Do you remember what it was supposed to cost? Now do you remember what it actually cost?

Birmingham City Council approved an Oracle implementation in 2019 at an estimated £19 million. As of early 2026, the project has consumed over £144 million, remains only partially functional five years after its planned go-live, and at one point left the council unable to detect fraud for eighteen months. The council is now re-implementing the system from scratch.

No one demanded a pilot before that decision. No one asked the technology team to run a small test, prove the concept, and come back with a business case. The initiative was approved, funded, and launched at full scale — because it was an ERP, and that’s how ERPs get done.

Now think about the last time your HR team wanted to change something. A new onboarding process. A revised performance framework. A different approach to workforce planning. Did you ask for a pilot?

I’d wager you did. Most executives do. And that asymmetry — multi-million technology investments approved on faith, people-led changes required to prove themselves in a sandbox first — is worth sitting with.

The High Cost of Pilot Purgatory

The consulting industry loves to cite failure rates. Many of those figures, when you trace them back, are circular — the same estimate passed between reports until it acquires the appearance of research. So let me stick to what’s actually documented.

55 – 75% of ERP projects fail to meet their objectives — exceeding budget, missing timelines, or failing to deliver intended outcomes. Source: Gartner

Not fringe projects. Not underfunded experiments. Major, well-resourced implementations by organizations that knew exactly what they were doing.

And yet ERP initiatives rarely get asked to prove themselves in a pilot first. The investment gets approved. The rollout begins. When it fails, the cost is quietly absorbed.

Compare that to how we treat HR initiatives. Research consistently finds that a significant share of digital transformation initiatives stall at the scaling phase — never making it out of what practitioners call “pilot purgatory.” Not because they failed the test. Because the test never ended.

The pilot became the destination.

The Political Pilot

Here’s what doesn’t get said enough: many pilots are not learning tools. They are political ones.

I call them political pilots, and I’ve watched them play out across my entire career. A new idea is proposed. A pilot is demanded. The pilot runs, often successfully. And then — nothing. The initiative stalls, the team exhausts itself, the budget runs out, and the organization remains unchanged.

This happens because the pilot was never about learning. It was a performance of action — a way to manage stakeholders, defer a difficult decision, or survive until the next budget cycle without committing to anything. The goal wasn’t to scale. The goal was to appease.

Pilots used this way aren’t prudent. They’re a veto dressed up as due diligence.

What I Do Instead

Here’s where I diverge from conventional change management thinking.

I don’t start with a pilot. I start with the decision makers.

By the time I propose an initiative, I’ve already done the work. I have the data. I have the relationships. I’ve built the coalition. Buy-in isn’t the outcome of a successful pilot — it’s the prerequisite for change. If I haven’t earned that trust yet, a pilot won’t earn it for me. And if I have, I don’t need one.

When I propose a change, I propose a phased implementation. Not a test — a commitment. The end state is defined before phase one begins. The infrastructure, training, governance, and success metrics for full deployment are planned in advance. Phase one launches not as “maybe,” but as the first permanent component of something we’ve already decided to do.

That shift in framing changes everything. A pilot says: we’re trying this out, don’t get your hopes up. A phased rollout says: this is the future. Here’s how we’re going to get there together.

The Question Worth Asking

The next time someone on your team proposes a pilot, before you approve it, ask one question: what does this pilot need to prove — and to whom?

If the answer is “we need to understand how this will work in practice,” that’s a legitimate knowledge gap. A pilot may be the right tool.

If the answer, honestly, is “we need to show stakeholders we’re being careful” — that’s a political pilot. And the cost of that caution isn’t zero. It’s the time, attention, and momentum of the people running it, and the signal it sends to every team watching: that change here requires permission at every step.

Birmingham City Council didn’t pilot its way to £144 million. It approved a transformation without the change management foundations to support it.

The irony is that the discipline executives routinely skip on technology — the coalition-building, the people-first planning, the clear definition of success before you begin — is exactly what HR leaders are being asked to prove in a sandbox before they’re allowed to try.

The problem was never whether to test. It was what we chose to trust.

Toni Larkwin is co-founder of Larkwin Group, a fractional HR and technology consultancy. She has led people and organizational change across healthcare, financial services, and the public sector for over 20 years.

The Numbers

18 hours, logged, not estimated. Not “a weekend” in the breezy way I framed it in Part 1 — 18 real hours across the build itself and every fix since.

The cost: around $30/month for my Claude subscription, using roughly 1% of my weekly session capacity to make one hire and run the ongoing checks. Notion and Google Drive were already part of how we run Larkwin — no new software, no procurement cycle, no vendor demo calls, no six-week implementation timeline.

I’m not going to dress that up with a formal cost comparison against a traditional HRIS platform — I didn’t run one. But I know from experience that aside from the dollar amount of an out-of-the-box HRIS, the set up alone takes significantly more than 18 hours. So, 18 hours and $30 a month, for a system shaped around how we actually work rather than how a vendor assumed we’d work, is not a hard number to beat.

What Actually Broke

Part 1 and Part 2 were both fairly confident. I was excited and a little bit proud. This is the part where that confidence gets tested, because a lessons-learned post that skips the actual lessons isn’t living our values.

