People-Led AI Transformation with Bryan Hong
Podcast Overview
-
Eric Guidice: Welcome to probably one of the best episodes that we are going to record of the Headcount Experts. We are joined by none other than Bryan Hong and Chris Mannion, co-hosts of the Headcount Experts. We are here to talk about the people-led transformation of AI. Bryan is pretty special, not only for his background and his come up in the people space, but in the role that he has taken on here. So I won't spoil it too much in the intro, but welcome Bryan to the Headcount Experts.
Bryan Hong: Yeah, thanks for having me.
Eric Guidice: For the people who don't know who you are or where you're coming from, give us a quick intro. What job do you have now? How did you get into this job? How do you describe it to your family at Thanksgiving? What do you tell them?
Bryan Hong: Yeah, it's a bit of a unique scenario. I spent roughly ten plus years in recruiting. Started out at an agency, went to a smaller startup for software value-added resellers, so providing ERP systems to small and medium-sized businesses. But I would say my current company is probably the cool company. I joined Astranis at roughly about a hundred people as a founding recruiter, took over the team, and built and scaled the talent function to about four hundred and fifty, went through a couple of rounds of funding, and we currently manufacture satellites here in San Francisco.
Roughly two years ago, I decided to explore the world of AI and quickly became very, I don't know if addicted is the right word, but quickly became very entranced by the opportunity, the leverage, and the capability extension that it provided. I really started pursuing that, building not just tools and automations, but larger agentic swarms, setups, and orchestrators, which then led down a rabbit hole of me wanting to really try to own that and provide solutions not just to Talent and People, but the entire G&A and anyone who basically wanted it.
Currently, I'm the Director of People Ops and Systems here at Astranis, overseeing all services across the entire People Org. Parallel to that, I have another team that focuses on delivering agents and AI solutions across G&A. I've been very fortunate to have a lot of great backers and sponsors internally at the company, and as of a few months ago, I was essentially converted over to an engineer, owning my own infrastructure so to speak, and deploying agents at will across two different teams. That is currently where I am at.
Eric Guidice: You would have lost my grandma around agentic swarms, but I think for this audience, they are going to get it. I always make this analogy: I like photography. In order to take a great picture before the iPhone, you needed to know how to take a great picture with a camera. You no longer need to do that, so now people can express themselves without this blocker of knowledge.
AI, at least for myself, gives me less access in my company because of the level of infrastructure my code would touch, but nonetheless, the idea is much easier to express myself, whether it's a design idea or an engineering idea using AI. It's really done a lot in the space, particularly with code or using the code base to deploy agents to make things more accessible. I love hearing about that access for a people-led leader. Chris and I have both moved into startups following a people or HR style role, and it's not common. I don't see a lot of founders coming from people, though I hope to see that transform.
What do you think the benefit is of you doing it? It sounds like you've gotten a lot of support. Why is the company recognizing a benefit from seeing someone with your background deploy in this new engineering capacity?
Bryan Hong: Yeah, I think in a lot of respects in the past, the idea person and the person executing the idea were generally two different people. AI has closed that gap relatively quickly and is making it so that product managers who envision a great product now have the ability to prototype it themselves. It's not always easy to communicate exactly what you want to somebody who may not be a domain expert or doesn't have the experience of working side by side within your field for a large amount of years, and they don't really understand what's happening in your day-to-day.
When you consolidate those two individuals into the person doing the work who can also build and optimize the work with a much higher technical capability, that brings a very new value to a variety of applications, primarily in corporate functions. The unfortunate truth in my opinion is that people and talent technologies were generally last in line for innovation. Last time I checked, I've not heard anyone tell me that a payroll system has changed their life. I feel as if there are many opportunities for innovation, optimization, and improvement around both the processes in which we in people and talent function, but also with what tools we should be using.
