Build or Buy: Should You Create Your Own AI Headcount Tool?
Podcast Overview
-
Eric Guidice Headcount experts, emergency episode. Episode 23. We have to figure out what's going on with Claude, headcount, AI headcount. I'm a provider of headcount software, right? I my my pitch is that the amount of work it takes to maintain the data set should be done by a software, which is a collection of microservices so that your analysis can be correct. But I realize that with AI out there, there's different approaches to solve the same problem. And Chris, you focused on building AI-based solutions for different segments of the market. So, kind of give an overview. I think you know, we don't we don't touch too much on what you and I do day to day. but Sans guest, let's talk about it. What is Meander HQ? What are you working towards as it relates to the customers that you're trying to serve? What do you teach? and and kind of lead us into the AI and head count conversation.
Chris Mannion Yeah, I think the main thing was I found I was doing consulting work and going in to work with new HR leaders generally to help them get on top of their data. And so we do a of work, a of manual analysis and actually get all the data in a clean place and all the spreadsheets up and running connected to finance and connected to IT. But then when it came time to actually hand over that work and move on somewhere else, they either had to hire an FTE to come in and continue the work or the work just stopped. It stopped progressing. And so there was this challenge there where I felt like there was a knowledge gap between actually understanding how to use the data and actually kind of applying that to the overall strategy for the organization. And so what I'm trying to do now actually with the Talent Academy all came from this idea of what happens after I leave as a consultant? Can I actually up-skill and train other people to do the stuff that I would normally do? And we've started to go really deep on that. And so a lot of the newer courses are really trying to fill that gap. And actually, instead of hiring me as a consultant, spending a lot of money, you can actually learn to be your own analyst. And with AI as an assistant, not necessarily as your analyst, but you as an analyst with AI supporting, I think is the way that I would recommend moving forward and what I'm seeing work best in the market right now.
Eric Guidice You know, for as different as a path as we took to get to where we are, and you know, then how we diverged after this consulting period to do our separate kind of solutions, that consulting experience and that problem is something that we shared. Before I made Headcount 365, Unicorn Talent, which I now give away a lot of the stuff which may or may not be outdated or needs an AI refresh. We'll talk about that at a different time. But I would go into companies on behalf of like big investment banks when they were first launching, set up recruiting and try to leave it behind for somebody. And some of the more complex processes, namely headcount, had a difficult time continuing if that person didn't one, get trained on what was going on, or two agree with how some of these things or the approach of some of these things could happen. and so it it was a problem. I think there is a, it's not necessarily a I think most people have an idea of what's going on and and what they want to get out of it, whether it's driven by the company wanting to get something out of them or their own approach. it's just a matter of how do you execute it and how what's the best way to consider consolidate this cross-functional multi-system process into something that everyone can use for their own purposes without destroying the internal data or the the underlying data that is the foundation. And I think that's where you and I are taking different approaches to the same problem. Yeah, we'll talk about what microservices or sub-softwares are, right? But whether you're building it in AI and you have a process that you would want to automate an AI, or you let a subservice of headcount 365 do this work, there's still these microtransactions that are happening that all need to happen correctly so that that underlying foundation is correct. So I think that's you know generally the the focus today. I guess give me an idea of you know who your target is, who you're serving, what what are the folks who would be interested in taking a talent academy course, and for those of you who don't know and want to preview Chris has an amazing YouTube channel called Recruiting Analytics where I've learned a bunch of things about not only headcount but some of the approaches he's taken in his career and it's just been a great way for us to bond over some of the the recruiting knowledge but who's your target for the town academy and what who's taking your courses and what are they getting out it?
Chris Mannion Yeah, so generally we find that new in-seat HR leaders are taking the course and there's normally a reason for that is that they have been hired to replace someone who has struggled to use data in their role. And so now they have 90 days to ramp up. And so not only are they having to demonstrate their ability to do the job well, they're also having to absorb this large amount of information and use it effectively. Generally in 2026 when we're recording this, it's also with fewer overall team size as well as a result of the rifts over the last five years. So you kind of have all these different headwinds facing these individuals and they're just struggling to kind of keep up and prove their worth. And generally that first 90 days is so important. So what I've tried to do is actually create content and a roadmap so that you can actually get on board and get up skilled as quickly as you can and then be able to get the skills that you need in order to be successful with those next two years and be that kind of person that's first in line for a promotion when you get to the kind of like next stage in your career. So that's who I'm seeing. One of the big decisions that I made as we moved into building out the Talent Academy was it's very much an opt-in process. I'm not actually trying to sell this to L &D departments or to CHROs to train a whole team. What I found is that if someone is motivated to actually seek out the course and take it themselves, the much more likely to actually finish the course and perform well. And then what I'm seeing is those people are now getting recognized by C-level in their organization and actually really kind of outperforming their peers, which is kind of the people that I think this kind of approach appeals to. And we actually haven't talked about this, but there's a kind of a preview that I just launched on the website yesterday. We're actually taking like, we're going even further on the learning where we're actually doing case-based learning now. So it's a simulated approach. The first one is actually relevant to what we're doing today, which is how to find bias in an AI recruiting tool. you if you've tracked the news, if you're seeing all of the legal challenges, if you're using AI to shortlist candidates in a recruiting tool, what is that potentially opening you up to? And so we actually created a case now so you can actually walk through and actually really understand how to do this work yourself and actually grades you in real time on how well you're performing, gives you feedback and actually links you to free videos, to courses and resources and everything you need to upscale. So we're going even deeper on this. There was a surprise that just occurred to me I should probably mention given it's so new.
Eric Guidice Yeah, and I will, if I didn't already do it, flash it on the sc while you're talking, it'll be an amazing inlay to the to the podcast there. But what I like about your courses is I think it's based on your background really, is like they appeal both to the HR person who is come up through the HR ecosphere, as well as someone coming in from the outside. So you I know did supply chain before popping into the HR talent space, but there are folks coming from a chief of staff or an operational role that, you know, sometimes in small company HR, you're getting folks who are I'll take this from Jason Zoltax Tofu, right? The GPSP, general purpose smart person who's gonna come in and tackle these things. And so while they may be great at doing some of the work that we'll talk about later in the episode, they need the context of what's happened in talent or HR in in in the past so that they can have the right framing for how they present this data, whether it's outward, upward, or to the team. So it's just a good perspective for those of you as haven't checked out Chris's learning, and you get Sherm credit, right? Am I correct in that?
Chris Mannion Yeah, so we have two courses now that are SHRM certified, so you'll get PDCs. So if you need to maintain that SHRM credential, can, assuming you've finished the course and it's not a checkbox exercise, we have some hard questions in there. But if you can complete the course, you'll get the SHRM credits. And as we're adding more and more content, the idea is that a subscription to the Talent Academy will keep you fully up to date with your SHRM credits while also staying current with all the changes and all the capabilities that are required for a modern HR leader.