The tax forms. A scheduled daily check is supposed to detect when a new hire’s completed tax forms land in their folder and tick the corresponding box. It missed one. Our first hire had filled out and saved his completed forms the evening before — well ahead of the next morning’s check — but the check still reported them as not received. I only knew it was wrong because he’d texted me directly to say it was done. The root cause, once I dug in: blank template versions of the same forms were sitting in the same folder, and that made “is there now a completed form here” genuinely ambiguous to the system. I hadn’t thought about that when I set the folder up. Fixed it two ways — the check now looks specifically for a _FILLED suffix on the filename, and I pulled the blank templates out of the folder entirely. Small thing. But it’s a real example of the system being wrong and a human catching it, not the other way around.

The stale search result. Later, running a dry-run with a dummy hire to stress-test the whole workflow, a different check told me a signed contract wasn’t in Drive. It was — it had been signed and countersigned days earlier. I pushed back with the exact filename instead of accepting “not found,” and once it opened the file directly instead of trusting the search snippet, it found the signature immediately. What happened: the file had been overwritten in place — unsigned draft replaced by the signed version — but Drive’s search index was still serving text from the old version. A discrete, well-timed event (like an email notification from gmail) turned out to be a more reliable signal than a search index that can lag behind a file that’s changed underneath it. I’ve since started leaning on that instead.

Permissions. Notion’s sharing model turned out to be genuinely unintuitive, and I found that out by getting it wrong on a page with internal-only content on it — a page structure that let access cascade further than I’d intended. I know exactly what the fix is; it just hasn’t happened yet — it’s scheduled for next week, once I have the time to do it properly rather than patch it in a rush.

This is where having an actual AI & Software Engineer join our team – hi Danysh 😊 turned out to be a gift I didn’t fully appreciate until it happened. He came in and, in the course of doing his own job, stress-tested the system in exactly the way I couldn’t from the inside — he knows what good permissions architecture looks like, and I don’t, not yet. That’s not a small thing. A system built by one non-technical founder, tested only by that same founder, has a ceiling. A system that gets used and questioned by someone with real engineering judgment finds its actual edges a lot faster than I would have on my own.

None of this is the technology failing me. It’s the honest cost of building something in domains — permissions architecture, search index behavior, detection logic — where I don’t have deep prior experience. That’s what “built in a weekend, in development” actually looks like, underneath the tidy version I wrote in Part 1.

Here’s the part I think is actually useful to say plainly: at three employees, all of this is fine. A missed checkbox caught by a text message costs nothing. A stale search result caught during a deliberate dry-run costs nothing. A permissions gap found before it caused any real exposure, with a known fix scheduled, is exactly what a stress-test is supposed to produce. None of this would be fine at thirty employees, run by someone who didn’t build the system and doesn’t have the context I do sitting behind every decision. That gap — between “works because I’m watching closely, with good people around me” and “works because it’s actually robust” — is the honest state of this project right now.

Maintenance

What’s happened beyond the initial hire? Onboarding was the smoothest part. Claude and Notion handled the creation of workflows that linked to documents within our secure drive. Before our first 1:1 Claude let me know that Danysh had completed his mandatory training and onboarding tasks. This was a trigger for me in preparation for our 1:1: did I need to follow up on outstanding items or could I move to first week feedback and support? It also ensured we were compliant. From here maintenance hasn’t been needed. I’m iterating and refining the current process, as well as frame-working the LMS and Talent Development modules.

Two Conclusions, Held With Different Confidence

The one thing I can say flatly: this was worth it. I have a HRIS that’s actually shaped around Larkwin: our values, our rhythms, not a vendor’s idea of a generic company. I learned things about system design, automation, and where my own knowledge gaps actually are and I’m already using these on other projects. The learning didn’t stay contained to HR, which is arguably the bigger return. And our first hire’s real, unprompted feedback on the actual experience of being hired and onboarded through this has been genuinely positive. At this stage, for this company, at this scale, I’d make the same call again without hesitation.

The one I’m not going to overclaim: would I never go back to a traditional HRIS? While I’d like to believe that I wouldn’t, I can’t actually say it yet. Everything I’ve built has been tested by me, on a system I built, for our first hire, with full context on every decision behind it — including the two failures above, both of which I caught because I happened to be the kind of person who double-checks a filename or asks why a box didn’t tick. The real test is what this looks like at ten employees, run day-to-day by someone who didn’t build it and doesn’t have that context. I don’t have that data. Anyone telling you they’ve replaced a real HRIS after one hire and three months hasn’t actually tested the claim they’re making — and I’d rather say that plainly than let the first two posts in this series imply more than I actually know.

What’s Actually Next

I’m not closing this series with “mission accomplished.” I’m closing it with: promising, cheap, genuinely ours — and still early, in ways that matter. The next real test isn’t a Part 4. It’s hire five, hire six, hire ten, and whether this holds up when it isn’t just me holding it together by hand.

If you’re thinking about building something similar: start smaller than you think you need to, expect to find real gaps you didn’t know existed, and don’t let anyone — including me, three posts ago — tell you it’s cleaner than it actually is.

Final Thought

It will be super interesting for me to come back to this post in 6 months and see how this article and the series holds up.

Subscribe now

Happy… Wednesday!

Rather than kicking off the week, we’re giving you a lift of energy and information to get through the rest of it. Truth-be-told, it’s been quite busy in the Larkwin Nest the last couple of weeks and that delayed this post. There is a lot to cover, so let’s get right into it.