Eric Guidice: Yeah, I couldn't agree more. Chris probably has a number of opinions coming from the analytics side of Wayfair, but our whole thesis behind the Headcount 365 environment was there's a data set that's not being used by the business. Now, whether you connect directly to the HRIS or you have someone who operates within that data set who wants to build on top of it or build tools from it, we try to streamline that process. But it's in the same vein of there being ideas that are great in people that don't typically have the budget, attention, or prioritization in the business that are still worth solving.
When there's a gap between the person with the idea and the person who needs to develop that idea, and the business is weighing cost-benefit, time allocation, or product allocation, now, especially during a time with discounted tokens during this growth stage of AI, the ambitious have this opportunity to exploit the discount that's meant to drive adoption towards AI.
For me as an HR leader, the reason why I'm excited about this, and I'm sure Chris is teeming with idea generation that he can bounce off of you, is that this is a very narrow window where we have the ability to prove what's possible out of a people and talent function, what's good with the data that comes from those systems, and how a business can use it to be profitable and more efficient. How can we add it to the executive meeting? How can we get people who typically see this as an afterthought or an administrative function to use it as a strategic function? That's my statement that doesn't have a question mark at the end of it. So Chris, help me out here. What am I trying to ask?
Chris Mannion: Yeah, I'd love to dig into a question and a use case because I think one of the things that AI has been able to do is really empower a lot of People Ops leaders to be their own analyst. Historically, I ran a team of analysts who would support TA leaders, and they always had to wait for feedback, data crunching, and then invariably the first run would never be exactly what they were asking for. Now there are tools available, and even just using Code for Excel, you can get to most of the information that you need.
One of the challenges that I'm seeing is that by entrusting a lot of the analytical work to an agent versus an individual, there is a potential to lose some understanding in that translation, to the point where you take that analysis to a senior leadership team and you have to defend that analysis in person. I'm curious if you've seen that in practice and if you have figured out a way to mitigate that. How can People Ops leaders use AI to be their own analyst while not getting caught out when the analysis they've done gets questioned in a meeting with the CFO?
Bryan Hong: Yeah, that's a really good question, and we have run into this in many cases. The most straightforward example is this: for our particular ATS, we have very limited purview within the default analytics module to have certain views. Our leadership expects a certain set of data, and the unfortunate truth is that view is not available by default. So in turn, we'd have to do this very roundabout way where we essentially connect it to an intelligence connector, going through an S3 or Redshift, so that we can go straight to the actual database and pull it into Excel.
Now, for whatever reason, you cannot automate the refresh of that, so someone literally has to go in every single week and click a button. For us to execute the weekly report, we built a custom MCP, and within that custom MCP, we narrowed in on a skill and defined a tool that the MCP uses that is available to the AI, which then executes Python. It becomes a very consistent, static execution of: here are the endpoints that I want, here is the formula to calculate it, and so on.
Now, if you were to ask the AI to look at other analytics, it will probably do what you are talking about, where if you ask it five dashboards in a row for the same dashboard, it may come back with three different variants. This is very possible. So for things that are very quantitative, you would generally want to err on the side of trying to staple or cement the actual methodology into the tool. That's where we've seen really great results for numeric-driven analysis.
Once you get to the qualitative side, this is where it starts to get more Wild West. I don't think there's any one way that you can definitively say, but a good example I had was when we started to hook up AI, we would ask it to go through anonymous feedback. The question was: are our interviewers interviewing for the qualities, responsibilities, and requirements that we've posted online? We found in many cases they were missing or misaligned.
For software engineers, sometimes the AI will come back and say you guys have an entire interview for debugging and that's nowhere mentioned on the job posting. That's very helpful. Otherwise, someone would have to sit there reading it. The other thing is around quality markers, but this becomes very complex where you generally don't want a large data set; it's more of a control group where you want to align a set of interviewers and build a framework. But even then, you have to take it with a grain of salt, because how do you judge or rate a confidence level for how AI is interpreting text? And for that matter, how much confidence do you have in the quality of text being written by every person?
