these things? And that's being very. Yeah, I think this is super cool.
Giannis: Yeah, indeed. I didn't do that for the capex analysis, but for the bank fee analysis and account funding process we have in a Document hub. And under the Document hub, I created a new category that's called customer process. And there I put basically the step by step process that the customers use along with the skillmd that we created for that.
Emma: Very nice.
Giannis: I can definitely add the Capex analysis as part of this document hub as well.
Emma: Yeah.
Giannis: And then we will have. Already three customer process and we're working together and personia to create more of those.
Emma: I think. It's fantastic. One small wish from me would be to at the top of each state. Hey, what problem does this help solve? It increases accuracy and calculate like we can use it like this to increase category accuracy. Or like.
Giannis: Just open. I would say just open one, open one. See how we structured these, because I think we already grew the problem at the header. And then we say basically what the current process falls short and how we are planning to improve it with AI. These are the three main pillars.
Emma: Searching for it. Process docs.
Giannis: I sent you one link, I think.
Emma: Thank you.
Giannis: So the bank. This is. I think I need to also add another file here. But that's the main three pillars that we follow.
Emma: Problem? Yes. Okay, beautiful. And I think this is fantastic. But I also want to capture, like, what we discuss now, like, hey, maybe we can automate this part of the flow to help with the categorization accuracy in product. Maybe we like, because I think we can actually automate a lot of this running in the background. And I think that would already create immense value. And I wonder if there's a way to do that, you know, leveraging the full, like, entrophic product. That we don't really surface to use as yet, but we can still deliver the value. And the value being the insights around what category? Like transactions that you might want to recap. We can still deliver that insight, you know what I mean? We can still deliver the output without having the full cloth entrophic experience. In our app. So I think that would be immensely valuable to already start finding, hey, what value can we like just deliver automatically to them? And I think for I might be wrong and I think they definitely going to want to do analytics themselves and all of that, but I think it might be helpful to have a lot of this automated, so we're proactive about notifying them. Hey, we run this analysis last night. These are the findings. This is our recommendation. You know what I mean? And then they can just like, oh, yeah. Let's refine the Opix a little bit.
Giannis: Yeah, I think that should work. However, one question on the part that you mentioned, what we're trying to solve with AI. Who's the target audience? Is it the end users or the engineering team?
Emma: You mean for this docs? Sorry.
Giannis: You mentioned that. You mentioned something that this document lacks and. That's a section that we say what could be useful for AI to leverage. So what we could leverage with AI. To deliver a solution for this problem.
Emma: Yeah.
Giannis: What would the target audience for that section be? The engineering team or the end customer? End user.
Emma: I think it's both. Right. So my example is solving a problem for the customer. Right. But it's also making it really clear for internal team. What are some options? So this is not just about. Oh, if the user uses Palm Chat, they should be able to conduct this full analysis. Like, that's one feature that would be amazing. We want to support that, of course, but you've extracted true value here. You've helped them arrive much faster in an insight. Which. Is these transactions are probably not correctly categorized. This is probably why, and here's our recommendation. And if we can just deliver that without them having to do anything, That's also incredibly powerful and in theory wouldn't require us to first build the whole Palm app chat perfectly. As you know, the Entropic product is working if it's possible to run something just on our server using a full Cloth code instance, for example.
Giannis: Yeah.
Emma: It was just.
Giannis: Did you see? What pairing with the agents.
Emma: Sorry.
Giannis: Did you check what Rodel is preparing with the agents part?
Emma: I only see what he demoed yesterday.
Giannis: Yeah. So basically, we will be able to create those agents to run on a specific cadence, so daily, weekly or whatever cadence. And they can send this prompt, let's say, for example, their bank free analysis or the categorization analysis. They can trigger the agent to run on the background and get results. I don't know every day. And I think, what would. Be very important is to create a new blog at the AI Digest that also gets this information from the results of those triggers, whenever those are available as well.
Emma: 100%. And that's what I mean. We need to be proactive. So we need to show the user we need something worth seeing. Not running it every day. Not like spamming the user. Hey, we run this. We found something. Here you go. And I think exactly that if it's agents, if it's what Rodell is building. I honestly don't care, but if that's going to support it, that's fantastic.
Giannis: Yeah.
Emma: But I think ideally I just want us to try and find where we can deliver proactive insights instead of having the users having to pull them out of us. However, I love it that we need to do this as well to learn, but I just feel like there are these little insights like you just told me now, okay categorization. It's an ongoing process. We already know that. We know that it's something we need to get better at. We need to know, move beyond the dashes that Jan has built to help them spot deviations. How do we move beyond the dashes? This is a fantastic example of that. Right. And how can we, in a simple way, deliver that value? And if it's what Rodell is building, if that's going to manage to deliver it, that's fantastic. I just don't want it to get lost because I think we can build a little then. Okay, so this is looking at capex. What? Would something. Looking at bank fees, what would that look like?