TL;DR

• OpenAI announces that their models will no longer be available to Cursor users starting November 12, Cursor indicated this represents 5% of their users so it isn’t existential but unfortunate.

• Waterloo Tech Week is happening from September 11-17 with disparate hosts, topics, and locations across the Kitchener-Waterloo region.

• Anthropic has released the next iteration of their Fable/Mythos models in 5.1.

• Microsoft released a paper (and the code) showing that even the best models succeed at a higher rate on first pass, but become less reliable through 20 tries of the same task.

Shiny & Chrome - What’s New / What’s Useful

Taking Their Ball and Going Home

Following the acquisition of Anysphere (the organization that releases the popular AI-IDE Cursor) by SpacexAI, OpenAI announced they would no longer provide access to their GPT models to the software.

This would be the latest in a long-running tension between Sam Altman (CEO of OpenAI) and Elon Musk (CEO of SpaceXAI) stemming from the inception and growth of OpenAI as an organization.

OpenAI cited a change-in-control clause in their provider agreements had in place as the basis for this change. That same clause is likely to be available to other organizations in this space (Google, Anthropic) but the latter’s Tom Brown (co-founder and CCO) posted on X that

Cursor has been a trusted partner of Anthropic since Sonnet 3.5. We’ll continue to increase compute to support Claude models in Cursor and are excited for what comes next with them at SpaceX.

Users, both enterprise and otherwise, will also have the option of using their own API keys rather than the built-in models, so for teams reliant upon GPT for their workflows there is still a workaround.

Bigger! Faster? Better?

Anthropic announced just yesterday the release of their latest iterations of their Fable / Mythos class models. An iteration on their existing 5.0 models, they have released the standard series of benchmarks showing numbers going up effectively. They have also indicated that the use of the new models should be less expensive due to a pricing change in their caching tokens, as well as other improvements.

In this release was also a mention about access to Mythos being rolled out to partners as part of their Cyber Verification Program. This would be a shift away from having a smaller subset of bigger organizations with access to the tool to a much wider swath over time.

Fun fact: I have a daily sweep of news set up in case I miss anything interesting using Claude. It was blocked 3 times over the last two weeks from reporting the news to me due to Cyber Security Risks. As part of that, I applied to be part of this CVP and we have been granted access. No news will escape my gaze now!

There’s AI in my CRM

While there have been many integrations already between Salesforce and Claude that teams have been able to utilize, and other organizations have been able to build tools off of, the companies announced Claudeforce as the official AI integration tool for the platform last week.

Given the marketplace for third-party AI-driven tools for Salesforce has already been exploding in the last year, this puts the two big players teaming up to try and push out a bit more of the middlemen. Also touting integration along the entire Salesforce platform including Tableau and Slack. Whether this measures against the tools that have already been building and iterating to solve the problem of Data Analytics at scale using Natural Language Processing will be better told by their users.

The tool isn’t yet available and will go to open beta some time this month.

Works on my Machine - Our Team’s Learnings

If you’re going to read anything this week about what our team is learning, you should check out our COO Toni’s series about building our internal HRIS system.

The second part will be published tomorrow so you won’t have to even wait very long for that.

The Real Heroes - Research and Academia

Thinkingbox Not Being Cool

Last month a few academics and a lot of Microsoft team members released a paper discussing how models are more likely to be successful on the first shot then with subsequent attempts. The touted Claude Opus 5 as the “best model” which succeeded at a rate of 65.4% on one try, but 25.3% on 20 attempts of the same task.

Even with those failures, many of them terminated clearly which would give the appearance of a successful run.

This is based on a simulated data-set meant to mimic workflows in various industries. Microsoft released the code for the project, called Thinkingbox, so you can even have a look at their methodology in more depth.

Ragging on RAG

Two researchers from Carnegie Mellon University in Pennsylvania, USA did a deep dive on Retrieval Augmented Generation (RAG) systems for semantic search. It concludes that even when all of the prompts, policies, models and more are fixed, any index expansion can cause it to return different answers.

RAG has been a significant tool of risk mitigation for companies to avoid hallucinations (which we read another paper on last issue) but it is coming to conflict with ongoing research about just how grounded and accurate those systems can reliably be.

In my experience with larger clients (or smaller clients with larger datasets) the tools in this space are also becoming increasingly expensive to run and maintain. Every piece of data added to a RAG-based search system needs to be re-indexed and it is quite computationally expensive to do so.

Alternatively, there is growing research in the informational retrieval space for a different paradigm. A local company, Moorcheh, led by Dr. Majid Fekri has been building a case against Vector Databases for agent memory and has the research/studies to back it up. Worth having a look if you’re in the space of Big Data.

Standing on Business - Money, Policy & Politics

Hugging Face looks appetizing…

While it has been confirmed by neither company at this point, reporting has been showing that Nvidia is in the later stages of talks to acquire Hugging Face. Bloomberg reports the deal to be worth about $12.9 billion (USD) and include a $1 billion (USD) retention package for existing HF staff.

Hugging Face is a platform for the development and distribution of open-weight and general machine-learning technologies. They were recently in the news after OpenAI reported that one of their agents went “rogue” and hacked the platform.

