If you haven’t yet subscribed, join the 11800+ smart, curious HR folks by subscribing here:

Subscribe now

Happy Monday,

FINALLY had time to record another practical video.
And this time, about perhaps the most common question I get when I’m about to talk about AI:

How do I create a PowerPoint presentation in our organization's template?

This is how I do it using the paid version of Copilot.

Let me know if you found it useful!

If you haven’t yet subscribed, join the 11800+ smart, curious HR folks by subscribing here:

Subscribe now

Happy Friday!

At this point, we’ve largely been talking about hiring people and adding agents to our organizations as two separate decisions. We recruit someone, and then we build or buy an agent.

But what happens when the person comes with their own agents?

Suppose you hire me. I bring my experience, my relationships, but also a few agents I’ve already worked with. Those agents may have access to knowledge I’ve collected, understand something about how I work, and help me get things done.

What, exactly, have you hired? And when I leave, what leaves with me?

The more time I spend with personal agents, the more I think about this. Working with them feels increasingly natural. And that makes the boundary between my tools and the organization’s tools harder to draw.

In Winston Brought Friends, I described trying GrokBot, Instinct, and Muse. (I’ve since added Dot and Cue to the list, and I will add AutoPilot as soon as I get access.) When a trip landed in my calendar, all three offered to help with the logistics.

Ben Thompson writes about agents becoming our interface to services in Apps, Agents, and Aggregation, and I largely agree, of course, since I’ve been using these agents heavily the last few weeks.

Where does my agent stop and yours begin?

So what have I done with the agents? Let’s take a concrete example.

Over the last three days, I’ve spent 6+ hours in my car. Driving to different events, driving to work in Gothenburg. I dictate this article now while driving, but that’s not the only thing I’ve done while driving; I’ve also answered all my emails.

Or me or me, my Dot has.

It’s easy, quick, and, for me, time well spent in my car. Now, I work for myself, so this is rather easy for me, but imagine if I were employed somewhere and did this. These bots crave context; it's what fuels them. But useful context might sit in my personal account as well as my work account. Should I be allowed to connect the two? It would benefit my work, but it would also mean my personal and professional life blend in ways that have all kinds of implications.

Why not keep two separate swimlanes? Of course that’s an option, but these agents benefit from context (work) and personality (you). If all there is is context, why bother hiring a person at all? You (hopefully) are the one being creative, answering emails, and bringing your thoughts to benefit the organization; the agent is a helper to get you there. So where will we draw the line? And will it even be possible to draw a line?

Because even if I decided to keep all my personal agents out of my professional life, I think it would be very hard to keep them totally separate. Information will, whether I like it or not, bleed into my agents. When it understands I’m traveling, it will ask where and why. When I ask for restaurant tips for a client meeting, it will piece things together. When I’m stressed and asking for last-minute advice on the meeting with my boss, it will get work context, in some shape or form.

So what happens then? And the above assumes that we, the organization, provide people with alternatives to our personal agents at work. If we don’t, well, a different kind of story. It’s a bit like when the smartphone entered people’s pockets. Suddenly, a ban on surfing Facebook on your work computer meant nothing; you just used your phone instead.

If we don’t make employees' lives easier at work by using agents like the ones mentioned above, we can be pretty sure some people will just use their own, with or without our permission.

So do we need to consider this now? Is this something that will happen “somewhere in a distant future”?

Muse is the most downloaded app in the US.
Dot has over a million users, just a couple of weeks after the launch.

These tools are convenient, and convenience usually wins.
So the time to start informing yourself and pondering (and maybe even creating processes) about what’s allowed, what the benefits are, and what the problems are is now.

If you haven’t yet subscribed, join the 11700+ smart, curious HR folks by subscribing here:

Subscribe now

Happy Wednesday,

A bit of an unusual article this week, because last week I sat down with Jason Averbook to talk about AI on the Now to Next: The Signal podcast.

What we talked about? The easiest way to find out is, of course, to listen to or watch the podcast Now to Next on Spotify or Apple Podcasts. But to give you an idea:

What's happened since the spring conferences, where we both attended Unleash in Vegas and HR Tech in Amsterdam. At the latter, we both spent time on the start-up competition, and the winner, Insygna, was building a system for managing agents inside an organization. I was a big fan of their product then, but as we talk about in the podcast, I never ever thought that we would be at a place a couple of months later where we have all these agents, just available.

Jason even says he has too many agents and might need to fire a few. I’m no better. Muse, Instinct, Grokbot, Dot, Hermes, OpenClaw, and around 40 sub-agents in Grok is maybe a tad too many… I'm not saying everyone has these agents right now, but a lot of us do, and what challenges come up when you bring them to work?

Personal agents stay personal for about two minutes, until you use them at work. Most organizations still won’t let you plug yours in. And when you leave, your agent has learned how the company works. Is that knowledge yours or theirs? We need answers soon, because the agents are already here.

I could go on and on, but I suggest listening to the podcast; I think it was a good conversation that not only cover agents but also covers what will happen to HR and how hard it is to predict the future.

Listen on Now to Next or Apple Podcasts.

If you haven’t yet subscribed, join the 11700+ smart, curious HR folks by subscribing here:

Subscribe now

Happy Tuesday,

Back to normal schedules now, with posts on Tuesdays/Thursdays and, hopefully, guides on Saturdays. Been so-so with these guides, but it’s still my hope and aspiration!

2027 will happen sooner than we think, and if you have a project that needs help from someone with a lot of know-how at the intersection of AI and leadership for 2027.
Do reach out!

What do I do? Well beyond keynotes, here are some of the things I’ve been doing this fall;

- 1:1 CxO Coaching
(We’ve built beautiful KPI dashboards, payroll reminders, integrations with HRIS systems)
- Ambassador programs
(It’s almost like our signature program now. We train YOU to train your people.)
- Leadership programs
(How to lead with AI front and center? How do you combine AI + leadership in a practical yet ethical way?)
- Workshops
(Everything from practical agent building to AI for boards to 90 minutes on Lovable.)

All of the above are grounded in practical, pragmatic ways of doing and leading organizations! Interested? Shoot me an email!

Now, let’s get to it.

Winston Brought Friends

Around New Year, I started building my own OpenClaw agent. Some of you met him back in February. Winston, my AI employee who never clocks out.

Back then, OpenClaw was all anyone on tech X could talk about. But you needed a command line and a pretty high tolerance for risk to run it. So I told you not to build one yourself, and I wrote that I was completely convinced we’d have this type of agent, secure and reliable, within a year.

Seven months later, half of that prediction has come true. The agents are here, as consumer products, for regular people.

Secure and reliable? We’ll get to that.

Meet the friends

I ended that post by saying Winston was bringing friends. They’re here.

Elon Musk’s SpaceXAI launched Grok Bot in August. It gives you a persistent agent with its own cloud computer that signs into your apps and keeps working around the clock.