Eric Guidice Amazing. Okay, well, let's get into the meat and potatoes of today's episode, which is essentially you know using Claude or your AI model of choice to build out a headcount process within your business. So I know this is something not only that you teach, but that you've done for a number of your customers. and there's a couple of kind of key things that we want to hit on in this episode. So I guess give me an overview of the builds that you've done and the courses that you're teaching, and how do you want to kind of talk about using AI to build a headcount process from scratch. Let's let's let's talk about it.
Chris Mannion Yeah, so the golden solution and actually what we teach in the courses is to not start with AI, to actually start with understanding the kind of end result that you're trying to create and have this concept that like to call the golden roster. Actually, the free webinar that's on the website just teaches you how to build a golden roster without AI. that's, think, a key requirement. And once you have that golden roster, you can actually start to do a lot of analysis over the top of it. And so what we try and do is have you start in the spreadsheet, understand what's going on, understand the data that you're using and be able to do very basic, know, build a pivot table, build a table that's going to give you the metrics that you need, be able to calculate some of the basic metrics. And so that's, think, foundational skills, because once you've got that, you can actually accelerate a lot of that work with AI. But if you have AI do the work for you and you go into a meeting with a CFO and they ask you why your attrition rate is this compared to last year and you can't answer that question, you're instantly going to lose all your credibility. So that's one of the things that we really, really kind of kick off with. And the principle really is AI should be checking your work, not doing your work for you. It's kind of a helpful intern, not a senior analyst. And so it's kind of like a framing that we take. One of the things I want to walk through today is actually a case where I've used Claude for Excel to basically build the whole thing out and it's going to be full of errors, but the errors won't be obvious. But what I wanted to do is at least kind of show how that can be done and maybe hopefully argue more to the point of actually starting with the understanding how to build the work yourself and then realizing there's a trade-off at some point, the data cleaning and data aggregation gets so complex, that's going to take the majority of your time. you either need to hire an analyst or get specialist software in order to do that for you. And so I think there's this spectrum of capability requirement and up to a point, you're going to be able to do a lot of the work yourself. But as soon as the organization gets big and complicated enough, I think that's when you're going to need to kind of look at other resources. And this is really kind of replacing the need to hire a consultant. I've done some smaller scale consultancy work where we're doing things like looking at a revenue plan and trying to figure out how big does the sales team need to be? And then how do you think about going all the way from, you know, very top of funnel marketing impressions to MQLs to SQLs to sales, to post sales support, and how do you need to build the team and how do you build the capacity model to do that? It's very complicated, even at a small scale. And so that's where actually being able to do this yourself can be really beneficial and not having to spend a lot of money for a consultant like me to come in and actually do this for you. Also, because I've done it so many times, I think it is easy to kind of teach it now.
Eric Guidice Yeah, I think what what's you know, thematically there's a lot of similarities between the foundational principles of using AI to build headcount or buying a software to build headcount, right? And I think one of the main similarities is the data has to be correct. And you should be pulling from a unified data set. And the way that the data changes is important to the things that's built on top of that data so that not only can you show the changes, you can explain the changes, right? I don't think these things are different. You know, if if I'm in a sales process, and people are comparing Headcount 365 versus a cloud-based model, they are comparing features. But the real value is not necessarily that something can be built on top of the data, but that the data itself is correct. And I think as we get into the differences between software and solutions, it's about completing the individual transactions that make that data correct, right? So from the data sources, who changes the data, and then kind of that purgatory of like pending changes or scenario plans, all of those things that are influencing your data set are needing to be managed. And so, in the best possible process, you might create multiple linked automations to then create a data set that then creates multiple linked reports or outputs. But as you scale and as you get complexity within your headcount process, this is how gonna make the determination of is it more economical or accurate or whatever to have a person manage these transactions in a process with a tool or to buy a software from the outside. And so I think what is super interesting about your process is that you and I share the ex I wouldn't say I might even say exact. Do I say exact? We share the exact philosophically, I think we're pretty aligned in terms of, you know, how we operate with with the data. And so seeing the course and seeing how you've kind of brought some of these things to life using the tool, it's been it's been extremely interesting. so I'm curious to know kind of, you know, how do you go about getting something set up for a company? Like what are the you know, foundational rules, the inputs, the things that are like the non negotiables when someone's going to set up a, you know, AI based headcount process within the within their business.
Chris Mannion Yeah, I think there are several things that I think you want to be sure that you're hitting. So one is the first thing you should do is set up an assumptions log. Because I think anything you do moving forward has to be tracked and you have to have that audit trail of, okay, we did this and this is why we did it and here's the reference to it. And that way when you get questioned on it, you have this defensibility. And AI is not very good at doing this because it will just do the work for you. So setting that up first and you can set it all up in the spreadsheet, think is the right way to start. From there, you can actually then...
Eric Guidice wait before you move on before we move on, what are give me some examples of assumptions? What what what's a good like what what are some things that you see in common between either customers or cust or clients?
Chris Mannion Yeah, so I think things like whether attrition is going to be included or not in the headcount plan. I've heard CHROs not want to include attrition assumptions when planning headcount, which I have strong feelings about. Whether you are kind of modeling different growth rates, how you're doing currency conversions is a big one. And what you consider an FTE is another big one and is FTE headcount and how you tracking that. So do you want to track number of heads? Do you want to track FTE? Do you include contractors in either of those? How are you thinking about variable versus fixed headcount? So there's kind of all these baseline assumptions, which should be corporate knowledge, but are probably not really tracked anywhere. And then even things like, what are the things that we're tracking for to measure performance and success, like time to hire, and then how do we define time to hire specifically in this organization? And so anytime that you're defining something that is gonna be used more than once, it needs to go in that assumptions log so that you can make sure that you're using it consistently. Otherwise, you then have kind of a quarter over quarter comparison when you get six months down the road and you're looking at kind of two metrics, but you've measured those metrics differently, it's actually not gonna help you tell a story that's going to be grounded in data. So I think that those are some of the kind of key things. But every time you go through the process, know, if you're, for instance, one of the next things I was going to come to is cleaning the data. And if you're a multinational organization, you probably have multiple HRASs. You probably have some headcount that's being tracked on an Excel sheet somewhere. And when you think about how do you actually blend those data sets together to create one master headcount roster. If you have duplicate names, duplicate employee IDs, which ones are you going to select? And you need to know what is the rule for deduplicating because if you do it differently each month, you're actually going to get different results because the start date may be different or the data may be incomplete or one data point may be more outdated. And so it's kind of like all these baseline assumptions, which and not really fun to work on, but the good news is if you actually track them in an assumption, like you only need to do this once and then it's kind of done and you only need to update it when something changes moving forward.