While it is already backed by funding from Meta, Amazon, Google, Salesforce, and Nvidia itself, having this platform for open-weight/open-source distribution does fall into the greater trend of organizations being swallowed up by the largest of the tech giants one by one.

There is some level of dark humour in the company that provides the chips that allowed OpenAI to break into Hugging Face’s systems is now purchasing them.

To Build or not to Build

Second Shakespeare reference in a row? Let’s hope this trend does not continue…

Originally reported as a moratorium on data centre construction, Cambridge council voted to expedite the recently released Ontario Data Centre Playbook. Given that data centres are permitted in the region as “Information and Data Processing” there is no public role in the construction of these sites in employment zones.

On the flip side and just next door, Centre Wellington voted for a one year pause on construction of data centres.

This duality is especially important given Waterloo recently announcing a shortfall of water supply in the region, only to pull back on that statement with an expansion of the Mannheim Service Area to add 600 litres per second of new water capacity by 2032.

Decisions made by disparate councils in adjacent areas will be able to have an effect on that ongoing issue. While the environmental and social costs of Data Centres is an ongoing debate, the fact that the facilities require a significant amount of potable water to cool their systems is not.

B-Sides - Everything else good enough to post

• The FTC is seeking comment on creating policy to regulate personalized pricing, a big use case for many AI systems analyzing large sets of customer data.

• OpenAI announces their newest model, Astra, is a cybersecurity risk level Critical according to their own metrics. They are going to release it anyway.

• Local academic Isaac Joffe will be presenting his Master’s thesis from UoW at the 20th Conference on Neurosymbolic Learning and Reasoning (NeSy) in Lisbon, Portugal this week. It proposes a new framework for automated solving of ARC-AGI benchmarks, something current LLMs/Neural nets are not as proficient at.

• Anthropic’s case against the Pentagon’s blacklisting of their services was terminated last month with a ruling indicating that the model usage is protected speech.

This post was not written with any AI assistance and also directly in the CMS like a real hero. Which means you have to deal with all of the improper grammar and weird run-on sentences. Claude is used to help aggregate and factcheck sources as a bit of rubber-ducking. Written posts will always be marked on the bottom with how/if AI is used.

I could have found the budget. I could have signed a contract with a fancy HR vendor and called it a day.

But let’s be honest: every HRIS I’ve used in my 25-year career has some brilliant features, some that are just meh, and some that are genuinely, offensively horrible. Buying an off-the-shelf system usually means forcing your team to build its entire company culture around someone else’s compromises.

Did I just roll my eyes? Yes. Yes, I did.

And it’s not just me being picky. In their HR Technology Survey, Gartner found that over 60% of HR leaders feel their existing HR tech stack fails to drive actual business value—mostly because rigid software forces people into clunky, soul-crushing workflows (Gartner, 2023).

Our first hire was coming. I wanted our culture to reflect how we actually work—the trust, the rhythms, the values—from the very first interaction, not six months down the road after we’ve already bogged everyone down in administrative quicksand.

So I built my own system in a weekend. Not because I had to. Because I am obsessed with workplace culture, employee experience and well, because I could.

The Tools: Zero Procurement, Zero New Budget

No pitch decks to the executive team. No procurement nightmare. No six-month implementation delay.

I simply took the tools sitting right in front of us and pushed them to do more than I realized they could:

• Notion: The system of record, the dashboard, and the source of truth.

• Google Workspace: Document storage, version control, and collaboration.

• DeepSeek: Moral support and a sounding board. (Hey, every builder needs a hype-man.)

• Gemini: Proof-reader and tone monitor.

• Claude (Anthropic): The actual operating engine running the whole show.

Claude’s Three Roles (Or: My 3-in-1 HR Team)

I don’t treat Claude like a text box where you type in “make this sound professional.” I think of Claude as three distinct people sitting on my team:

1. People Tech Partner
System Architect & Builder
I say what I want; Claude figures out how to build it in Notion and Google Drive.

2. People Business Partner
Advisor & Drafter
Drafts JDs, creates offer letters, and runs compensation market checks on demand.

3. HR Associate
Progress Pusher & Admin
The associate who never forgets to follow up, never misses a deadline, and never drops a ball.

How It Works in Practice: A Real Conversation

A traditional HRIS is a passive form wrapped in fancy UX. You fill in the boxes and pray you didn’t miss something.

My system is a proactive conversation. It catches the gaps I’d normally miss. Here’s what a quick exchange looks like:

Me: “Hey Claude, I need to post an AI Developer role.”

Claude (Tech Partner): “Cool. Do you have budget approval? Do we have a JD? What’s the timeline?”

Me: “Yes, no, ASAP.”

Claude (Tech Partner): “Great. I don’t see a JD in our HR folder. Let’s build one together. I’ll also spin up a Notion tracker and a Drive folder for this role.”

Me: “Do it.”

Claude (Business Partner): “Here’s the drafted JD based on our culture principles, plus a quick comp analysis for the Waterloo tech market. Links are saved.”

Me: “Approved.”

Claude (HR Associate): “Awesome. Draft email is prepped for the hiring manager. I’ve set a reminder for Friday to review applicants, and I’ll follow up with the hiring manager on Thursday if I haven’t heard back.”