Instinct doesn’t even have an app. You text it on WhatsApp or iMessage, like you’d text a friend. It’s still invite-only, and it’s reportedly in talks to raise money at a $10 billion valuation.

And then there’s Meta’s Muse. It launched on September 8, overtook ChatGPT as the top free iPhone app in the US within days, and passed 3.4 million downloads in just over two weeks, according to Sensor Tower.

If the rumors are right, OpenAI will show its own always-on agent at DevDay later today. (Everything points to this!)

So, simply put, the world is moving to this kind of agent, just as I predicted earlier this year.

Three agents, side by side

I’ve tried a lot of stuff along the way since Winston, Hermes Agent, among others. But for the last few weeks, I’ve been running GrokBot, Instinct, and Muse in parallel.

Which one is best? I don’t have a strong favorite. They’re more alike than different. They draw the line in different places for what they’ll do without asking, and we all have different thresholds for how much of our lives we let them into. If you limit them, they are not that powerful, but I 100% understand that people might be hesitant to invite them to see mail, calendar, etc.

What they do share is that they’re ridiculously easy to use and (if you connect them to stuff) very, very powerful. You give them a goal, they figure out the rest.

I’m going to Abu Dhabi to work with a client in a couple of weeks. When that trip landed in my calendar, all three started working, to different degrees, on getting me there and how the logistics fit together. And they all suggested booking it for me. Not “here are some options, go book it yourself.” More like “give me your credit card details and leave it with me.” (Yes, all three suggested that.)

I haven’t seen that before, not like this.

The interface matters more than we think. Instinct lives in your WhatsApp, Muse has its own app, but more importantly, as mentioned, none of them sit around waiting for you. They look at your calendar, notice what needs to happen, and suggest what they can do. You get the feeling it’s guiding you, not the other way around. (Once again, dependent on you giving them access to your digital life.)

From my phone to your workplace

Which is exactly why I think this is coming to work.

If you have something in your private life that keeps track of things and suggests what to do next, why would you settle for less at work? We saw the same movement with ChatGPT. People got used to it at home and brought the expectation with them to the office. I think we’ll see the same push from inside organizations with agents.

Here in Europe, Muse isn’t even available yet. But once your people have had an agent in their pocket for a few months, they’ll start asking why they don’t have one at work.

This is where I see the biggest obstacle. In a work context, it’s hard to let go of data and control in a good and manageable way. And these agents want a lot. Your email, your messages, your calendar to start with, but then meetings, HRIS, ERP, and CRMs as well, of course.

The early reports show why that matters. Instinct users have described an email sent without their approval and a successful prompt-injection test through email. (Winston would relate. He once sent a newsletter twice, “to be sure.”)

And it’s not just early bugs. Meta says so itself in its safety write-up for Muse. Prompt injection is still an open problem, and Muse will sometimes make mistakes.

Then there’s the data. Muse uses sanitized versions of your conversations to train Meta’s models by default, unless you find the switch and turn it off. One hands-on review found no way to turn its memory off entirely, and Muse kept nudging the reviewer to connect financial accounts, email and identity information. Instinct users have reported copies of their emails sticking around after they disconnected Google. And Instinct’s terms appoint it as your agent, authorized to enter binding agreements on your behalf.

The rest of the internet isn’t sure it wants them either. Amazon blocked Muse from shopping on Amazon.com on September 20, calling it an unauthorized AI agent. An agent that works through a browser is only as reliable as the sites that let it in.

Now move all of that into a workplace. Your agent commits to something on your behalf. Whose commitment is it? It reads a thread full of employee data, and unless someone flipped the right switch, a version of it may end up in a vendor’s training data. In Europe, that’s GDPR territory. And if a platform you depend on decides your agent is unauthorized, the process stops overnight.

And organizations aren’t exactly ahead of this. According to Deloitte’s 2026 research, only one in five companies has a mature governance model for autonomous AI agents.

So I get the hesitation, yet if the tech is there, that’s usually where people want their org to be. And despite all of this, I can see the benefit.

Imagine a leadership agent

Just think about you being a leader with your own leadership agent. It sits alongside you and proactively supports you. It doesn’t replace you as a leader. But it takes in the context and tells you when something is slipping. You haven’t had a one-on-one with two of your people in three weeks. You promised to get back to someone on Friday, and you didn’t; should I do it based on the context I have?

Like having your own personal assistant at work. I think a lot of people would like and appreciate that.

But as with all these tech tools that we will have at our disposal, that will make our leaders’ lives easier, I think it’s important to set firm guidelines for how and when to use them. To make sure the bots are trained with the organisations best, front and center. They’ll need to be built on a clear structure for what good leadership means in that particular organization.

Do that work, along with a strong data strategy, and I think we'll get better leaders out of this. (And I know that the data piece here is a hard nut to crack!)

What happens to the leader’s job?

Which brings me to the question I think matters most. How does this change what a leader does?

If an agent handles the reminders, the follow-ups and a good chunk of the admin, leaders get time back. The question is what we do with it.

One path is letting more people do more. People get more room to act on their own, and each leader takes on more direct reports. We’re already heading that way. Pave’s data shows managers averaging 4.9 direct reports in March 2026, up from 4.4 at the end of 2023.

The other path is letting managers take on more operational responsibility again. Fewer layers of coordination, more time closer to the problems and closer to the people.

Both are tempting. Both come with a price. When a manager is stretched across twelve people instead of six, coaching, feedback, and development are usually the first things to go. And a manager who goes back into operations has to protect the time for leading, or it disappears into the work.

AI frees up administrative capacity, but the need for judgment, coaching and accountability stays. (But as said last week, it might be looking different) That leadership load should determine how many people someone leads. And if your agent sends that follow-up for you, it’s still your follow-up. (I thin. Instinct might beg to differ.)

So the answer will probably differ from role to role. But it should be a choice we make on purpose. Not something that just happens because the tools showed up.

Early this year, I wrote about two things. Winston and OpenClaw, and whether an AI would be a better boss than yours. Now it feels like everyone is talking about both!

I think they’re about to become the same question.

So if you work in HR or L&D, start with something less shiny than which agent to roll out. Could you describe what good leadership looks like in your organization clearly enough for an agent to support it? And who decides what that agent gets to see?

That’s the leadership conversation we need to have this fall. Winston brought friends. Let’s give them something useful to do!

If you haven’t yet subscribed, join the 11500+ smart, curious HR folks by subscribing here:

Subscribe now

Happy Thursday,

Let’s dive right into it this time.

Jason Averbook’s piece on judgment and AI got me thinking! I highly recommend reading it before reading this piece (and also make sure you subscribe to his substack).

His concern is that we remove the opportunities to practice, then expect people to exercise the judgment those opportunities used to develop. Someone still has to take responsibility when the AI gets something wrong.

I agree with that concern. Our exchange in the comments left me wanting to explore something further, though. As the work changes, how does judgment itself change?

I brought up farming, it’s analogy I’ve been thinking about for quite a while. I live in farm-country in Sweden and of the farmers I know, I wouldn’t call any of them lay, quite the opposite. But imagine dropping a farmer from today into 1926 and saying, “There’s the farm. Get on with it.”