Giannis: Yeah.
Emma: Or looking at these categories, what would that look like? And so on and so forth. So I think that would be really, really powerful.
Giannis: Y. Eah. So maybe let me update you on a conversation that I had recently with Jen, and then I have to jump.
Emma: Yeah, of course.
Giannis: I had a very transparent and honest conversation with Jen that I was feeling for the past weeks. Before the workshop we found that we were just attending, let's say to new Dasport request. Let's build this dashboard. And that last part and this change here, this chains there. That I felt for some time that it was not that helpful in the end or impactful. Maybe it was helping them in a way. But not creating a big of impact. And that's why we decided to change completely the structure of the weekly meetings, at least with on and personio, and instead of focusing on weekly meetings, to have one to one sessions with each of them. Where we go through, let's say, Julia, Rodrigo, Samandas, Lucius. Day to day. And let's say, consume as much process and create those process documents as much as possible. And then convert into the customer process docs that we have in notion right now.
Emma: Very nice.
Giannis: And same for personio. Hopefully same for prospects and incoming customers as well.
Emma: That's fantastic.
Giannis: I think that's going to change completely the way that we interact with customers. And how we also deliver value to them. Because Saint is completely. It's not just okay, we make an incremental change. We fundamentally change the way that they're going to work, their way of working is going to change massively. So this is what.
Emma: You're preaching to the choir. It's in my doc. It's in the blog post. I 100% agree with this strategy. The first thought that popped in my head the second you shared that first skill was exactly that. This is what we need to go all in. And this is, in my opinion, the main thing we should focus on and just try to build for closing the gaps there in terms of correctness. Governance and customer like customization, right? So if you look at the doc, I've set up three pillars that we need to really, really deliver on, and in its correctness, And this is something I'm speaking to engineering about a lot, given his data. If the answer is allowed. Right. So treasury is a complex domain. They have policies, they have restricted ways in which funds can move between entities. There are a lot of boundaries that needs to be modeled out throughout our database, throughout our systems, because we can come up with a really great recommendation. But it's violating policy or it's not physically possible, or it's like, you know, not, you know, legally possible. So we need to make sure that the governance layer is there as well. And then thirdly. The most ironic part of it all. SARS has never been a subject for hyper personalization or hyper customization. Now we finally live in an era where you can build hyper customized SaaS. And I think we need to figure out how that works. And that part also requires exactly what Lucia sent in her latest long post that I shared. She said exactly that. Right? She wants to feed a lot of these nuances, a lot of information. She wants to feed into pom, and I think she expects us to remember that. You know what I mean? So that's the company specific institutional knowledge that just needs to be in there to make it a lot more custom to the context that the business is operating in. And that's and I think that's going to be key for us to figure out how we kind of.
Giannis: Yeah, I think most sorry.
Emma: Get all of those three pillars and the fourth one that I think is the biggest one is actually evals and learnings. And how do we make sure that the processes that you're getting from the customers now, sure, you build the reports.
Giannis: Yeah.
Emma: When do we get. Like, when do they tell us whether or not it looks correct or wrong? You know, all of these things.
Giannis: I think in two ways. The first way is just plain and simple. Ask them and go into detail into that. But then we have implemented, let's say, products analytics, like with posthog. We can see how often they interact with it, how often they use it. And, yeah, basically, we have many ways to track that. In the end, I think what we should do is create templates that they can run, because I think these people are not very educated on what AI could do or the capabilities. They always consider this as another AI chatbot that I'm asking questions, but not as, let's say, assistant in a way. So I think this would be important for you as well.
Emma: I agree.
Giannis: As soon as we create those individual customer processes, we could generalize them in a bit and create some example.
Emma: Hundred. Percent.
Giannis: That we push.
Emma: 100% starter skills, starter, template, starter, whatever we call them, like, you know, run it, make it yours, make it. But this is when we hopefully get a new designer on board. This is why we need a really strong point of view on the sign, someone who owns that user experience. So that's going to be a big part of whoever joins us after Tomoya.
Giannis: It.
Emma: But 100%. 100%. I agree. Couldn't agree more. But. Okay, so that's that. But otherwise, in the company I know we have to improve. Me and Rodell are continuously working on ad hoc reporting. Yeah, that feels pretty much like. We'll cover. But the data storage for the variance analysis, I didn't quite understand. This day. Manual variant.
Giannis: Sorry.
Emma: I'm on slide 22 for manual variance analysis. Data storage they have. And then they say palm in house reporting for variance analysis. Snapshots with visuals and KPIs. Are they talking about comparing forecasts to each other or.