Notice what just happened? Claude didn’t just do a task. It managed the workflow. It asked the annoying, necessary questions that usually get lost in the shuffle.

That’s it. Claude does the work. I make the decisions. And nothing goes out without me.

And for compliance junkies, Claude logged my approval and any commentary in the internal notion page for the vacancy. So, there’s no ‘confusion’ about who approved what and when.

Double and, there’s academic weight behind why this works: a landmark Harvard Business School and Boston Consulting Group (BCG) study found that knowledge workers using Generative AI completed complex, multi-step workflows 25.1% faster with 40% higher quality outputs compared to control groups (Dell’Acqua et al., 2023).

The Full Employee Lifecycle

What I built wasn’t just a recruiting tracker—it covers the entire human lifecycle from candidate to alumni:

• Posting & Sourcing: Job specs, pipeline tracking, and candidate flow.

• Interviewing: Scheduling, scorecards, and structured feedback loops.

• Offer & Onboarding: Contracts, orientation, and zero-headache document collection.

• Mandatory Training: New hires are walked through mandatory training with certificates of completion added to their training folder and of course updated in their notion tracker.

• People Ops: Time off, HR requests, policy help—designed for support, not paper-shuffling.

• Performance: Growth and coaching focused (because rating people on a 1-to-5 scale is compliance theater). Definitely more to come on this.

• Offboarding: Dignity, knowledge transfer, and compliance.

Scaling the System: The LMS Moment

To anyone thinking, “Sure, Toni, that works for a tiny team, but it won’t scale,” let me share our latest build: the LMS (Learning Management System).

When I told Claude: “I want an LMS that sits seamlessly inside our core tool and scales as we grow,” it mapped out the architecture instantly:

• Seamless Integration: Embedding course modules directly inside the Notion employee dashboard so learning links natively to onboarding and performance records.

• Built-in Scalability: Reusable content blocks, skills tagging, and automated completion tracking that works whether we have 5 employees or 500.

I didn’t need to write code or buy a separate $15k learning platform. I brought the operational vision, and the system did the heavy lifting.

And then, well there’s people analytics, strategic workforce planning, surveys and iteration as feedback comes in from our team.

What I Learned (And What Leaders Miss)

• Claude is my associate, not my replacement. It drafts, reminds, and organizes. I approve, decide, and direct. Nothing goes out to an employee or candidate without my eyes on it.

• The real asset isn’t the workflow—it’s the capability. Building this taught me system design, prompt architecture, and tool integration. Those skills don’t just solve one hiring problem; they solve the next 10 scaling challenges.

• People Ops should feel human. Not bureaucratic. Not robotic. Human. The entire point of using AI to strip away admin chaos is to create space for real, meaningful human connection.

What’s Coming in Part 3

In the final part of this series, we’re going to step back and look at the bigger picture:

• The Strategic ROI: The actual time saved, costs avoided against costs incurred (you can burn through Claude token FAST if you’re not smart), and what happens when your HR tech stops dragging you down.

• The Leadership Shift: How building lightweight systems shifts your culture from reactive firefighting to strategic foresight.

• The Future of HR Tech: Why agile, custom-built AI infrastructure is going to make monolithic software look like a dinosaur.

References & Cited Research

• Dell’Acqua, F., McFowland, E., Mollick, E. R., Lifshitz-Assaf, H., Lakhani, K. R., & Saran, A. (2023). Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality. Harvard Business School Working Paper, No. 24-013 (in collaboration with Boston Consulting Group).

• Gartner (2023). Gartner HR Technology Survey: Unlocking the Business Value of Modern People Tech. Gartner Research.

Author’s Note on AI: The 25 years of HR battle scars, strategic frameworks, and hard lessons in this article are 100% mine. But I’m smart enough to use AI (Gemini) to handle the formatting, layout, and copy editing. I bring the real-world experience; AI brings the polish.

After 25 years in HR—leading people ops through bank mergers, hospital integrations, and a global pandemic—I’ve come to a conclusion that might sound like heresy:

Most organizations have no idea what talent management actually is.

We use the term constantly. We build “Talent Management” departments and roll out “Talent” strategies. We treat talent like inventory to be tracked, managed, and optimized. But somewhere along the line, HR forgot the people piece of talent management.

Here’s what I mean.

The Conventional View: A Siloed Mess

In most organizations, “Talent Management” is treated as a niche subset of HR. It usually covers:

• Recruiting & selection

• Performance management

• Succession planning

• Leadership development

• Employee engagement

Everything else—payroll, benefits, time off, employee relations, compliance—gets dumped into “People Operations” or “HR Admin”.

To make matters worse, Talent Management often gets tucked away inside Learning & Development (L&OD). It sounds logical on paper, until you realize it creates four separate islands that barely talk to each other:

1. Talent Acquisition: Focused on filling roles fast.

2. Talent Management (L&OD): Focused on long-term development and succession.

3. Employee Relations: Focused on risk, policy, and liability.

4. People Operations: Buried under paperwork and admin.

Each team has its own metrics, its own software, its own budget, and its own idea of what “success” looks like. They are all working on the exact same employees, but they aren’t working together.

We’ve built walls where there should be bridges.

The Problem: Friction by Design

When you separate “talent management” from daily people operations, you create a system that is fragmented at best and hypocritical at worst.