They would recognize the job. They would understand plenty about what needed to happen. But without the equipment, forecasts and other tools they normally use, they would (probably) struggle with parts of it.

I wouldn’t take that as evidence that they were a worse farmer or that they are lazy farmers. They simply learned to farm under different conditions, using different methods. Some of their knowledge would transfer. Some of what they needed in 1926 might be unfamiliar.

Jason’s response was that “the farmer still owns the harvest.” Whatever happens to the tools, the farmer remains responsible for the outcome.

Absolutely. But the methods for taking that responsibility have changed too.

Consider deciding when to harvest. A farmer a century ago might have relied heavily on reading the clouds, noticing the wind and remembering what similar conditions had meant before. Today, a farmer can combine local observations with forecasts and data that earlier generations couldn’t access. The World Meteorological Organization describes how these services support decisions about planting, harvesting and other farm work.

The farmer still has to assess the situation. But using a forecast well requires different knowledge from producing your own prediction by watching the sky.

(And we don’t normally expect the farmer to recreate the weather model before relying on it.)

They need to understand what the forecast means for their decisions, including its uncertainty. They need ways to assess whether it’s useful. Much of the knowledge behind it sits with other people and inside systems they use.

That is already a different way of exercising judgment.

Agriculture has changed alongside a considerable increase in food availability. FAO figures put the global average at around 2,360 calories per person per day in the mid-1960s, rising to more than 3,000 in 2023. Those numbers don’t establish which older skills we can safely lose. But they make it difficult to argue that knowing how to work with fewer tools is, by itself, the measure of a capable farmer.

I think we need to leave room for a similar change in our own professions.

We can remain responsible for a result while using different methods to produce it and assess it. What counts as knowing the craft can (and probably will) change quite substantially along the way.

When we talk about AI and skill erosion, I sometimes struggle to tell which kind of change we’re worried about.

Someone might be losing a capability they still need. They might also be spending less time on something that has become less useful, while developing another capability. And, of course, they might be losing the old skill without learning anything to replace it.

We need to be able to tell those situations apart.

The IBM study Jason discusses surveyed 1,500 HR and workforce leaders and 8,800 employees. It found widespread concern about skill erosion and argues for protecting opportunities to develop human capabilities.

That gives us reasons to investigate. But reported concern doesn’t establish what has happened to someone’s ability to do the work. We would need to examine what they can do, where they struggle and how that changes over time.

The same applies to verification.

If hallucinations become much rarer and AI becomes demonstrably more reliable at a particular task, will we still need the same amount of human checking? Will we check the same things in the same way?

I’m not proposing that we stop verifying today. I’m asking how verification might change as the technology improves.

Being responsible for an outcome doesn’t, on its own, tell us how to check it. The consequences of an error matter. So does the ability of a human reviewer to catch it. We may find better ways to test results, monitor performance or identify the cases that need another opinion.

Some of those methods could involve AI too.

That would change what people need to learn. It could also change how they learn it.

We’ve been working on this with a client through a series of leadership labs. Leaders explored the consequences of different decisions, with AI as a natural part of the decision-making process.

The AI reached its conclusions almost immediately. The leaders had to decide how quickly they were prepared to move with it.

When should you speed up? When should you slow down and spend more time working through the reasoning? What could happen if you acted on that decision?

Those were the questions we practiced with them, and continue to practice.

A quick answer creates an opportunity to move faster. It also puts a choice in front of the person responsible. They need to follow the argument, consider the consequences and decide whether they have enough to act. Sometimes that means keeping up with the AI. Sometimes it means pausing a conclusion that arrived in seconds and spending considerably longer examining it.

I think that ability to change pace deserves more attention in how we develop leaders. Moving slowly through every decision could waste much of what AI makes possible. Moving at the speed of every AI response could leave too little room to understand what we’re about to do.

In these labs, we give leaders practice making that choice while AI is part of the work. They remain responsible for the decision, but they’re learning to take that responsibility under different conditions.

That’s the kind of learning I’d like us to explore more: working through decisions with AI, examining what could follow, and practicing when to move quickly and when to take more time.

For those of us in HR and L&D, that means being precise about what we want people to learn. “Critical thinking” is a pretty broad instruction to hand someone designing a development program.

Pick a task people will be responsible for. Identify the decisions they need to make and the mistakes they need to recognize. Then give them different ways to practice, including with AI, and see whether they can explain their choices and handle an unfamiliar case.

I’d be interested in what happens when we test that. Particularly when the person who does the work well uses a method the rest of us never learned.

If you haven’t yet subscribed, join the 11700+ smart, curious HR folks by subscribing here:

Subscribe now

🎧 → Spotify
🎧 → Apple Podcast

Today, I have Sam Pasupalak on the podcast. It’s a short but sweet one (and a highly interesting one, if I might say so myself)

Sam is the co-founder and CEO of Skyfall AI, where his team is working on an AI CEO. In this episode of FullStack HR, we discuss what that would mean and how he separates the operational work from the human relationships involved in leading a company.

The team is still working on the research behind that ambition. Sam explains where he expects progress first and which problems he thinks will take longer.

Here’s some of what we get into:

• Deciding what to work on. Sam sees today’s AI tools as useful for execution once a task is defined. He explains why planning across teams and over longer periods is central to what Skyfall AI is trying to build.

• Anticipating the consequences of a decision. We discuss world models, which Sam describes as a way to imagine possible futures before acting. His team wants to apply that approach to businesses, where conditions keep changing.

• Learning while the business changes. People absorb new information, remember what matters, and forget what no longer helps. Sam explains why building that capacity into an AI system remains an open research problem.

• The work he believes stays human. Even with his ambition to automate operational decisions, Sam sees relationships with employees and their motivation as human responsibilities. We discuss what that could mean for managers and CEOs.

• Where automation might arrive first. Sam describes small online SaaS and direct-to-consumer businesses as early candidates. He also says Skyfall AI has acquired businesses to pursue that work, with results still to be announced at the time of our conversation.

Listen to the full conversation with Sam Pasupalak from Skyfall AI on FullStack HR.

If you haven’t yet subscribed, join the 11700+ smart, curious HR folks by subscribing here:

Subscribe now

Happy Friday,

Earlier this year, I analyzed Swedish job ads. I wanted to understand what was happening to the opportunities available to people starting their careers.

Using public data from Arbetsförmedlingen and JobTech, I examined selected occupations that typically require higher education. Between 2022 and 2025, ads with junior-related wording fell by 48.6%. Total ads in the sample fell by 40.5%, so much of the decline happened in an already weakening market.

The larger difference was between occupational groups. Junior-profile ads fell by 66.2% in the group I classified as more exposed to generative AI, compared with 36% in the comparison group.

These were ads containing words such as junior, trainee, or graduate. The measure misses entry-level roles advertised without those words. It also tells us nothing about who got hired or which employers used AI. The economy, interest rates and the tech cycle could all help explain the decline.