There are a lot of dimensions to that, but it does open up a lot of new doors to answer questions that in the past we probably never would have tried, because reading through employee pulse surveys, performance feedback, or large bodies of text would be near impossible. Python is pretty close, but really you're just looking for signal marker words at that point. Outside of that, it opens up a lot of new doors. I don't know if Chris you have, but I have not seen an amazing qualitative use case where I'm 100% confident that it's quality and 100% believable. But perhaps you've seen something that's exciting.
Chris Mannion: Yeah, the point that you made that I really want to highlight is where a lot of people go wrong: dropping data into an LLM or chatbot and getting an output, but the variance of the output depends on the randomness of the model, making it quite hard to understand what the black box is doing.
The point you made, which is actually how I do most of my work now, is using AI to build out a code base. The code base provides repeatability. Once you've built that code base out, you're not using tokens every time you run it, versus sending a two-megabyte CSV into an LLM where you're going to burn through tokens pretty quickly. That's a really interesting use case. You can start to experiment with known starting points and see if interviewers are actually hitting key points in the structured hiring process or going off-script.
The way we used to do that would be a huge People Analytics project to sample transcripts from interviews and use very basic NLP methods to look through the transcript and figure out if we're hitting keywords or matching sentiment. What you're suggesting is building that infrastructure out so every time you have an interview, you can just run it through this process and get an output. I haven't seen anyone doing that at scale.
Some things I've gotten quite excited about are when a model is used on top of a known scoring framework for things like performance reviews, where there is a clear requirement for the level of detail expected. The model compares a performance review submitted by a manager to the definition HR has scoped. If the score drops below a certain level, the review automatically gets sent back to the hiring manager to update. That's not something you could realistically do with a whole team of HRBPs, but the model allows that human-to-human interaction to be more productive. That elevates what talent and People Ops are really good at, which is empowering individuals and growing them. I'm curious if you've seen anything else in that sphere or experimented with anything that's been beneficial.
Bryan Hong: I would say one of the more beneficial qualitative analyses we've done is trying to identify the right talent and matching that with signals in the hiring process to correlate to actual high performers. There were quite a few findings that were surprising. One of the most surprising was the opposite: people who typically don't perform well were generally all very good interviewers, almost perfect. With AI at this point, that might be even more prevalent because everyone has help. But we found that most of our highest performers at one point probably had a small error during their interviews. That was probably one of the most surprising facts, but it helped us define what we should be looking for and how we curbed focus areas for our interviews.
Eric Guidice: It's interesting to hear how you set a baseline that allows people to pull the hiring trigger once someone is above a certain average after doing a sample group. There are these automations extracting a people idea into an AI transformation so it's broadly understood. As I think about how the business starts to care about this, it's about the findings you bring back to them, changing behavior, and drawing it back to the headcount conversation.
There's capacity demand in headcount, but there's also capacity demand in interviewing. If you can prove via these findings that to get to the optimal choice you need less effort, then you need fewer recruiters and hiring managers to meet demand. Your capacity demand forecast for headcount looks wildly different, and your cost of finance goes down. There are broader implications to a people-led finance process that are super interesting.
On the interviewing side, it is correlated to headcount, but specifically regarding headcount management, what is your general approach to headcount and how have you applied these AI tools to optimize, change, and make more efficient the way you do headcount today?
Bryan Hong: Yeah, headcount is a very tricky topic, especially for startups. If you're Apple, it's quite easy to measure ROI per employee per iPhone. What we try to do is understand the resources we need to meet business objectives, paired with a method to justify that a resource is required. Then there's a shuffle of prioritization at the executive level.
What I've been able to do over the last couple of years is address the fact that typically we'd have to wait until the planning process is completed to start executing, meaning from a talent perspective, we'd be behind three to six months. So we built a capacity planner for the team using about fifteen inputs, including internal benchmarks, external benchmarks, and market data. We paired that with a point system (1, 2, 3, 5 points). Each recruiter has a capacity of points; senior recruiters have 15 points, junior recruiters are closer to 10 points per quarter.