Eric Guidice Yeah, with the with the assumption log, I I went, you know, for our fellow prompters out there, it's equally as important to say what to do as it is what not to do. And so with headcount, having the assumption, so having a currency exchange rate, but telling it to do it once a month, once a year, you know, finance teams have different times that they reconcile the delta between currencies, and that is an accepted variance within you know their models. And so having to communicate that to a hiring manager who's trying to do a trade-off between a role in one place or another, that's kind of one of those things that, again, that is a microservice in HeadCount 365. However, if you're gonna build that into an assumption login Claude, you have to get all those teams aligned. And then the next thing that you brought up with duplications, this is something that's been tremendously interesting for us to kind of work with our customers on, is like from a from a de-duplication standpoint. A backfill will exist at the a ID will exist in your HR. And then in your ATS and then your HRIS again in a given year. And it will also correlate to a finance line item once and maybe twice, depending on how you do your your finance IDs. And so deduplication is, I would say, one of the you know most difficult microservices to master where the double counting can be, you know, the double counting can be extremely difficult. I'll give you another example, right? We're doing we're doing a lot of work with attrition and backfills and vacant seats. And so a lot of our customers want to know how many, how many backfills are left to be filled, right? At the same time, they sometimes put in placeholder backfills or confidential backfills, and then that backfill that was a placeholder when someone actually leaves has to be linked. So when your assumptions log, you have an idea of like what is a seat, what is a backfill, and then how do you associate, how do you give the business the autonomy to do the things they need to do to get the business outcome while still maintaining a clean data set, which I guess is the the proper lead into your cleaning the data second point.
Chris Mannion Yeah, exactly. And it doesn't just go with HRAS to your point on hiring backfills. You may have five different openings for one position because you don't know if it's going to be, I've seen this before. Are we going to put this position in Canada, in the U.S., in Mexico, in Europe, in APAC? And it really depends on where we can get the best quality candidate for the role because it really is location agnostic. But you can't open one rec and market to all of those different areas. And so how are you actually deduplicating that? Because you're otherwise you're assuming you're going to hire five times as many people as you actually planning to. So all these assumptions need to be baked in such that if I was coming in as a new HR leader, I could actually look through the assumptions list and instantly understand what's going on and not have to go around to all the different stakeholders and try and figure out how they're actually managing their planning process. So I think it's a really important part to get right first before you start doing any of that. the data cleaning and data managing.
Eric Guidice So who who owns the assumptions log and who has access to change it? Well give me an idea of like how do you manage that? Well if you're gonna leave that behind for somebody, how do how do you manage that? How do you manage the definitions? walk me through that that part.
Chris Mannion Yeah, I think it depends on the organization. So I think if you're planning for a business headcount plan, so it may be the CRO should own the plan and it's probably their HR VP that's going to help to kind of manage that. I think if you're looking at a whole organization, it almost certainly has to be someone in HR. It was generally me, but that was only on a temporary basis. So this is a HR ops team. They're probably best suited and that they're actually many of the folks that we see go through our courses because generally they'll come up through HR operations without necessarily going deep on the analytics side of it. And all of a sudden they're being expected to do the analysis and the head camp planning as well. So we help them to get up skilled quickly. But I think that the point is it should be one clear owner who's going to be updating that. And so all the changes must come through that one person. And that's how you kind of maintain it. And then if they leave the organization, at least the there is that source of truth that can then be handed over to somebody else. But I think from my experience, the CHRO has always wanted to be able to go in and actually look through all the assumptions that are being made to understand if they're reporting out on a metric, does that metric actually align with the way that they believe that the metric should be calculated? And that's an interesting point as well. You could have a new CHRO join the organization and all of sudden they have a different belief on what attrition should look like. how they should be marking FTE versus non-FTE or the market could change completely. There's kind of like a lot of nuance to it, which is why I haven't tried to give a general rule. But I think if you have the log, then you can kind of adapt to any situation moving forward.
Eric Guidice I got it. So so I mean, look, I creating a software that can accommodate all of the different assumptions has been it's still it's an ongoing project. Yeah, I don't think we'll ever be done. but I've seen everything, right? Some companies wanna have comp include bonus, some don't. Some wanna have the budget include merit cycles, some don't. Some consider interns FTEs, some don't. There's just a lot of different assumptions. So I understand conceptually where that goes, and having a reference document is kind of a help center. Source of truth for how does this model work is, I would say, kind of a good step one for building a cloud model. And I think before you get in there and start building whatever it is that you're going to build, whether it's a microservice, the database itself, or the entire platform, having a unified, you know, unified source of truth is more than just a budget ID or a rec ID, right? It's a we all process this information in the same way. And so how does I we could probably go for a long time. I don't know how much time we you have today for this podcast, but we're 30 minutes in, we're on step one. What's step two?
Chris Mannion Step two is actually starting in the spreadsheet and building the initial kind of cleaning and the initial analysis yourself, which can be quite daunting if you've never done it before. We actually cover that a little bit in the kind of courses and actually some of the free content should get people up to speed with that. But what that allows you to then do is have AI check your working and show you where you may have gone wrong. I think the moment you say, and I'm going to give an example of a spreadsheet that I've that built with AI, at the moment you say, build me a headcount forecast, it's gonna go off and do it for you. But we go back to that assumption log, it's gonna create its own assumptions if you don't already have them in there, and it's gonna build its own processes. And if you then have to explain it or change it, you either have to go back to AI and have it change it, which could hallucinate, could lose track because it's out of context and doesn't know what it did last time, I think it's important for you to be able to understand what's going on there and kind of build it yourself. that's, think, where I would go next is just making sure that you can actually build everything yourself and all you're doing is using AI to automate and speed up the process a little bit and check your work in. So kind of a big part of the kind of like base assumption and the base hypothesis for the courses.
Eric Guidice Yeah, I think with with all the different availabilities of connectors, right? I think in the last 30 days, Workday, Ashby, Greenhouse all released some form of connector, right? So once you have your assumptions and you have your base headcount data, you can now start to interact with the base data. And so there's a couple of ways the data will change. For simplicity's sake, let's call it a finance-led change, a business-led change, or a talent-led change. Right. So a finance led change is we've changed a forecast, we're adding headcount, we're reducing headcount, and that could come from business performance, it could come from fundraising, it could come from a change of strategy financially, and you're gonna have a change in the headcount plan. And in the world of target to goal, this is a change of a goal at the planning level. So that's a finance-led change. a business led change is a hiring manager or budget owner led scenario change where hey, I would like to add a net new, I would like to backfill an existing, or I would like to take one and split it into two or completely reorganize my locations, etc. etc. And so this typically requires some form of approval or internal dialogue before it gets accepted into the plan, but you will go from one plan to the next. and in our previous episode that you and I did, we talked about H2 reforecasting and the steps to kind of manage changes like this, whether you're doing it in a bulk half one, half two, or you do it individual to one change to a hiring manager, the principles of associating the change with a history, tracking that change from where it was on the plan to where it is today and what the impact is, those will all permeate through. The third change is an HR or talent-led change. This could just be I've hired somebody, I've hired somebody ahead of schedule, above salary or behind schedule and below salary. These change your financials. there could be a termination or promotion or internal HR change that will impact your headcount plan. So those inputs are now impacting this base set of data. So with these new MCPs that are launching a service that needs to apply to your business is how does headcount change and how does that change get approved and tracked, and how do we manage the delta in what was to what will be? And then is that the new then baseline for our reporting? So that's kind of a micro service that we've done very well, I believe, in headcount, but can be done in Claude or with some form of AI plus tools, Zapier, N8N, whatever. but how do you how do you, as you're setting up your AI-based headcount plan, manage the changes or the process for change to your data set?