Giannis: Oh, yeah. So that one is basically to find a way to delineate on the variance analysis because now we have a global view. Plus KPIs. KPIs mean that now we have only variance, percentage and various amount. But essentially what they want to track is additional KPIs like map or weighted map, etc. So if we could surface these metrics, let's say that would be very useful for them. And the last part, that Simon is already informed. And again, the forecasting team is already aware about this issue. Is the bias? Are we over forecast? Are we constantly over forecasting? For example revenue or payables? Or under forecasting. Are we optimistic? Pessimistic, let's say. Because it's good to be a bit pessimistic on revenue. I think it's better to be pessimistic on revenue and not over optimistic and the same and vice versa for expenses.
Emma: And I think we'll see.
Giannis: Something like that.
Emma: Yeah, yeah, yeah. That's the class. That's the classic posterior for. For Treasury. But I think Lucia expressed many times that. And I see it in scenario planning to be enabled for the growth assumptions. And I see probably since they're growing so fast and they're growing like, not always we can just. We can't just rely on the historicals for on s growth journey. So I think they're keen to have more influence and more say in growth projections in general. But I think that makes total sense. And then I feel like. What is the decrease on forecast accuracy from week five to 13? Oh, it's a challenge. Yeah. We're more accurate in the short term. Okay. Maybe.
Giannis: No, we are less accurate in the short term.
Emma: But it says decrease.
Giannis: There was an example.
Emma: But it Sundays from week five to 13, we have decreased. That's what it says on the slide. But happy to hear. I'm just confused.
Giannis: Yeah, okay. So the specific example we brought is that there was one occasion. Maybe it was one occasion, but good to investigate. Is that? If week 13. Let's say 10th of March. Sorry, if week 13 is one week prior to the forecast date. So the latest forecast. Week 13 is the latest forecast. In week five. Which was a much earlier, let's say, forecast for that particular date. The accuracy was much better than the latest forecast, which was week minus one.
Emma: Aha. Okay. Interesting.
Giannis: And we need to find out why that was. To be honest. I tried to check with Simon and Art, but we couldn't find a specific reason for that. Maybe it was an isolated scenario, et cetera. Maybe we also did. So the data set that they picked was between December and January. So the seven weeks between December and January, and that was exactly when we also did the migration. On the categories.
Emma: Okay?
Giannis: So few things could be off as well. So I wouldn't call this a very representative data set, if you may. So.
Emma: And maybe. Yeah. Maybe not urgent, at least for now either. In terms of maybe so wait and see if it beating, but yeah, fair enough.
Giannis: Absolutely. I think that should be the way to go. We didn't observe it in any other, let's say, occasion or scenario. If this surface is again, then we should take a deeper look, but in my opinion, We should leave it in the back burner. If it pops up, let's pick it up again.
Emma: Super nice. And I love the Ambitions 2026 Plus. It's really clear they want insights on cash concentration, optimization. They want FX hedging. And identifying openness.
Giannis: That was my part. So these were basically. Things that we discussed in the past. What are we releasing? Usepalm They lacked it and they wanted to put it as ambitions.
Emma: Cool. The ethics hedge stuff. Would love to learn more about what you've discussed with on because it's currently on the potential candidates for coming up next on the Roadmap. But. We first need to have a better way of kind of just tracking ethics in general. That's fine. And then longer term forecasts, blah, blah, blah. Yeah. And they want to use it in Gemini. I see. It's not like.
Giannis: I think that was a bit of misinterpretation. They use company wide Gemini as an LLM. So I mentioned basically that this is an MCP server that I connected to Claude, and they asked me, okay, could we use it on Gemini? I said potentially, yes, because at that point we didn't have the pulse ready. So I would assume that eventually we would release it too, as an MCP server. But in any case, I think if we can transfer these capability in Pulse inside the app, it serves the same purpose, if not better.
Emma: Yeah, I think one of the best is that if we keep it internal as well, we, we have a chance of building the harnessing and context optimization, all the tooling more around the treasury domain. So for sure.
Giannis: It doesn't matter. I don't think it matters for them whether they use it on Gemini or in Pulse. As long as they can. Better results.
Emma: Of course. All right. Thank you so much. This was super, super useful. I hope I didn't have the opportunity up for too long. I think this was way more than 15 minutes.
Giannis: I know, but it was important. But I really could have done now. So really happy to help. If there's anything, let's pick it up tomorrow as well. And you probably understand already that I'm really a believer of this AI capability on how customers can use it. But let's see how to push this forward for more customers as well.
Emma: Yeah, yeah, please read the article. It's our internal strategic direction. So I think it's good for you to. To be.
Giannis: Yeah, absolutely.
Emma: Happy for feedback, but this is super useful and I think, yeah. I'll add it to the product docs and we'll take it from there. Thank you so much, Giannis.
Giannis: Perfect.
Emma: Have a good day.
Giannis: No worries. Nice talking to you.
Emma: Likewise.
Giannis: You too. Cheers. Bye. Bye.