Points are distributed to individual roles representing difficulty. A role with a high level of seniority, rare skill set, and not in the local region would be a 5. A more basic administration level with a surplus of local talent would be a 1. Knowing the recruiter's historical performance over time, we understand their average contribution. We can map out target deliverables per quarter and reciprocate to executives: if this is the amount of individuals you need, here is what the current team can deliver. If we need more or less, it tells you how many and what types of recruiters you need to add. It turns the conversation into a clear trade-off discussion about staffing and growth.
Eric Guidice: So you're using AI to drive the algorithms behind workload distribution, the staffing model for recruiting, and ultimately the conversation with business executives about trade-offs between recruiting with existing resources or supplementing with outside recruiters, contractors, or agencies?
Bryan Hong: To be clear, it was created and implemented before AI, which took a lot of manual verification. But at this point, Nick, who is currently leading talent here, is definitely leveraging AI to accelerate that process. Instead of looking up every benchmark and pulling all the data manually, it has been built into a system that facilitates the process in a much more expedited way using AI.
Eric Guidice: Have you delivered a headcount agent out from your tech stack yet, whether direct or adjacent? Is there something you've deployed in that way?
Bryan Hong: We have a very unique outlook on headcount here, so the way we plan would not facilitate an agent being very helpful, at least on the planning side. As Astranis builds satellites, we build them for commercial ISPs and data providers around the world, but also for the government. We have swings depending on whether programs land, which changes the trajectory of headcount rapidly across lots of programs and customers. Trying to do a formal, end-to-end, high-confidence headcount planning run is less valuable given how often swings happen. As with most startups, we do the best with what we can.
Eric Guidice: As someone who set out to solve that problem, I respectfully disagree, but only because I have satellite manufacturer and government contractor customers where a percentage of headcount is associated with net new contracts using algorithms incorporated into the process. I'd be curious to know your feedback on whether those are valuable, but we can take that offline.
When we look at people-led AI transformation, there are two main things: new ideas that haven't been extracted into the business, and the ability of a people team to drive adoption regarding the relationship between work and labor. For you specifically, how do you fit into the org chart? Who do you report to, who are your peers, and do you centralize AI deployment under a central leader or distribute it to individual teams and subject matter experts?
Bryan Hong: Here at Astranis, I currently report to the SVP of People. I have a manager who oversees People Operations and Systems, handling standard administration like payroll, ChartHop, onboarding, employee lifecycle services, and employment verification. Then I have two individuals: one who studied data science and acts as our AI engineer, and a forward-deployed engineer who works directly with teams, hiring managers, and leaders to define processes, optimize them, apply AI, and handle enablement training.
We currently have two different infrastructures: we own one, and central engineering owns theirs. Core engineering focuses on infrastructure, security, governance, identity, observability, and integrations. Solutions engineers or forward-deployed engineers are super-builders who deliver on specific workflow processes.
For areas like finance or people, prime candidates transitioning into these roles are former NetSuite administrators or Workday analysts because they have technical aptitude and understand workflows. My path from recruiter to full-blown engineer is less common, but there is high demand for people in the operations space who understand both workflow and technical engineering.
Chris Mannion: A framework from the SIOP Lead and Edge Consortium highlighted how AI impacts planning and future roles like talent engineer, go-to-market engineer, or finance engineer. It breaks down old task lists, automates what can be automated, and rebuilds new positions. The challenge is getting existing people to fit into these new roles. What would you recommend to someone coming from a non-traditional background in terms of tangible training, skill development, or projects to lean into?
Bryan Hong: My best advice is to use AI. I am a firm believer that you should use AI to learn AI. A lot of people get intimidated, which is 50% to 60% of the battle. Once you start using it, don't define yourself as the limiter. If you don't know what questions to ask, ask the AI what questions you should be asking. You have an unlimited knowledge partner.
I have a training I've built tracking my own journey from custom GPTs to YAML schemas for APIs, data-mining my own sessions so others don't waste time or have frustrating nights. I'll release that soon.