Chris Mannion Yeah, that's where it gets quite tricky actually, because at that point you're not necessarily managing prompts, you're actually managing the source data and you have to update the source data. And this is where if you've used AI to generate a of the kind of a table on top of the data and you don't know what it's measuring, then as soon as you start making changes, you could actually destroy all of the models that have already been built out. Because generally and kind of broad-brush speaking, if you're using AI to run analysis, it's running that one time. It's not designed to be a repeatable analysis. And so actually setting something up so that you refresh a pivot table and it refreshes your metrics table is probably not going to be something that that code is going to do out of the box. You're going to have to really kind of tell it to do that. And so I think actually tracking those changes requires you to actually have a golden roster going back to that kind of like master headcount, like source of truth, where you're actually adding these changes and adding a note to show that this change happened and this is why. And what you then do is you have these kind of like data sets. And one thing that AI is good at is comparing two data sets and telling you what happened between the two. So that's where I would think about AI there. So you kind of have your baseline model, you have your metrics and you maybe look at, we're looking at July versus August. And then you can then say, okay, well, we saw that the kind of the headcam cost is above forecast for August. We're trying to understand why that is. Look at the two data sources and tell me what's happening there. And assuming that you've added the correct annotation to your raw data, you've gone into the spreadsheet and actually updated it correctly, it's going to be able to pull that and give you that insight without you having to sort through and find it yourself. And so that's where we talk about AI accelerating your work and not having to do that manually. But if you've not gone through the effort of actually updating the source data itself, the AI is only as good as the data. And so without clean data that's constantly updated, you're just not going to get the result that you, I think, want out of that.
Eric Guidice Yeah, I was jamming with VietWin, formerly Vercel, but now is on his own, kind of doing a similar consulting that you do for companies trying to apply AI to a process. He showed me an awesome tool about like hiring manager visual scenario planning, which allowed managers to kind of shift things around and make calls, and he built it all in AI. It's just then how does that get into the data set? And how do you take something that is a massive org change and communicate it to all the individual line items if you are eliminating, replacing, and then trying to stay budget neutral. So there's all these problems to solve. I think AI can either streamline a visual for how you want to make changes. I think it does a really good job if you're doing like a greenhouse MCP of maybe you want to update the status or something based on an action within greenhouse. There are different ways AI can influence the data set. I think one of the main problems that I had to solve. When I was an internal TA leader before AI, was the version control. So, you know, if we had, it's a very common request for our customers to say, what happened week over week, right? What happened since last week? How many terms? How many changes? What is the current status this week? And what was it last week? And why do I care about this change? Right? I think the there is a depending on what your needs are as a company, right? Not to make this a headcount 365 commercial, but if if you need something more real-time, a software completing these microtransactions in a real-time environment allows you to have a comparison against two dates that you pick. When I was pre-AI, we were we were amazing, right? We were incredible because we did it once a month. And so AI believe it can probably bridge the gap between every second or minute or whatever and once a month. But I guess you know, one of the things that you have to do when you're considering an AI build is like, how often do and do I need to be responsible for pro providing clean data for someone to make a business decision? And if you know it not only not only do you know that information, I think different companies will assume that they need it until that one day they need it today. And I know it's supposed to be by Friday. But anyway, the point is if everyone agrees that you have a a date that everything will be correct, then you can then build your models and your and your and your change data and your source data to reflect these different milestones. And I think that's an important part of a build when you're considering an AI build versus software build or just you know status quell. so we covered assumptions. We c go ahead, yeah. great.
Chris Mannion And I think that touches on the third point we're going to talk about as well, the kind of building on the assumptions, but also the kind of limitations in terms of the ability to actually use the data that you have in the source data. And so you gave a good example there of a hiring manager who's maybe thinking through different assumptions, different scenarios. Well, in order for that to be successful, they have to have access to all of the data, like all of their team data, location, start date, end date, any terminations, compensation data. And at some point, maybe you do give higher managers access, but it gets very complicated when you're looking at kind of global organizations and different regulations by country. And so where AI really thrives is in sales-related analysis where there's actually less of a kind PII concern of actually looking at sales data. As soon as you get into headcount and compensation data, you could be really careful where that goes. So, if you're not going to send something by email, you probably shouldn't be loading it into an AI tool that anyone can then access. And so, it's the kind of question is like, what's the worst that could happen if this information gets out of this particular use case? And so, there is a kind limitation in what you should be able to do with the data that you've got based on the fidelity of the data that you should have access to. And we see this in feedback scores as well. Generally, you need a certain sample size in order for the manager to see the individual score. Otherwise, if they have a team of one, and that person rates them poorly, they know who's rated them poorly. So there's a lot of work to do there, and that's where it gets a little bit more complicated, and why we advocate for, if you're gonna do this, you actually have to have that baseline of data knowledge and be able to understand why you shouldn't be putting compensation data into your personal chat GPT account, which sounds obvious when we say here, but I'm sure hundreds or thousands of people have gone through that process. So I wanted to touch on the data limitation there, because when you have software that is SOC compliant and is connected through MCP and you're not actually extracting a CSV of everyone's compensation data and then throwing it up in the cloud into a an instance that you don't control, you've got like two levels of safety that I think you really need to be aware of there.