But I did show a pattern, and a similar pattern to what we’ve seen happen in the US for example with the “Canaries in the coal-mine” paper that was published last year.

Then I read this new working paper from Danmarks Nationalbank.

The researchers linked Danish firms’ responses about AI use to monthly employment and wage records. They followed companies that first reported using AI in 2023 and compared them with companies reporting no use in 2022 or 2023. They accounted for each firm’s earlier employment trend, its size, and developments within the same industry.

That gets us closer to what happens inside an organization after it starts using AI. It still doesn’t prove that AI caused the difference. Companies choose both their technology and their staffing strategy, and those decisions can be part of the same reorganization.

But by September 2025, employment at the adopting firms was about 11% lower relative to those earlier trends and the comparison group.

Eleven percent sounds like a lot but the authors offer a useful illustration which I really like and that I also see across a lot of the clients that I work with. Imagine a company that grows from 50 employees to 53, while its earlier growth suggested it would reach about 60.

More people work there. The company is still hiring. It has nevertheless fallen below the path it was on. (Estimating this is hard but I do like how they try to do this in the Danish study!)

What also changes the picture is the the size of the company. The decline was concentrated in firms with fewer than 100 full-time-equivalent employees. Their shortfall was roughly 15%. Larger adopters showed no statistically clear decline in total employment relative to their trend, although their employment mix shifted towards less AI-exposed occupations.

What interests me more is how the adjustment seems to happen.

The researchers found a sharp decline in employees with short tenure relative to trend, while longer-serving employees were retained to a greater extent. They interpret this as adjustment through reduced hiring rather than layoffs.

From an HR perspective, it is easy to imagine how this could play out. A manager decides the team can handle the workload without filling a vacancy. A planned position gets postponed. The next budget includes fewer additional hires because the existing team is expected to do more with AI.

Just as I found in my study, in Denmark they found that the shortfall was concentrated in occupations with tasks more exposed to AI. Young and university-educated workers also showed larger declines, although these groups overlap. The researchers cannot neatly separate age, education, and occupational exposure into independent explanations.

But if this holds true and if we hire fewer junior people because experienced employees can do more with AI, how do people get the experience we will later expect them to have?

That is my concern and worry, rather than a finding from the paper, but I think it’s a valide one. The researchers haven’t followed today’s missing hires through their future careers. But the question belongs in workforce planning now, alongside the savings from the positions we decide not to fill.

I’m also torn here. Keeping a task manual just so someone can learn it seems like an odd response to better technology. People should learn to work with the tools they will use. At the same time, handing someone an AI tool doesn’t automatically teach them to recognize when its output is wrong. We still have to design work in which people can develop that judgment.

There is some reassurance in the Danish results. The researchers find no clear wage response, and the measured effects among adopters are too small to show up clearly in employment across the economy. Denmark also has a fluid labour market, which may help people find opportunities elsewhere.

I don’t think we should assume every country or company will handle the adjustment equally well. The study covers early adopters over a limited period, and the firm survey excludes businesses with fewer than ten employees. It leaves plenty unresolved.

If you are CxO or in HR, I would start with a fairly ordinary exercise. Look at the positions you filled, the vacancies you chose not to replace, and the planned hires that never reached an advert. Ask the managers involved what changed. Lower demand, budget pressure, a different organization, or an expectation that AI would absorb the work?

Then look at which opportunities have disappeared for people at the beginning of their careers, and what opportunities we have created for them to learn instead.

Worth while thinking about when going into workforce planning for 2027 which I think a lot of us are about to enter now.

Sources: Johannes Sundlo, Vad händer med de juniora jobben, updated August 2026; Simone Maria Bonin, Saman Darougheh and Andreas Kuchler, AI Adoption and Firm Size: Employment Effects in Monthly Administrative Data, Danmarks Nationalbank Working Paper No. 223, 16 September 2026.

If you haven’t yet subscribed, join the 11700+ smart, curious HR folks by subscribing here:

Subscribe now

Happy Wednesday,

Last week’s big debate in the AI circles was this tweet (is it still called a tweet?) from Jacob Coxon. Discussions have ranged from the idea that this is solely hype from Anthropic (pre-IPO) to the idea that he is an activist doing this.

I don’t know the exact truth behind this, but what matters to me is that people I’ve followed for a long time and respect didn’t dismiss him outright. Like Ethan Perez (Alignment team lead at Anthropic)

And yes, we can argue that this comes at a time when Anthropic seems to be falling slightly behind OpenAI, is preparing to go public, and might want to make it harder for competitors to enter the space. Or they sincerely want to regulate it.

On top of this, Dario Amodei published a long article (worth the read) titled Pacing the frontier, where he argues that we need to slow things down and put regulations in place. This has also sparked debate about why they keep building it if they see these risks.

I like parts of Ben Thompson’s perspective (I usually do). He wants AI policy to focus on concrete risks without assuming uncontrollable superintelligence is inevitable. He argues that society should preserve human freedom and continue developing AI while addressing problems as they arise.

I find this pragmatic and grounded. Let’s not over-regulate and get caught up in the labs' hype cycles too much.

But. Even if we decide to slow further development, organizations still have to deal with what these systems can already do. What we have right now is good enough to make a real economic impact once organizations start scooping up compute.

As an example, last week I needed to change DNS settings for whatever reason. I know how to change DNS settings; it’s not that hard, but it requires a lot of clicks. So I figured I’d let Astra do it through computer use.

And it did, in roughly the same amount of time it would have taken me. I could do other stuff instead, and my part was to hand it over and check the result.

Across a team, handing over recurring tasks like that could mean taking on more work with the same people. How much you gain depends on what the tools cost, how reliably they work, and how much time you spend checking the result. But if those numbers add up, you have more room to decide how to run the business.

A couple of days ago, I spoke to an organization that had barely started with AI. They were stuck and didn’t know how to move forward, their CEO doesn’t yet see value etc etc. The usual objections. Interestingly, I’ve worked a lot with an organization in the same industry. They are investing heavily in AI, not only by having me train their lead team and managers, but also in raw compute. I would put them at one of the organizations in Europe that have invested the most in AI models outside the tech companies. And that is starting to pay off; we're already seeing better margins as an early effect.

That would give them choices the other organization doesn’t yet have. They could keep more of what they earn, invest more, or compete on price. If they can deliver comparable work at a lower cost, they can put pressure on competitors long before anyone announces an AI-driven restructuring.

That pressure could eventually affect jobs through fairly ordinary decisions. The next hire might no longer be necessary. Someone leaves and isn’t replaced. One company wins more business and hires people while another loses work. You won’t necessarily see an instant dramatic wave of AI-related layoffs, even as companies change how many people they need.

Organizations are, as I wrote last week, slow and messy. I’m not seeing widespread AI-driven job losses among my clients today, as I said, including those furthest ahead. That friction makes the timing hard to judge, but I struggle with treating it as reassurance about what comes next.