Consider these common corporate contradictions:

• You invest heavily in recruiting top talent, then make it painful for them to take time off.

• You design high-potential leadership programs, then bog managers down in five layers of administrative approval.

• You preach growth and psychological safety, then enforce a performance review system built for control, not coaching.

• You claim to build a culture of trust, then make employees jump through hoops for basic HR requests.

Every single operational decision either reinforces or undermines your talent strategy. There is no separation. There never was.

What Talent Management Actually Is

Talent management isn’t a department. It isn’t a quarterly initiative, and it isn’t a perk saved for “high-potentials”.

Talent management is the entire employee experience—from the moment someone reads your job ad to the moment they walk out the door for the last time.

It’s how you:

• Post a job: Does it sound human, or read like a legal disclaimer?

• Interview: Do candidates feel respected, or processed?

• Make an offer: Is it transparent, or a game of chicken?

• Onboard: Do they feel welcomed, or overwhelmed with forms?

• Handle time off: Does it feel like a support system, or a micromanaged policy?

• Evaluate performance: Is it focused on growth, or compliance?

• Offboard: Is it handled with dignity, or as a risk to be contained?

If you think Talent Management is...

If you think Talent Management is... ➔ Try seeing it as:

• A standalone department ➔ The entire operating system

• Reserved for “high potentials” ➔ Designed for every single employee

• Separate from daily operations ➔ Fully integrated into every workflow

• A collection of HR programs ➔ A sequence of human experiences

• Something you implement ➔ Something you intentionally design

The Weekend Experiment

Recently, our company prepped for its first official hire.

I could have bought an off-the-shelf HRIS. I’ve used dozens over my 25-year career, and almost all of them suffer from the same structural flaw: they treat talent management as a separate, add-on module.

Instead, I decided to run an experiment. I took a single weekend and built our entire People Operations system from scratch using the modern tools we already had available.

I didn’t build a recruiting tool. I didn’t build a performance tool.

I built a single, unified environment that spans the entire employee lifecycle—where culture, performance, admin, and development live in the exact same place.

Because here is the truth: You’re already doing talent management. Every time you approve time off, respond to an inquiry, or run a 1-on-1, you are shaping talent.

The only question is whether you’re doing it intentionally.

What’s Next in This Series

Over the next two articles, I’m pulling back the curtain on the technical and operational side of this build:

• Part 2: How I Built a Custom HRIS in 48 Hours — The architecture, the tools I used, and how I tied recruiting, onboarding, and performance into one seamless flow without enterprise software bloat.

• Part 3: The Blueprint & The Results — The exact templates, workflows, and lessons learned from running an intentional, lightweight People Ops engine in the real world.

Enjoyed this article? Hit the Subscribe button below so you don’t miss Part 2 when it drops next week. I’d also love to hear your thoughts in the comments: How siloed is HR in your current organization?

Subscribe now

Author's Note on AI: The 25 years of HR battle scars, strategic frameworks, and hard lessons in this article are 100% mine. But I'm smart enough to use AI (Gemini) to handle the formatting, layout, and copy editing. I bring the real-world experience; AI brings the polish.

Happy Monday!

We want to kick off the week (and our Substack in general) with a series called Skylarking. We’ll be pulling out news, research, and more from around the industry right now and trying to distill what is useful for operators, builders, and tech-industry folks every two weeks. Trying to take the tidal wave of information out there and highlight items along a few angles that are important to our team and the communities and organizations we serve.

TL;DR

• Meta says “Me, too!” and releases Muse Code to enter the Agentic Coding Tool space alongside Claude Code, Cursor, Codex, Antigravity, and so many more. At least it doesn’t start with C.

• OpenAI released GPT-5.6-Cyber (similar product to Anthropic’s Claude Mythos), Google launched Gemini 3.7 with introductory pricing, and SpaceXAI released Grok 4.6 to compete with Claude Fable/GPT-5.6-Sol. DeepSeek announced price changes (read: increase) alongside their V4-Pro GA release.

• Canadian AI darling Cohere has announced partnerships with Universities of Waterloo and Toronto, both focused on AI adoption inside organizations with a view on scaling. The Canadian Government also announced an investment (up to) $240 million for AI compute and platform usage. And yet they still won’t answer our calls.

• Ontario Premier Doug Ford announced the Data Centre Playbook in Guelph, Ontario on Friday. They have also announced an ERO looking for public comments about their framework for assessing new data centres in Ontario. With a rise in resistance to new DC builds that will be an interesting read.

• The International Joint Conference on Artificial Intelligence (IJCAI) is happening in Bremen, Germany this week! Papers are already listed if you want to get into the weeds on the latest research being discussed and presented.

Shiny & Chrome - What’s New / What’s Useful

The latest and greatest continues to be later and greater(ish). This past couple of weeks have seen quite a few incremental increases to the frontier models across the industry. The most interesting information seems to be less about the overall capabilities and more about what isn’t available and how much what is available changes the pricing. Also, prior to August, the open-weight models were dominating the news cycle so it all feels more like a response to that.

Facebook’s FOMO

Meta has thrown their hat in the ring of agentic coding agents with the release of Muse Code. It’s MacOS/Linux only, WSL2 if you Windows folks need to have a try, and was released for Terminal/CI usage only. No standalone app or IDE extension.