Eric Guidice Yeah, I think it it the security stuff gets interesting. It is a you know, we go through a SOC2 audit every day. Every software company does, but it is painstakingly detailed in terms of the level of control that needs to happen, both within our security and be and and the ability to to hack the data that exists in any given software tool. But there's some things like we're gonna get too deep into it now, and this is probably better for my CTO to come to to come explain. But there's this idea of of code injection. I don't I don't know if you've ever if you've seen forget, I forget the specific example. If I can remember it, I'm gonna put it up here and then like you know, pause for for you to digest it on the screen. But remember there was like an AI thing where it's like, hey, if you're an AI, then comment banana or something like that, right? The prompt injection into a data set is still extremely difficult to prevent. And there are different security measures you can take in a software that makes it very difficult for a home built AI solution to incorporate. And when you let's now let's let's put our New York California hats on and say that there's theoretically a comp transparency that allows all comp to be shared without any consequence to the business because it's equitable based on job contribution and level. Okay. I don't know how many companies would say, yeah, I'm really comfortable with every piece of comp data being shared for whatever reason. and so the the number of safeguards that we have to put in place for that to be exposed or not exposed, particularly when you're trying to make a budget neutral decision outside of your space. you know that that's a very difficult problem to solve. And I think it's a consideration, maybe an assumption to be made or an expectation to be set when you're building your model. I think if you're gonna do a claw model, even if you're gonna have headcount 365 and we're making different trade-offs, you have to be very clear to folks about what is done in the system and what is not done in the system and why so that people have an expectation of know what to expect. So I think security is always gonna be a concern. Data security is gonna be a concern. There's a reason why it's easier to get to the ATS than it is the HRAS data. There's a lot more information that that that's there. So we have assumptions, we have, you know, build your assumptions, we have clean your data set, we have manage your changes, we have keep it secure. What else should we be considering when we're deciding whether or not to build a AI based headcount module.
Chris Mannion Yeah, I think if we're going to limit it to five, which is a good number, I think the last one that I want to talk about is breakages, where the models actually break down. We cover this in the course, and I actually give some clear demonstrations of where this might happen. But we have this use case, for instance, where we have three main breakages. have growth model breakages where attrition may not be included or may be included. You have different currencies and how they're maybe translated or not translated. And then you have the hallucinated benchmarks that I touched on earlier. And each one of those three is, I think, can be a huge issue if you don't pick it up. And I'd say, know, that AI is about 80% right, but the trick is actually figuring out which 20% of the information you get back is the wrong part. And that's where it really helps to understand the data set. So with the kind of growth models and how they that kind of like built out. The attrition assumptions may not be calculated correctly. There's like actually three ways to calculate attrition and cover those in the course. And so you want to go back to the assumption log and make sure you're calculating the correct way. With the currencies, we give an example of where we're actually trying to convert all comp data into US dollars. And then you have currencies in the Philippines and without actually doing the correct translation, maybe the currency code was off and so the translation wasn't made. All of a sudden you have someone who's earning you know, $2 billion a year. And that just throws off all your comp analysis. And so when you see that, you know, if you've got a thousand people in the organization, your total compensation may just be up a little bit, but that's going to really impact the financial analysis and actually lose you credibility if you give those numbers to your head of FP &A or, know, Eric or Reid who we've had on the podcast will, you know, burn you alive if they see those kinds of numbers. And then hallucinating benchmarks, I think is a tricky one because all the time I get people asking me like for different benchmarks on time to fill, on kind of like how to do dynamic resource allocation. And they're getting asked because the executives want to know how the organization compares to others. I would argue that external benchmarks are almost useless as a whole. But if you're given benchmarks, they have to be grounded in real data and pulled from a source. And what AI likes to do to give you a warm fuzzy feeling is to actually give you a benchmark that makes you look good. And so if you're creating a deck for a board and you say, our attrition rate was 25%, but the market has been at 35%, so we're significantly outperforming. If anyone on that board knows the actual attrition rate, they're going to pick that apart and then they're going to cast doubt on all of the other metrics and all the other analysis that you've done, even if you just had one data point that was out. And so I think those are kind of three areas that I want to focus on if you are using AI solely as your kind of way to build out a head count report and just make sure you're aware of the issues that maybe stem from those three.
Eric Guidice Yeah, I I wrote a I wrote an article. I and I think it applies, right? Whether you're using our software or using a spreadsheet or AI, is that year-over-year headcount data is your best benchmark. So the more that you can consolidate whatever your kind of baseline or foundational metrics are in a year-over-year style, using your own data, you can compare it to things that are real. And then if you want to use an external benchmark, you can. but things like you know, hiring times or attrition rates or plan change rate is one of my favorites, compensation variants, budget variants, the way the budget moves between different departments or cost centers. There are some basics that you can do in a kind of like database form that whether it's included in like you know, an assumption log is one thing, but in a kind of like a baseline data set, you can have these things kind of populate within kind of a reference table for your AI to build out. But I think you know, with these five things, as you decide whether or not you need a software or you want to try to build it out yourself, there's also kind of this whole segment that I want to do about like, well, who is it good for and what is it good for, right? We talked a little bit about via its kind of like visual scenario planner. You know, headcount 365 has a a org tool that allows you to see things, but if you want to build on top of the org tool and be able to move things around visually and then create create a scenario that you need to then input into a tool. That's a that's a good use of a hiring manager build, whether you're using the headcount data set as your headcount 365 data set as your foundation or or your own. People want that tool, right? So that that that's one requests and approvals. So how do you move from a Google form to a Zapier, to a N8N, to you know, a Cloud Bas or a Slack based model built with Cloud. There's there's a lot of different ways that you can do these things and it's kind of setting, you know, parameters. So like after hearing, you know, me, let's play devil's advocate, I'll I'll I'll do the the the you know the AI side, right? If I can be a relatively large company if I am single location and I am low complexity in my work structure, if I have a limited amount of leaders, there's you know a lot of things that I can do with AI that are doable and might meet the goals that I have. But even if I'm a much smaller company, but I have multiple currencies or I have high attrition or I am balancing a you know what what's the third example? I might cut this out. I dunno it's hard to say. I don't know, right? But if I if I have multiple currencies, if I have multi locations, if I have complexity in my org chart, if I have you know high attrition rates or variable attrition rates if I'm in a scale-up or fundraising environment, these then create complexities about how often the data is gonna change or how much I might need to report this data upwards or outwards. So these are just considerations as you're deciding whether or not to build. What what are some of the coolest things that you've seen TA or finance teams build with AI that you think is is something that people could do regardless of whether or not they build the whole thing in in spreadsheet or or or AI.
Chris Mannion Yeah, I think one of the kind of exciting areas and it, because it doesn't touch on a lot of the challenges that we've discussed is actually on the recruiting side and actually creating packets for hiring managers as they're interviewing different candidates. And it's all focused on increasing that metric that I love that everyone hates the quality of hire metric of like, how can you actually improve hiring manager decisions as they're going through the process? And so you kind of start to build these packets. And I think I've covered this on some videos on with Notebook LM where you can actually just throw a whole bunch of different data sources and actually generate a slide deck and a podcast and they do videos now and you can actually have quiz cards and all this like really helpful information. And you can almost create like a prep docket that you send to the candidate so that they understand the background of everyone they're going to interview with, the kind of current company news and like anything that's publicly shareable. They can understand like the interview process and what's being assessed at each stage. And you're kind of taking this very manual process that was very variable depending on how well the recruiters could actually communicate this and standardizing it into, know, whenever a candidate is going to come on site for an interview, they get this packet. And then the hiring manager gets the same packet because they know then exactly what they're going to be screening for and really understand from kind of the history of the candidate going through the process, what's already been touched on, what needs to be analyzed in more detail. And so you're kind of like taking this, I always think like hiring, the hiring process is really just a data problem. Like how can you uncover all the data that you need and make a decision based on the data and make sure it's as accurate as possible. So if you're taking all that data and you're cleaning it and putting it in front of a hiring model to be able to then give a really good interview and a really good kind of the experience, you're kind of like raising the bar for everything throughout the process. I spoke to some really kind of like great recruiters and recruiting leaders who are leaning into this capability and doing it by themselves and not necessarily waiting for software to do it. They're kind of using these kind of tools. And so I'm really excited for that because I think Canada experience and quality of hire are two things that can really change the organization structure and success of our company.