I would still be extremely surprised if we don’t see job losses grow in the next 10–18 months. My bet is that we’ll see this happen in the US first, since we have more regulations to abide by in the EU (for better or worse). I could be wrong about the pace, and some uses may turn out to cost too much or need too much supervision to make economic sense. Those are useful objections to work through.

But what frustrates me is hearing influential voices use the lack of widespread job losses as a reason to relax, especially when they spend very little time using these tools themselves. If you’re advising an entire profession or organizations on the limits of AI, you should probably have some direct understanding of where those limits are today.

(I’m trying, very, very, very, very hard usually not to dunk on other people and instead show what good looks like, but it’s getting harder and harder.)

And we need that understanding inside our own organizations too. If your CEO doesn’t see what you’re seeing, give them something they can judge for themselves. Take a piece of real work. Show what the system can do, what still requires a person, where it fails, and how much effort it takes to verify the result.

Then work through what would change if a whole team started working that way. What would it be worth? What would it cost? Would it change the work you take on or the people you need to hire? You can start having that conversation even if you disagree about my timeline.

If we wait for widespread job losses before getting involved, decisions about how to reorganize the work, which roles are needed, and whether that next hire happens may already have been made without us.

Welcome to FullStack HR, and an extra welcome to the 36 people who have signed up since last edition.

If you haven’t yet subscribed, join the 11600+ smart, curious HR folks by subscribing here:

Subscribe now

🎧 Listen to the episode 🎧

Spotify
Apple Podcast

Candidates use AI to write their applications. Companies use AI to filter those applications. And somehow, both sides end up unhappy.

Greenhouse calls it the “doom loop.” Having recently gone through a hiring process myself, I can relate.

Today, Sharawn Tipton, Chief People Officer at Greenhouse, joins me on the podcast to talk about what we can do about it.

Here is some of what we get into:

• Why the CV can miss people who could be great at the job

• What an AI interview feels like, and what happens if a candidate opts out

• How Greenhouse involved its employee resource groups in testing voice AI

• Why every hiring decision still needs someone who owns it

• How Sharawn and her team use AI in their own HR work

We also talk about something I suspect many of us recognise. It is now very easy to produce more work. Whether that work is useful is another question.

Sharawn’s grandmother had a saying: “Anyone can have a recipe, but not everyone can cook.”

Turns out that applies pretty well to AI, too.

Listen to the episode for the full conversation, including what Sharawn thinks HR leaders should stop doing on Monday.

If you haven’t yet subscribed, join the 11600+ smart, curious HR folks by subscribing here:

Subscribe now

If you’ve been following FullStack HR for a while, you know I’ve been writing about how AI will impact HR since January 2023.

I’ve been both positive and negative. And while it’s probably fair to put me more on the “AI hype” end of the spectrum, this was one of my conclusions in that first article about ChatGPT:

Since then, I’ve relentlessly tried to put the pieces together: where is this going, and what does it mean for HR and the work we do?

I’ve done so by talking to people, both at the labs and in other organizations. I listen to so many podcasts and YouTube videos a week that I sometimes wonder if it’s healthy. On my bike, running, commuting, cooking. (Always at 2x speed, just to get more information in.) I read the news every morning, every night, and usually in between.

And then, perhaps most importantly, I test the models deeply and thoroughly. Ask my wife, and she’ll tell you about the late nights.

Along the way, I’ve encouraged you to play around with the technology yourself. To use it, understand it, and form your own opinion.

Having worked with GPT-6 Astra since launch, mine has become considerably firmer.

I’m now convinced that, from a technical perspective, AI can replace a substantial share of the work we do in front of a computer. In my own work, I’m seeing it complete tasks at an equal or better quality than I would expect from HRBPs on my team.

I thought it would take us a bit longer to get here.

My bad.

The computer use capabilities are mind-blowing. I remember being blown away when this launched. Fast-forward to today, and Astra can do (almost) whatever I do on a computer, fast.

That’s my experience with the work I’ve given it so far. It’s also why I’ve become so convinced that we’ve crossed the bridge where we, on one side, hypothesize about what AI could do and, on the other, conclude what I can do. We are now on the can-do-side.

As an example, we’ve been talking about MCP and connectors here, and sure, they’ll continue to make your life easier. But if your HR system lacks one, the AI may now be able to use the same interface you do and act on your behalf.

(Yes, simplification, I know. Access, permissions, and getting the result right still matter. But technically, once again… the capability is there.)

So if we are on the can-do side, what should we expect? Just because we are on that side doesn't mean we'll be replaced from one day to the next. Other factors affect how an AI operates. Governance, data structure, laws. Things we’ve put in place to safeguard our organizations, and things we still need to take seriously.

But when an AI gets stuck, we need to understand why. It might lack the capability. It might lack the information. Or it might be perfectly capable of taking the next step but lack permission to do so.

Those are different problems.

My guess is that more of the limitations we’ll encounter will come from everything around the model. That doesn’t make them small problems, but it does change what we need to work on.

Astra isn’t perfect. Neither are humans. And as said, technical capability alone won’t mean mass layoffs overnight. Organizations still need to figure out how to use it, how to check the work, and who takes responsibility for the result.

But if I’ve been a bit on the fence about when and how this will impact us in HR, I’m more convinced than ever that it will not only impact our work; it will change how we work, or whether we work at all.

We’ve been marching towards this for a while. I’ve spent years writing about it and trying to understand it. Even so, working with Astra has made me reconsider the timing. What I wrote about in January 2023 is materializing faster than I expected.

The pace is not slowing down, and I’ll keep trying to figure out what that means for HR and how we work until there is nothing left to figure out.

We’ll see when that happens.

And to the skeptics out there: if GPT-6 on Ultra has failed you, please let me know in the comments. I would love to hear where it’s falling short, because I’ve struggled to make it fail at the work I’ve given it.

Tell me what you tried, where it got stuck, and what you had to fix.

Welcome to FullStack HR, and an extra welcome to the 12 people who have signed up since last week.

If you haven’t yet subscribed, join the 11600+ smart, curious HR folks by subscribing here:

Subscribe now

Happy Wednesday,

I feel fully back from summer now, slowly but steadily moving back into the daily routines of working with organizations and leaders almost daily.

It’s great to be back meeting teams and leaders; it’s where I form and hone most of my thinking, and I’m going to go against the jante law and say that I’m getting pretty good at working with organizations on AI and leadership.

The convergence of having spent my life in HR with great mentors and colleagues, paired with working in the tech industry, and being insanely passionate about HR Tech and how it will impact work since 2010, has led to this, and I feel happy and fortunate to be here.

If you need help rethinking your leadership programs around AI, reach out! And the offer from summer still stands! If you need a second opinion on HR Tech, leadership programs, or anything else regarding AI, I’d be happy to help for free. (3 people have done so so far!)

That said. We need to talk about skills again, so let’s get to it.