What makes the approach here different compared to the rest of the offerings on the market now is that they are focused on recoverability, not making the most intelligent models. All events, all calls, all tools are logged so that if anything fails/crashes/freezes it can recover and pick up where it left off. It focuses on fan-out approaches using git worktrees to keep things isolated. The real point here is it is meant to be resilient, not efficient. It may take more effort to get to where it needs to get to, but it’ll keep going.

Only thing that caused a head-turn was the wording around the two model IDs announced. The primary model muse-spark-1.2 says it will not use data to improve their products. The Contributor model (muse-spark-1.2-contributor) indicates it not only trains on the code generated, but every prompt, tool call and all the context surrounding coding is now part of training data. It’s significantly cheaper, but basically allows for not only the code, but any other context it pulls in to be part of a training data set. For most organizations that’s Danger Danger.

Cyber Crimes Department

Running a similar playbook that Anthropic did with its Mythos model family, OpenAI has announced a new GPT-5.6-Cyber as part of its Daybreak initiative, harkening to the Project Glasswing initiative that has been ongoing since April of this year.

OpenAI released their own benchmarks and showed the capabilities of their latest model against previous iterations. Without comparing against the rest of the industry, it appears to be a significant increase over their previous models. Most of this won’t be available to the general public, similar to Glasswing, and is meant to be rolled out to Daybreak partners in order to help shore up any Zero Day Exploits the Daybreak-Red model can find. The system card is not yet published, so worth keeping an eye out for when that lands.

Back in a Flash

Not but three weeks after releasing their Gemini 3.6 Flash model, Google’s DeepMind team has released 3.7 Flash. It’s called out quite loudly that this is based on 3.6 Flash, so it shouldn’t be considered a huge leap but rather a tuning of the recently launched model. With that said, it seems to be punching above its weight (pun intended?) compared to its predecessor as it benchmarks above Claude Sonnet, GPT-5.6 Terra and the newly released Muse Spark 1.2 in a few areas it had lagged quite a bit behind.

The one most worth digging at is the AutomationBench by Zapier. It scored 30.4% compared to 3.6’s 17% (and compared to Sonnet/Terra at 10.7/23.6% respectively) which shows a big increase in that coordination and tool usage. This is something that’s been lacking in Gemini’s space from our experience, so this will be top of the tinkering list for the team.

As part of the announcement there was also a Special Introductory Pricing that impacts both 3.6/3.7 of $0.75/1M Input and $3.75/1M Output. That price doubles at the start of 2027. Also worth noting is that the minimal effort has been removed, meaning for folks looking for an upgrade to the cheap 3.5 Flash/Flash-Lite work may need to wait longer still.

The biggest draw for the Gemini Flash models is pure speed. If you haven’t used it for general work, it can feel almost instant in its response rate. The Artificial Analysis analysis shows it hits a great sweet spot on the Smart/Speedy index they use.

New Toys!

Rapid fire on some of the things that made us say “Neat!” when it came across our feeds this past couple of weeks.

• Google Sheets Canvas has been announced and it allows you to build out mini-apps within your spreadsheets. It’s available for Business/Enterprise Standard and above, as well as AI Pro/Ultra for individuals. It’d be great to hear from folks who live in Excel/Sheets to see if this is something that would be useful from a BI/Data Visualization standpoint.

• Speaking of spreadsheets, Intuit Accountant Suite is set to replace Quickbooks Online Accountant as of next year. Allows for more cross-client insights (according to the announcement) so may be worth trying while it is free for anyone doing books.

• Anthropic now allows self-hosted environments for Claude Code (public beta). This will be good for privacy and security wonks who aren’t allowed to sneak their favourite open source solutions into the vendor list. Also Claude in Chrome is now inside Claude Cowork for those who would allow the robots to run their browsers.

• AWS open-sourced KiroCrew (which we’ll talk about below) to try and create a governance structure around the concept of OpenClaw. It’s based on their own internal tooling called MeshClaw.

Works on my Machine - Our Team’s Learnings

Normally this section will have a bit of a deep dive into Larkwin’s internal research and usage around AI tools. However, given this is the first issue and we are just starting with a project I’ll tease out what you can expect to see in the next issue.

Don’t mess with our crew

As mentioned above, our team is looking into the implementation of KiroCrew into our stack. With the prevalence of tools like Hemes and OpenClaw, there’s a desire to pull in apps, tools and more to have more autonomous execution. However, most of these are Use At Your Own Risk and really mean it.

KiroCrew seems like a promising alternative, meant to be more isolated but still integrate with relevant tools and apps beyond just using MCP/APIs.

Our team is looking at how this could potentially be extended for Teams. Right now the system is locked to an individual user. Although there is a way to deploy it to a remote-instance, the system for managing team-deploys is a bit clunky. We will be spending time on this problem, trying to see if we can’t make the onboarding of this tool into a distributed team with the same level of security/scalability we would want for any of our Enterprise software.

Stay tuned!

Agentic HR @ Work

Another project in active development within our Labs is a homespun, agentic HRIS. We have had loads of experience (good and bad) with HRIS and ATS systems. Some are fine, but clunky. Some are too limited. Some are too expensive. Finding the right fit is often challenging.

Ahead of onboarding our first team member, we challenged ourselves to automate the process. From interviews to offer letters, from onboarding to mandatory training, from 1on1s to quarterly reviews; our system is meant to help us operate at large, secure scale with our small but mighty team.