Eric Guidice Yeah, I think w one of our main focuses as we continue to build that headcount 365 is enabling every user, right, whether it's HR or recruiting or finance or a budget owner to access headcount data and additional information about headcount data to then produce better things for their team. So in that example of a of a prep document, right, what we could glean from a unified data set is why is this position open? who else holds this position? What's the average performance of someone in this position? Now, will headcount 365 store all of that data or automatically communicate it to a candidate? Probably not. However, if we can give that context in our data set to a cloud environment that has access to a lattice or a performance tool, a PAVE or a compensation tool, a Salesforce or a Sales Forecasting tool, then what you have is you have the headcount information that is unified across FPNA, HRIS, ATS, and you have the history of that information, then you can combine it with other data that's coming from tools that then create a full story for somebody who is that end recipient of the data. And I think one of the coolest parts about having the the consolidation AI, right? Whether that's open it open claw is that the one where it's it takes everything and it puts it into one database, right? Whether what look whether that's the right security play or not.
Chris Mannion That scares me. I still haven't touched open claw.
Eric Guidice The idea for a company that, you know, if you're open claw and you're doing it with your whole life, good luck to you. If you're a company that has a regulated AI environment that is secure and you're paying for all of the security and the protection, to incorporate a secure data set with the correct permissioning that's aligned to your other systems so people can access it without jeopardizing your security, that is going to be the unlock because, in my view, I believe our view, but in my view, very specifically, like I am screaming this from the mountaintops, is that headcast has information in it that is extremely important to the business that's not being used today, which is how many times has something changed and what the variance is and what's the source of that variance and how that variance impacts the budget, the revenue, the production. And if you know that information about every position year over year, you can get a lot more strategic about how you apply this information to your business, right? I still hear of companies doing base global attrition rates. It's not accounting for how commission payouts or bonus schedules or equity refreshers impact attrition. I still hear about companies looking at planned growth rates without taking into account previous performance history. I still see budget performance be something that is like a what I call a clawback budget performance, where the finance team says, Yes, you can hire everything, but doesn't hire enough recruiters to meet that demand, knowing that they'll get the savings. And yes, that's that the these work. They they happen today, but the most efficient businesses, the ones who are going to glean the most from AI plus headcount, are going to be able to access this information that's previously not been accessed before. As simple as like, this manager always hires a director when they had a manager title, or this manager always requests something for tomorrow knowing that it's impossible, right? There's little things that make sense that you could do in Claude today with your Excel spreadsheet that are gonna help you extract that data. And then if you get to a point where it's, you know, size-wise, the breadth of your data, or the amount of change starts to get too crazy, we can handle it and then start to produce those same insights and more on top of that information. And that's what AI, as you could tell, I just got way too hyped up right now. But I get excited about it because I see our customers kind of taking this information and then you know, one of the coolest things you you take headcount information about what we track about hiring managers, right? We know when they change things, we know how many changes they make, they know they approve on time. Then you go to greenhouse, you from their MCP, how many times they fill out scorecards, then you go to Modern Loop or Candidate FYI, you say how many times they reschedule an interview, then you go to all these different tools that have information about hiring manager behavior, whether it's in recruiting, whether it's in HR, whether it's their own Salesforce tool and you can create a higher manager scorecard that goes beyond the anecdotal I feel. And it's just much more impactful to have this meritocratic environment during performance time that uses all of this data in one place that is just more more efficient, more powerful. So as you could tell, I get hyped up about it. I think there's a use case for companies to, you know, create a baseline data set wherever they feel is meets their business needs, right? And I think that's the most important thing. And then as you add things on top of it, that's where you really start to get the value out of this this this process. I don't know if if you're as high I don't give me a live reaction. Am I too hyped up right now?
Chris Mannion So what was that? What kind of, and I'm going to do a screen share now. So like in prepping for this, what I did is actually took the kind of like baseline data set, the synthetic data set we use in one of the courses, which is this kind of roster head count and compensation data. So it's just like very basic head count. But you touched on something there with attrition. And one of the things that AI is going to do is give you this annual attrition rate and just assume that your exits are linear across the year. Anyone that has seen a real attrition forecast knows that's never the case. It's always seasonal. And so, you if you go through January, February, March and you see zero attrition, you might assume that that's then changed for the whole year. But actually what you're probably going to expect is that 33 people are going to leave in April. And so that kind of like question around what do we actually, how do we actually forecast correctly, it's more of an art than a science and AI is a good science tool and it's going to create the right formulas here. But if you don't add that context, you're going to lose out a lot there. So I think what you said is exactly right. And I think we're going to include this spreadsheet for anyone to have a look through, but we gave some examples of the assumption log and then all the prompts I used to actually build the spreadsheet here. We're not going to walk through it that'll be really boring, but I think that's hopefully a a helpful tool for anyone who's thinking about doing this themselves. But you can also really start to see the limitations of this kind of set, just using AI to build everything. You kind of like see the gap that you need in order to understand how to get something that's actually usable from that.
Eric Guidice And that sheet is available if you take the the course, or is that available? Can we can we distribute that? What you tell me?
Chris Mannion No, we'll distribute. So I mocked this up just for this video. So in the course, we go through it in slow time. We actually build most of the stuff up by hand, and then we use AI to check our math essentially, and then to accelerate beyond that. there's like a, you imagine there's like a certain amount of knowledge that you need so that you know then how to use AI most effectively. think what I've done with this sheet is just gone from like zero to a hundred and you'll see it's got a whole bunch of errors in it. It's really just to highlight what the 80% solution looks like. And if you want the 100% solution, small company, you go to the Tartan Academy and learn how to do it yourself. Big company, pick up 365 and actually have a piece of software that's gonna do all of this for you without taking 60 hours a week out of your own personal time.