And yes, I know we have talked about skills before, but we need to do it again since skills are evolving and are in almost all AI tools today.

ChatGPT has them.

Claude has them.

Copilot Studio (and Copilot Cowork) has them.

Skills are how organizations extract value from AI tools right now.
(At least if you want to extract value beyond the individual use cases.)

But if you look under the hood, what is a skill then? At it’s core, a skill is a set of text instructions that tells a model how to do a certain thing. That’s the essence of it! A text file.

How you create one differs a bit between the different tools, but the substance is identical everywhere. Written instructions, saved somewhere the model can find them, reused every time that task comes around again.

That’s it. (No secret witchcraft.)

Two ways to create one

So how do you create one? I see two ways. (There are probably more ways, but this is how I do it.)

The first way is to let the model analyze what you already do.

If you have a process document of some sort, where you’ve clearly defined the steps that process takes, you already have a skill sitting there in draft form. Feed it in and ask for a skill. Most organizations have far more of these documents than they think, quietly rotting in a SharePoint folder somebody created in 2019.

The second way is to just do the work you want done.

You do the work once, together with the model, and let it capture the workflow while you go. You correct as you move. You say what went wrong, you explain what you want different, you get picky about the format. Then you ask it to package the whole thing into a skill.

I prefer this method, but we are all different, so use whichever one you can start with today.

A concrete example: performance reviews.

Let’s take an example, because this sometimes gets abstract fast.

Say you’re about to run performance reviews in your organization. The obvious move is to build a skill for the review itself, and sure, you can. But the more useful move sits one step earlier.

Build a feedback-gathering skill.

A skill that asks the manager a set of very specific questions about each of the people reporting to them. Not “how has it gone”, which produces nothing. The questions you’d ask if you were sitting in the room with them. What did this person deliver over the past six months? Where did they surprise you? Where did you have to step in? What have you been avoiding telling them? What would you write if you knew they’d read it tomorrow?

The skill asks. The manager answers. The skill structures the result.

Then you send that along with your performance review guidelines.
”Dear manager, you’re about to head into performance reviews. Use this skill to extract what you already know about your people before you write a single word.”

Two things already exist in your organization: your questions and your guidelines, and this skill makes them run and be repeatable. Sure, you could build an agent that did this; sure, you could buy a tool that did this, but you can also just create a skill, and it will take you what, 30 minutes?

If you want to try this skill, you can download it here and test it out.
(You could also adapt it to your own process; just ask Claude or ChatGPT to adapt it to your process.)

How to upload skills

Here’s how.

Claude

ChatGPT

And this is the hard way. You can also just tell Claude / ChatGPT to learn the skill and point it to the file.

And you can do this for whatever you want! (Create a skill that is.)
This isn’t a special category of task reserved for clever people. Create a skill for whatever recurring thing you have or where you want to be specific about something.

And yes, you could potentially connect this to your performance review system or HRIS.

Because for a skill to work at its most effective, it needs access to tools and systems.

A skill that writes a manager briefing from nothing is a writing exercise. A skill that writes a manager briefing from the real engagement data, the actual team composition, and the actual notes from last quarter is a different animal entirely.

That’s what connectors and MCP are for, and I’ve written about that separately, so I won’t repeat it here. Connectors and MCP, explained for people who have never touched either.

The chicken and the egg

Which brings us to the question: What do you do first? Do you create skills, or do you connect stuff to your AI system?

Ideally, you connect first. A skill with access is worth considerably more than a skill without it.

But a connection on its own doesn’t do anything either. A connection also needs instructions. Nobody’s HR function got better because someone plugged in a system. It gets better when something knows what to do with what’s now reachable.

So the answer is that you’ll do both, and you’ll start with whichever one you can get your hands on this week. If your systems are locked down and IT hasn’t come back to you yet, start with skills. Find workarounds to the information you need. If you already have connectors sitting there unused, start writing instructions for them. Neither order is wrong, but standing still and doing nothing is.

And as said, it’s very easy to create skills. (Almost too easy.)
You describe the work, you correct it a few times, you ask for a skill, and you have one.

The perhaps hardest part is knowing what to make a skill out of. But I suggest you start by picking the thing you’ve explained to an AI more than three times in the last couple of days.

That’s the whole selection criteria. If you’ve written “remember to...” or “always structure it like this...” or “you missed the same thing again, my friend” to a model more than a couple of times, you’ve found your first skill.

Do the work once. Let it capture how you did it. Save it.

Then go connect something to it.

And oh, with clients and friends, I’ve created 16 others skills for you.

Free to use, no sign-up required.
BUT if you do like them and feel that this has been useful, as “payment” I would love if you shared this post on LinkedIn.

Share

Download the skills here.

(For how to install them, scroll up!)

If you haven’t yet subscribed, join the 11600+ smart, curious HR folks by subscribing here:

Subscribe now

Andy Doyle has a title that I’ve never seen elsewhere: Chief People and Agent Officer. Today he joins me on the podcast to talk about the title and the numbers behind the title. Kantar has around 12,000 employees and around 15,000 agents (!), so the agents already outnumber the people, and that gap widens every week. And who built them? Well, most of them were built by colleagues who wanted their own workday to be easier, not by the tech or IT department.

In January 2025, Kantar went from 300 Copilot licenses in pilot to 2,000, and three months later everybody had one. They made that bet before they knew how the productivity would pay for the licenses. (Spoiler from the episode: I really like this approach.)

But I’m getting ahead of myself; here is some of what we get into.

• Why they run AI training every Tuesday

• Change influencers, experimentation, and permission to fail

• Kantar’s Agent Factory and the path from no-code to pro-code

• The roughly 70 employees outside IT who have been trained to build agents

• Examples from HR data, HR service delivery and recruitment

• What happens when an agent starts being shared and needs stronger security, controls, and ownership

Andy also talks about the 30 (!) agents he has working for him personally. He also has one sentence for everyone who hasn't started yet, which I will let him deliver himself…

Listen to the episode here

🎧 Spotify →

🎧 Apple Podcasts →

Hit follow while you are there!

Connect with Andy on LinkedIn → https://www.linkedin.com/in/ajdoyle/

And are YOU doing interesting stuff with AI and want to be on the podcast? Reach out! johanness@sundlo.com

If you haven’t yet subscribed, join the 11600+ smart, curious HR folks by subscribing here:

Subscribe now

Happy Tuesday,

This is the written version of today’s podcast episode. If you’d rather listen, it’s on Spotify and Apple Podcasts.

Next week... I'm closing out the Startup People Summit hosted by none other than Matt McFarlane.

The program is 🤯, and I'll wake up a bit (too) early so I don't miss anything.

My talk will center around that you can just...do things now. Build things.
And how you do them. And why you need to do them.
And why you shouldn't do pilots.

See you there? —>

And on Friday this week, I’ll do an update on the latest in the world of AI from a leadership perspective. What do you need to know? What’s the latest updates?
100% free and you only need to be subscribed to this very newsletter.
(Which I hope you are!)