There are also great communities like Promptmates, founded by Jason and others. They take cohorts of people and put them through a process to ramp up on AI. I saw someone go from not being able to prompt in a chat to having her own GitHub with automations in a couple of weeks.
I would warn against flashy YouTubers. There are great creators like Nate Herk who do awesome non-technical automation tutorials, but be careful of videos claiming to automate your whole life over a weekend. Building something personal is easy; deploying at an enterprise production scale is very difficult. Be mindful of data security, privacy, and legal risks before deploying solutions at work.
Eric Guidice: We love Promptmates. Jason showed me his intake session work back in the day, and we'll put a link to their resources in the description.
Let's move into some reaction content on the latest in the headcount space. Article number one is by Benjamin Encz, founder of Ashby, titled "Why an MCP Alone Doesn't Cure Weak Recruiting Reporting". He outlines where MCP-based reporting works, but notes quality and cost tradeoffs. What's your reaction to Benji's article, Bryan?
Bryan Hong: A generic MCP will fall trap to a lot of those issues. But if you understand how to optimize an MCP for your own use case, a lot of that goes away. An MCP might call every API endpoint searching for an answer, but if you build tool bundling, skill definitions, and optimization rules, you don't run into data or cost problems. Permissioning is a real problem, but overall there are cool things you can do with AI and MCPs.
Eric Guidice: It comes down to how many companies can afford the time and engineering cost of having a Bryan build custom setups versus buying an out-of-the-box solution like Ashby. Chris, what are your thoughts?
Chris Mannion: Data ownership used to be my biggest frustration running tech stacks, having data locked up in a system where you couldn't pull it out for joins and analysis. Also, token spend goes up if you constantly send raw data to an LLM. Ashby pre-processing data before sending it to an LLM is a smart approach for most people who don't have dedicated data teams.
Eric Guidice: Next reaction: Parker Gilbert, co-founder of Numeric, posted about finance engineers, stating that the best finance engineers aren't accountants who do things slightly better, but engineers applying systems thinking. What's your reaction to that shift, Bryan?
Bryan Hong: I agree 100%. Basic AI functionality like building a dashboard or simple analysis will soon be embedded natively into products for free. The individual who can architect workflows and translate outputs into business impact, strategy, or pivots will drive real value. The days of just building a dashboard or fixing an Excel spreadsheet are coming to an end.
Chris Mannion: I agree as an engineer who moved into talent via supply chain. Systems thinking applied to a domain is extremely powerful. Whether you come from engineering or accounting, having both technical understanding and domain knowledge is key.
Eric Guidice: Last one: Destiny Thompson, Chief People Officer Consultant, discusses framing workforce reshaping around AI, pointing out that layoffs are citing AI as a reason, but savings aren't necessarily coming from AI itself. Bryan, what's your take?
Bryan Hong: AI is definitely saving money in certain areas, but whether that allows you to cut resources depends. Some startups would rather hire two senior engineers who use AI really well over ten new graduates because of output magnitude. At Astranis, we haven't had layoffs, so it varies by company.
Chris Mannion: Data on AI-driven restructurings shows several large companies made big cuts and then had to double back and hire later. AI automates tasks within roles rather than cleanly eliminating full positions, because oversight, reviews, and transition plans still require human involvement. It's a multi-year transition rather than an overnight workforce reduction.
Eric Guidice: History shows that reliance on technologies can lead to changing cost structures over time, and token costs may shift as adoption deepens.
We are at time. Bryan, I'll give you the floor to plug Promptmates, your own projects, training, or open roles at Astranis. What would you like to share?
Bryan Hong: If you want free builds and resources across people and talent, check out my GitHub under the username "onepromptman". I share builds, n8n workflows, Claude plugins, and frameworks there.
Eric Guidice: We will link Bryan's GitHub below. You can find Bryan at Astranis, in the Promptmates community, and on GitHub. Thank you for joining us on the Headcount Experts podcast. Chris, any closing statements?