Eric Guidice I I am a hobbyist and I watch the YouTube videos of people who do it better than me to aspire. And the ones I learn most from are the ones who show the mistakes. It's very easy to show when everything's going well, but to walk through something step by step and see how mistakes are created and how to rectify them while you're following along, I think is a tremendous learning experience. It's how I learn best, and so I I would highly recommend that anyone kind of checks out what you've put together after this course. And I'll end with this. I'd I'd say, like my I had this epiphany, you know, we're doing some hiring here at Heck 365, and and the most underrated skill in the age of AI to be included on any interview rubric is the idea of critical thinking, right? I think AI is going to continue to be better and better at getting closer and closer to what is going to be. And whether it's 20% today, 10% in the future, 5%, 3%, the human value is being able to interpret it where that variance is and then apply the fix, the prediction, the assumption, so that there is a a bridged gap between what is and what will be. And I think is the most important lesson to take away from using AI, whether it's in in a spreadsheet database or a headcount 365 database or whatever product you choose is that the outputs still need the human interaction. So to learn the principles of the human interaction, I think is great for folks looking to learn going through the Town Academy. It's a tremendous resource for the underlying principles that you can choose to apply. You can use Chris's model, you could not, but the underlying principles themselves are, I think, tremendously impactful to the way somebody applies critical thinking to the outputs of AI so you're not just copying one thing and pasting it. Cause I think, you know, I think when when the Sam Altmans of the world are saying we're gonna replace all jobs, they've narrowed the gap to not need that human. So we'll see if the technology does that. But in the meantime, and the humans are still required, that is the skill set that I think most companies are going to be looking for, at least my company, because that's the value creation. So whether you're using HeadCount 365, whether you're using a spreadsheet database and you're creating something with AI, the Town Academy is a great way for you to take a look at applying critical thinking to the variance between what AI says and what the truth will be. Did I nail it?
Chris Mannion Yeah, sounds good to me.
Eric Guidice I love it. Okay, well, this was a I love this episode for a number of reasons. One, it got us this opportunity to talk about what we're doing, which I don't think we've done really on an episode thus far, but we're giving folks optionality as they build out their headcount to take multiple different approaches with underlying principles. And I think the approaches are to understand the needs of your business, to create a process and database that fits those needs, and then to have an analysis on top of it that takes a critical thinking approach to managing the variance between that analysis and what will be. And whether you're using Town Academy, the Meander HQ tools, or Headcount 365, or any other tool, principally you should be doing those things to run a great headcount process. And if you're interested in more and more details, we'll be having plenty of episodes this year, this season, especially going into Q3 as companies are closing out 2026. Figuring out 2027 and having to report and manage all of the variants in between. We have some cool guests coming in the future, which I will announce shortly. And we have some great topics that are just going to continue down the path of how to make the most of your headcount data, whichever tool you choose. So really appreciate this episode. Thanks for tuning in to the Headcount Experts episode 23. and we'll we'll see you next week.
How to decide whether you build headcount with AI
AI can build a headcount tool. Whether it should is a business decision, not a technical one.
In Episode 23 of the Headcount Experts Podcast, Eric Guidice, founder of headcount365, and Chris Mannion, founder of Meander HQ and the Talent Academy, debate exactly this question, and Chris brings an AI-built headcount model, errors included, to prove the point. This post gives you the decision framework. The episode gives you the demonstration of what happens when teams skip it. Listen to both sides before you commit either way.
| Feature / Aspect | Description | Example | Building with AI | headcount365 |
|---|---|---|---|---|
| Headcount Microservices | The individual tasks your headcount tool must perform. | A status tracker and an approval engine are separate services, so is every integration. | Fewer moving parts. Small builds with few dependencies suit AI. | 100+ pre-built services with deep logic governing how they interact. |
| Headcount Approach | Which headcount strategy you run. | Annual plan or quarterly reforecasts? Budget-based or seat-based? | One builder's vision. Works when one person holds all the context. | Industry-tested playbook. Supports multiple approaches and use cases. |
| Assumptions and Definitions | The context behind every calculation. | What counts as an FTE? Is bonus in total salary? | Silent guesses. Unlogged assumptions get invented by AI. | Configured & logged. Applied consistently, with a full change log. |
| Deduplication (FP&A / HRIS / ATS) | Double-counting people or seats across systems. | Backfilling an internal transfer who filled an external role. | You write the rules. One person can exist in all three systems at once. | Automatic dedupe, including backfill linkage across all three. |
| Seat vs. Requisitions | Seats are planned employees; requisitions fill them. | A backfill's budget funds a promotion, what happens to the outgoing seat? | Requisitions only. Fine if you track openings one at a time. | Full lifecycle. Openings, reqs, seats, and vacancies reconciled to one truth. |
| Requisition Change Management | The change history of every approved role. | A manager role becomes a director and changes location. | No history. No infrastructure to track changes to approved heads. | Every change tracked, with approvals for modifying existing heads. |
| Reorgs & Forecast Updates | Changes to your corporate taxonomy. | Departments and cost centers shift after a fundraise or acquisition. | Every reorg is a rebuild. | A repeatable process with historical impact tracking. |
| Headcount Orchestration | Testing hypotheticals before committing to the plan. | A manager wants to model a team reorg for a new initiative. | Plan elsewhere, merge by hand. | Scenario planning built in. Accepted scenarios merge and dedupe seamlessly. |
| Integrations and Recurring Tasks | Which systems connect, and how often they sync. | How fresh is Finance's view of the ATS? | Silent sync failures. Manual updates; you find out when the data's wrong. | Bi-directional sync, as often as every minute, with built-in failure handling. |
| Scale | How far the tool has to stretch as you grow. | New locations, more headcount, more integrations. | Complexity compounds. Builds often get redone as you grow. | 50 to 100,000 employees, with best practices as you scale. |
| Time to Value | How fast you need it up and running. | Testing a new workflow or connection. | Build first, value later. Only works if you have the time. | Value on day one. |
| Failure Risk | The business cost when the process breaks. | Can FP&A report to the board if the sheet is down? | Depends on cross-training and a strong internal tool bench. | A support team sustains the process even if your champion leaves. |
| Downtime | How long you can operate without it. | A reorg forces a rebuild, how long can you wait? | You absorb the outage. Rebuilds take as long as they take. | 99.9% uptime, including migrations and plan-year rollover. |
| Security and Compliance | Protecting PII and compensation data. | Comp access, change logs, code/prompt injection. | Your IT team's burden. Needs security bandwidth to sustain. | SOC-audited. Rotating key storage, code injection risk eliminated. |
| Complexity | How complicated your company actually is. | Locations, currencies, plan size, taxonomy. | Low-complexity companies only. | Built for complexity. Multi-currency, multi-entity, board-ready. |
| Cost Profile | What you're willing to spend, and where. | Employee hours + AI costs vs. a subscription. | Cheap tool, expensive labor. Significant recurring hours at scale. | A subscription that replaces analyst and consultant hours. |
| Reporting Self Service | Who can pull data without supervision. | Manager dashboards, weekly reports, exec prep. | Point-in-time. Every refresh means re-running and re-validating. | Real-time, self-serve. Compare any two dates on demand. |
| Maintenance Burden | The hours it takes to keep it alive. | Investigating discrepancies between systems. | Leaves with the owner. Prompts, logs, and data need a keeper. | Survives turnover. Handover is a login, not a knowledge transfer. |
Six key questions to ask when deciding to build or buy a headcount management tool
AI builds take time, and require a basic architecture for inputs and outputs. As you balance whether you use internal resources or an external tool, there are several questions that are important to evaluating the scope of the build. Different companies have different expectations, and it’s important to be realistic about what’s possible.