In May, I wrote that AI is a work tool, not a perk. The argument was simple. If you buy Copilot, give everyone the full license, because the alternative is building an A-team and a B-team inside your organization.

Three months and a lot of client work later, I want to come back to it. Not because I’ve changed my mind, but because I expected (hoped?) to see the gap start to close by now. I don’t see that. What I see is organizations fuelling it.

Nobody is doing this out of malice

I don’t think this is done out of “evil”; there is no bad intent anywhere in the chain. Nobody sat in a meeting and decided who should be allowed to pull ahead. (I sincerely hope this at least) Instead, it’s a budget line, a pilot that never ended, a license count that made sense in October 2025. Good people, sensible decisions, and an A-team and a B-team at the end of it.

And I get it. We’ve always had fast adopters. People who are extra curious, extra invested, who want to learn new things, whether that’s to have an easier day at work or to stay ahead. That part isn’t new and isn’t a problem.

The problem is what the organization does with that difference, and right now, most of them are widening it rather than closing it.

Enter almost any organization, and you can sometimes visibly see or at least feel both groups from the same spot. One group is far ahead; they’re running fast, they have superpowers, and they either pay for their own Claude and ChatGPT accounts or they’ve nagged their employer into giving them a proper license.

The other group has barely started, or they’ve tried it once or twice but haven't found lasting results for various reasons.

Same company, same floor, sometimes the same team, yet so different results or thoughts.

So how do we end up here?

Someone knocks on the door and asks for a business license. They get it, because they asked and because they clearly know what they want it for. The person who didn’t knock doesn’t get one. Nobody made a decision about who pulls ahead. The distribution follows who was comfortable asking.

If you call yourself an equal opportunity employer, that has to include the tools. Not only for the ones who are loud and vocal about wanting a paid license. You roll it out universally, or you’re letting confidence decide who gets to be productive.

When I look at the organizations where this has gone sideways, the common denominator isn’t budget. It’s that nobody defined why this matters.

Why do we care about AI? Why is this an investment we need to make? Because it is an investment, a conscious one, and it has two halves. The tooling is the easy half. The other half is helping every person answer the question “what does this mean for me and my job on Monday morning?”

Without that second half, you’ve bought licenses and called it a transformation. (yey!)

You can’t be strategic if you can’t be practical

And before we talk about this, please keep my sample in mind. My sample is organizations that want to work with AI in this shape and form. The organizations I work with lean forward by definition, so don’t draw too many conclusions from what I see.

But what I will say is that the ones doing this well share one trait. Leadership has a working understanding of what AI can and cannot do. Not a vendor deck; hands-on with the tools. (If you’ve been reading this newsletter since the dawn of time, this is hardly a surprise, but I will keep repeating this until we have all lead teams onboard. Current estimate is May 2057.)

But yes, that’s my firm belief after many (many) lead team sessions this year. If you can’t be practical, you can’t be strategic. Without a basic, rudimentary grasp of how the thing works, it’s close to impossible to get any of this moving.

And the pace right now is punishing. It’s been fast since ChatGPT launched, but the last six months are something else. Every month you leave the B-team where it is, the distance grows.

The fix hasn’t changed since May. Give everyone the full tool, train everyone, and tell people why. What has changed is how much catching up your B-team has to do once you finally decide to.

If you haven’t yet subscribed, join the 11500+ smart, curious HR folks by subscribing here:

Subscribe now

Happy Thursday,

On Tuesday, I promised you real cases from people who are experimenting and driving AI forward in their companies. Not “AI this and AI that”. People who build, measure, and kill what doesn’t work.

Today you get the first one:

Joe Shahmoradian is Director of Talent Acquisition for EMEA at Morningstar. 15 years in TA, a journalism degree he never used for journalism, and one of the most hands-on AI builders I’ve met in this field.
Everything we talk about in this episode is stuff his team runs today.

Some highlights to give you a taste!

His team cut post-call admin by 75% with AI transcription on screening calls. The first tool he tried got killed because it treated small talk and candidate answers with the same diligence. How he found the right one, and why he’s now pointing the same model at interview quality and fairness, is in the episode.

He built Amira, an AI teammate that works like a journalist and scans the market around the clock. Market intelligence that used to take five days now goes out as short video briefings in about 90 minutes, before stakeholders even ask.

Joe built an AI clone of himself before paternity leave, with access to his emails, chats, and every document he’d ever made. Lying awake one night, he started wondering what would happen if someone maliciously prompted it to send an email in his name. So he got up and tested it…

We also get into how Morningstar moved AI adoption from enthusiastic volunteers to structure, with an AI Academy, a champions community that grew from five people to nearly 200, and an idea hub with ownership.

Plus Joe’s answer to what he’d do in his first 90 days at a company starting from zero. Spoiler, it starts with governance and it skips the demo-agent theater.

Listen to the episode here

🎧 Spotify →

🎧 Apple Podcasts →

Hit follow while you’re there. The next guests are already recorded, and the bar is the same for all of them. People who do stuff.

Connect with Joe on LinkedIn, he is always happy to talk about this.

And are YOU doing interesting stuff with AI and want to be on the podcast?
Reach out!

Welcome to FullStack HR, and an extra welcome to the 72 people who have signed up since last week.

If you haven’t yet subscribed, join the 11600+ smart, curious leaders by subscribing here:

Subscribe now

Happy Tuesday,

Before we dive into the episode, I’ve been looking at your survey results on what you want FullStack HR to be this fall - thanks to all who filled out the form!

First, it’s clear that most of you want practical, tangible stuff. AI tools and prompts that you can use directly- I will do my utmost to provide that, keeping the same cadence as before summer and sending those practical updates out on Saturdays.

Secondly, you also want news and trends. To accommodate that, what I will trial is a monthly update here on FullStack HR, live. I will do three and then evaluate, turn out, etc.
The first one will be next Friday at 3 PM CET / 9 AM EST. You will need to be a subscriber to watch it, and I will send out the invite shortly.

Thirdly, real cases from other companies. This is in the making, and I’ve recorded two podcast episodes with, in my view, insanely interesting podcast guests who do stuff. Not talking about “AI this and I that”; these are people who are experimenting and driving AI forward in their respective companies.
I will post the first episodes in the newsletter, but make sure you hit the follow button either on Spotify or Apple Podcasts - I can guarantee that you don’t want to miss the episodes!

Are YOU doing interesting stuff with AI and want to be on the podcast? Reach out!

Ambitions are high for the fall, and I hope you’ll tag along!

All that said, let’s get to it.

I’ve been slow for a couple of weeks. I wrote all summer and needed two weeks off to gather my thoughts. Instead, I discussed…this topic. A lot of conversations, most of them circling the same thing. Human writing versus AI slop.

The debate got new fuel on August 14, when Anthropic started watermarking everything Claude writes. It’s an invisible signature built on Google DeepMind’s SynthID approach, where the model’s word choices get nudged into a pattern a detector can read later. Anthropic rolled it out globally to meet the EU AI Act. Some people say they’ve already reverse-engineered it and can strip it out.
I’m not so sure. It also isn’t the point.