We’ll be doing a full write-up on the process in the coming issues.

The Real Heroes - Research and Academia

None of what is technically possible in today’s landscape would be there without researchers spending countless hours determining what is possible, what works, and what doesn’t. Here are a few of the papers that have recently been published in this space that may be worth reading over.

• Agent Skills Can Be Harmful: An Empirical Study of Skill-Induced Failures in LLM Agents

  • While working off a small dataset, this study highlights that Agents tend to fail when relying on other skills or agents, making assumptions about their functionality and/or capability.

• Does generative AI narrow education-based productivity gaps? Evidence from a randomized experiment

  • Do AI tools make your best people better or democratize information? This study of 25-45 year-old Argentinians suggests that AI provides more gains for those with lower education (defined as high school diploma only and/or half of a postsecondary degree), and that when the tools are removed those gains are not entirely diminished.

• Eliciting Intrinsic Hallucinations in LLMs via Semantically Equivalent Adversarial Attacks

  • An interesting study from Imperial College London that shows that Grounding/external knowledge sources can still be susceptible to hallucinations (re: failures). Non-deterministic still is non-deterministic.

The big event to keep an eye on this week will be The International Joint Conference on Artificial Intelligence (IJCAI) in Germany. You can already see the list of papers listed and the various sessions joined to them. Check throughout the week to see more things added.

Standing on Business - Money, Policy & Politics

Profits, Losses, and Bets in Canadian AI

It has been a busy couple of weeks in the country, and definitely in Ontario. One of Canada’s biggest players in the AI space, Cohere, has partnered with both UofT and UW to build out curriculum based on adoption and integrating AI systems into scaling or scaled organizations. It echoes the sentiment from leaders in this space such as Daniel Wigdor (Professor at UofT and founder of AXL) that Canadians lead in the area of Applied Computing and we should be applying that expertise into AI.

Canada has continued to see investment at the Federal and Provincial level. The Deputy Prime Minister announced that they would invest up to $240M into Cohere to scale its AI compute capacity (via CoreWeave, see below). The Federal Government also announced that FedDev would fund $40M into various Ontario organizations to support their growth, with some of those directly mentioning automation/AI systems.

Worth noting re: Cohere is that the $240M is noted as a “partially conditionally repayable contribution” administered by ISED, the largest line item being buying compute from CoreWeave, as reported by The Logic. This means most of this money will be going into the US-based company, rather than a Canadian supplier.

At the Provincial level, Ontario announced their Data Centre Playbook with a view on being transparent on what possible vendors and contractors would need to do in order to be considered by the province. They also have published a call for comments on their assessment framework that closes on September 12th, 2026. Meanwhile, Alberta open-sourced The Velocity White Papers to show their process and mission to make their Government “the most AI-enabled public service in North America.” Which is… a metric to measure against.

Canada has also lost a bit of sovereignty with AMD acquiring Toronto’s inference chip-maker start-up Taalas. While it hasn’t closed, and the release stresses retaining and growing Canadian talent, it is another company that will ultimately have a parent interest elsewhere. KW has also lost Convictional, an e-commerce business that did a hard pivot to be “Slack for the AI era.” Founder Roger Kirkness posted about their journey, the decision and the lessons and information they hope to deliver back to the community.

All this is compounded with two recent StatCan surveys. The first indicating that 19.2% of businesses used AI in the past year (up from 12.2%) but close to half indicated it is not actually relevant to their business. Paired with AI adoption, while showing a 16.8% labour productivity gain, falls closer to 5% when controlled for complimentary capabilities such as training, cloud infrastructure and more.

Two houses, alike in dignity

Anthropic’s recent announcement that they will be imperceptibly watermarking generated text from their Claude models has caused a stir in the industry and wider AI Discourse. They released a more detailed explanation of how it works and noting it is in response to the EU AI Act, alongside other major providers. This highlights that this is likely to be (or to have already been) implemented by other players that operate in the EU (most of them) and isn’t unique to Claude.

On the complete flip side, Google recently enabled users to remove visible watermarks from Gemini’s generated images, videos and music. While in the same week announcing an open source library for adding content provenance information to generated images, text and more. The project, Credentio, is aligned with the Coalition for Content Provenance and Authenticity to help make authorship more transparent. Google has already published a system for verifying AI-generated images, videos, and audio through Gemini Apps.

There continues to be this push-and-pull for the enablement of more and more content creation, while establishing that transparency around that content’s origins is also crucial. Expect this to continue to volley around and be influenced by policy makers in nations worldwide.

B-Sides - Everything else good enough to post

• Google Meet now has a button to allow for taking notes for in-person meetings. Folks using competitors like Otter will ask “They didn’t have this already?”

• GitHub released a new open framework called Agent Plugins 1.0 that allows you to combine MCP and skills into a single package that can be installed.

• DeepSeek’s new pricing includes Off-Peak hours that perfectly align with North American working hours. High five?

This post was not written with any AI assistance and also directly in the CMS like a real hero. Which means you have to deal with all of the improper grammar and weird run-on sentences. Claude is used to help aggregate and factcheck sources as a bit of rubber-ducking. Written posts will always be marked on the bottom with how/if AI is used.