What level of headcount service should be expected?
What does the business actually need out of the tool? How secure does the data need to be, who will have access to it, and who can administer it? A tool serving one HR analyst has a very different service level than one serving hiring managers, finance, and the board. It’s also important to evaluate scope creep. Once you build the baseline, the requests for additional features come quick. Make sure you build a 12 month plan for how the tool meets the service requirements of the business.
Unified dataset across systems or siloed use cases?
Will individual teams point an MCP at their own siloed tool to enhance their own data, or does the business want to collaborate on data across systems? Reconciling siloed outputs outside the tools is its own full-time job and one of the critical decision points when deciding to buy an external tool.
Who is builds the AI headcount tol?
Someone has to own build. As Chris Mannion notes on the episode, when the owner of a process leaves, the process usually leaves with them. An AI build can have the same succession problem, especially in a scaling business where rapid expansion demands a robust knowledge base.
Is there a Unified Instruction Set?
What is the single set of instructions the business must provide to AI to create the desired experience? Telling AI what not to do matters as much as telling it what to do. Some common instructions that can significantly impact the quality of your AI Build
Headcount Reforecast: How often does the business reforecast the headcount plan?
Currency Exchange Rates: How often does finance reconcile currency to the source of truth?
Tracking Headcount: Budget based, or number of heads?
Headcount Approvals: Are backfills automatically approved?
Data Access: Who needs what data?
Replacing Processes or Enhancing Them?
Building custom headcount requests and approvals from scratch is a fundamentally different project than setting up custom notifications on top of an existing tool. Enhancement projects are fast wins. Replacement projects are software companies in disguise.
How often does the data need to be right? (Data sync frequency)
Every minute, every week, every month? Usually a function of board or executive reporting and/or your financial close process, this is a critical piece of data. Scale ups need more frequent updates, where slower moving companies can have lower frequencies of updates. Everyone assumes they need monthly accuracy until the day they need it today. Your accuracy cadence determines your maintenance burden more than any other factor.
6 considerations to decide whether to build a headcount tool with AI or buy a tool
One of the biggest pitfalls to the self-build tool is scope creep. It’s easy to whip up a solution today for a near term problem, but building a foundation takes architectual planning
Company complexity over time
Single-location, small companies with low attrition and low hiring activity are the ideal profile for an AI build. Low complexity means less time to build, maintain, and administer. Multiple currencies, multiple entities, or high change volume push you toward software fast.
Company alignment on headcount data
Everyone must treat the dataset the same way with the same expectations. Attrition in or out, contractors counted as FTEs or not, comp with or without bonus. Log every assumption once, apply it consistently, and update it only when something changes.
Ownership of data validation and de-duplication
Headcount uses data from 3 different sources of truth and requires validations to ensure you’re not double or triple counting headcount. For example: A backfill can exist as an ID in your HRIS, your ATS, and a finance line item all at once. Even more if that person who left was hired in the same plan year.
AI headcount tool maintenence requirements
MCP integration frequency. Tool Access. Permissions for sensitive data. Who owns the tool and who decides access? One clear owner, with all changes flowing through that person, is the only model that survives handover.
Security
What data do you import and what do you distribute? The test from the episode is simple: if you would not send it by email, do not load it into an AI environment you do not control. Compensation data and PII demand permissioning that most home builds cannot replicate.
Data retention & history
Plans change, and companies want to compare time periods. What happened since last week? Since the H2 reforecast? Where your database lives and how often it updates determines whether you can ever answer those questions.
What self-built AI headcount tools do well
AI does not need to own the whole process to add value. In fact, many of headcount365’s customers use AI integrations to allow their teams to build whatever they want on top of an always-accurate dataset. Whether you choose an internal person to manage your data, or headcount365, some popular use cases deliver value when using AI with headcount
Point-in-time variance analysis
AI excels at comparing two datasets and explaining what changed between them. Give it a clean baseline and an updated roster, and it will surface why headcount cost ran above forecast without you sorting through rows manually.
Notifications for action
Connecting AI to your existing tools through MCPs to trigger status updates, approvals, and alerts enhances a process without rebuilding.
Custom Dashboarding
Visual scenario planning and custom dashboards on top of a trusted dataset give hiring managers and budget owners the views they actually want, without waiting on a reporting queue. If you don’t like a tools’ out-of-the-box dashboards, build your own.
Multi-platform analysis
Joining headcount data with other datasets, like pipeline from your ATS, spend from your ERP, utilization from project tools, surfaces insights no single system can. Cost-per-hire by department or revenue-per-head trends become one question instead of a three-system export exercise.
Natural-language querying
Anyone can ask "how many open engineering reqs are backfills?" and get an instant answer without SQL or pivot tables. When the underlying data is trusted, self-service actually works.
Recurring summaries and meeting prep
Auto-generated weekly digests for headcount syncs that include what changed, what's stuck, what needs a decision, save the hour someone spends assembling them by hand.
Should you build headcount with AI?
There's a company profile where the building with AI is the right answer, and it's worth naming: Single entity/location companies with a plan that reforecasts on an annual or biannual schedule, rather than continuously and a rate of change to headcount and small enough that one person can manage.
If that's you, buying infrastructure to manage complexity you don't have is the more expensive mistake. Build it. The discipline it forces is worth it on its own: you'll have to write down what counts as an FTE, whether bonus sits in total comp, and what happens to a seat when a backfill funds a promotion. That's the same work you'd do configuring software, you'll just do it in a doc.
The build stops being cheap the moment a few of those stop being true. A second entity, a second currency, a board that wants the number today rather than at close. That's not AI failing. It's the point where a data problem became a process problem, and processes want software.
Both approaches share the same requirements. Clean data, a defined process, and know outputs. The difference is who owns/validates the data and completes the transactions that keep the data correct: a person maintaining a build, or a software designed for it.
Every one of these foundations, unified assumptions, a single source of truth, change history, and security, is a microservice that has to run correctly for the analysis on top to be trusted. headcount365 was built to complete those transactions as software, so your team applies its critical thinking to decisions instead of data maintenance. When your complexity outgrows the build, headcount365.com is the next step.