Chris Mannion: Bryan is a great example of someone making the transition many people are looking to make right now. Hopefully, listeners can apply these insights. Thank you so much, Bryan.
Bryan Hong: Thank you for having me.
Eric Guidice: Instead of doom and gloom around labor markets, we should focus on new positions, opportunities, and the coming economy. Bryan is a leading example. Really appreciate your time.
Bryan Hong: All right, thanks guys.
Eric Guidice: All right, give it two seconds to wrap up.
What happens when people teams have engineering resources?
AI eliminates the operational divide between functional domain knowledge and technical execution, allowing corporate functions (G&A, HR, FP&A) to construct tailored software solutions without competing for central engineering resources. Bryan Hong is a shining example, leading an effort at Astranis, a San Francisco-based aerospace company that builds small, low-cost telecommunications satellites for geostationary orbit (GEO)
““In the past, the idea person and the person that can execute the idea were generally two different people. AI has closed that gap... When you consolidate the person doing the work with the person who can build and optimize the work, it brings a new level of value to corporate functions.” ”
What do the headcount experts talk about with Bryan Hong?
1. Deterministic Architecture Over LLM Prompts
Passing raw CSV files or database exports directly into an LLM chat interface for quantitative analysis introduces structural risk. Language models operating without constrained execution paths produce variable outputs across identical datasets and generate unnecessary token expenses.
Enterprise-grade reporting requires a deterministic software architecture:
Model Context Protocols (MCPs): Custom MCPs act as constrained integration layers that connect language models directly to underlying data warehouses (e.g., S3, Redshift).
Programmatic Execution: Rather than allowing the LLM to calculate metrics probabilistically, the model calls structured Python scripts containing defined mathematical formulas.
Data Consistency: Caching the analytical methodology within code guarantees identical outputs across runtimes while preserving data governance and access control parameters.
2. The Emergence of the Domain Engineer
Organizational structures are evolving from generic operations roles toward domain-specific engineering positions, including Talent Engineers, Finance Engineers, and GTM Engineers.The optimal enterprise model uses a hybrid structure:
Central Infrastructure Team: Owns security layers, identity governance, model hosting, observability, and core API connectivity.
Forward-Deployed Domain Engineers: Subject matter experts embedded directly within business units who build workflow automations, custom agents, and functional tools.
3. Algorithmic Capacity Planning and Headcount Allocation
Headcount forecasting in high-growth or volatile environments requires moving beyond basic historical ratios. Modern talent architecture applies point-based algorithms to evaluate recruiting throughput against enterprise delivery goals:
Role Complexity Indexing: Open requisitions are assigned difficulty weights (e.g., 1 to 5 points) based on required skill rarity, geographic constraints, and seniority.
Bandwidth Mapping: Senior talent acquisition team members are assigned maximum point capacities per quarter based on historical performance baselines.
Executive Trade-Off Models: Algorithmic capacity forecasts transform executive headcount discussions from speculative estimates into clear financial trade-offs regarding internal hiring costs, agency allocation, or schedule delays.
Headcount Content Creator Reactions
Benjamin Encz | Founder, Ashby -MCPs vs In-App Reporting
Benjamin Encz argues that deploying a generic Model Context Protocol (MCP) over unstructured recruiting data generates high token costs, security risks, and variable outputs unless the underlying data is explicitly pre-processed and optimized for AI execution.
Parker Gilbert | Co-Founder, Numeric - Finance Engineers will be more like engineers than finance who works with AI
Elite future accountants will transcend manual reconciliations by adopting software engineering principles to design, orchestrate, and manage automated financial systems.
Destiny Thompson | CPO, Verano - CEOs are not seeing the cuts from AI they expected
Destiny Thompson argues that although many CEOs are using AI as a justification for widespread layoffs, actual cost savings aren't materializing because companies are simply adding AI tools to outdated operating models instead of fundamentally redesigning their people strategies.