Substack shipped its version of the same idea in July. Any reader can hit the three dots on a post and see what percentage Pangram thinks is AI-assisted. I have my objections to Pangram, but they seem to do a decent job. (With emphasis on decent.)

So we have watermarks on one side, detectors on the other, and a long argument about transparency in between. My question is simpler. Why do we care?

And I’ve landed on the idea that I don’t care whether a human or a machine produced the slop. I care that it’s slop.

I use AI constantly. To gather my thoughts, to get new ones, to summarize what I already think, to challenge it. That’s where the value sits for me.

I have also let AI write entire pieces on its own. Those tanked. You could tell immediately. Not thought through. But I’ve written pieces with no AI anywhere near them that got exactly the same response. Rushed. Boring. Is this even true? Same verdict, no machine involved.

The problem was never the tool. It was me not thinking hard enough.

Which brings me to how I made this piece. I dictate in the car, then I let Claude or ChatGPT turn the dictation into an article. I’ve worked this way for years. The thoughts are mine, the detours are mine, the chain of thought you can hear on the podcast version is mine. I have no idea what Pangram makes of this particular text, and I don’t care. Sure, it’s sometimes a bit off, and I’m guessing that it will flag this as partly AI-generated, which it is. I mean, AI wrote this, but the thoughts and the rambles are mine. (As you can hear in the podcast version.)

It is what it is, but the part of the Substack feature I like is the box next to the score, “How I make this,” where you write your own explanation of how you use AI. That turns a verdict into something you can learn from, and there’s plenty to learn from how other people work.

What puzzles me is the off switch. You can disable scanning on your posts, and readers get “AI detection unavailable” instead of a number. I see creators I’m fairly sure use AI turning it off. For me, that reads as shame or uncertainty.
Say it instead! I use AI here; here is where, here is how.
That’s a better post than the one you’re protecting.

So the summary for me is, don’t be afraid to use AI in your writing.
I’d prefer you did, if it makes the piece better. Get good at prompting. Get good at making your point land. Own the result, and be open about how you got there.

If what comes out is sloppy, that’s not the AI.
That’s you not having thought it through.

If you haven’t yet subscribed, join the 11,600+ smart, curious HR folks by subscribing here:

Subscribe now

Curious to learn more about Claude Cowork and ChatGPT Work?
I run the third cohort of the popular Cowork for HR starting on the 27th of August.
Three sessions, and the first 10 to sign up get a 30-minute private coaching session with me.

Join now

For the last couple of years, as a kind of sanity check, my co-worker Sara and I have been reviewing HR ads.
First, we did it manually. Row by row. (For Sweden only, but still.) Looking for whether anyone happened to drop in at least one mention of AI. No one did in the beginning, and that was maybe not surprising at all- new technology, uncertainty about whether this was hype or something more long-term sustainable.

But last year when we did the Swedish count again, there were still…almost none.

During the summer, I did the Swedish count again. (When I say I, in this case I mean Claude + the excellent open API provided by the national Swedish Public Employment Service)
There were still… almost none.

But what if we widen the lens? Looking beyond Sweden, what happens then?
In how many job ads will AI pop up? I started to count (but again, when I say I, I mean Claude + Monid), and all in all, we’ve analyzed 3,350 job ads in HR.

Before we dive into the data, it’s perhaps good to answer the question of why this matters. Is it at all interesting to look at? Unsurprisingly, my answer to this is yes, for several reasons.

First, I believe this tells us something about the true mindset of organizations right now. My assumption is that if AI is important and/or adopted in an organization, that will be reflected in how they hire. I know this is a somewhat flawed assumption; it could be that AI is important even if it’s not in a job ad or that AI skills are just seen as a commodity in the same way as we expect people to work with email these days.

Yet, I don’t think we’re there.

I think this is one of the best indications to see where we are with the use of AI in HR.

(And yes, I know that putting it in a job ad is easy and that it could be merely words, but I still think it shows some kind of ambition.)

Secondly, research on “learning-by-hiring” shows that recruiting people with relevant expertise can be an important way to introduce and spread knowledge1 within an organization, particularly when new hires collaborate closely with existing employees. It is not the only route to upskilling, but it can be a key component. So, are organizations making enough use of it?

But the data, then?

As mentioned, I’ve analyzed 3,350 HR job ads across the US and 20 European countries. It’s a seven-day snapshot from late July and early August, so think of it as a signal rather than the complete picture. But what does that signal tell us?

5.4% of them require AI as a hard requirement.

And an additional 4.1% ask for it as a skill that is desired, bringing the total to where an AI skill is asked for to 9.6%.

If we look at the total mention of AI, we see this in 18.5% of all job ads. That could be mentions such as “we work with AI” or that the organization is selling an AI-enabled product.

I’ve also analyzed the different roles within HR.
L&D is the leader here. The difference is not vast, but it’s there.

If we widen it to mentions, we see that 18.5% of all job ads mention AI, but if we dig a bit deeper, we see that a lot of these have little to do with the job and are describing the employer. Not ideal but not super bad either; I’ll take that over no mention at least.

Ok, but those who mention that AI skills are a good thing to have - what kind of skills are they looking for?

You’ll see the results below, but I do want to point out a couple of things.

1. It’s interesting that HR Leadership is so low in analytics and insights with AI. Why? Isn’t this a core feature for HR Leadership, or are we putting the HRBP:s to the test here?

2. Same goes for responsible and lawful AI; my expectation would be that this would be super-high for HR Leadership. (Maybe that was more of a hope?)

3. HR Generalist? Congrats, you're building agents now!

4. It’s very clear that we’re looking for people who can build stuff inside the HR core workflow and not machine learning experts.
   

Tooling then? When AI is asked for, what kind of tools do they prefer you use when working? Claude!
Before I get too excited about seeing Claude on top here, it’s worth noting that this dataset is super small. But still, Claude is on top.

Also a big shout-out to one of my very first sponsors, Metaview, for making it into three job ads!

If you want to see the full dataset, head over here. You’ll see whether Europe or the US is in the lead, and what that looks like, country by country.
You can also download the data, upload it into whatever LLM you prefer, and see the methodology and caveats behind this.

What to make out of all of this?

It’s very tempting to scream that this is disastrous and that we need to do better, but that would not be a great analysis. There could and probably is more to this than first meets the eye, and I’ve laid out a couple of these ifs and buts above.

So what is it then? It’s this.

Not hiring people with AI skills won’t be the end of your organization. After all, it’s still a new technology, a new skill. But it’s a missed opportunity.

I understand that vetting such skills is hard if you don’t have them yourself. So maybe don’t make them a requirement, but at least signal in the ad that these are skills you value. That you and your team are curious about.
(After writing the intial article, I got nudge by Martyn Redstone about Sova Assessments AI Readiness - worth checking out to solve this problem!)

Of course, you need to match that job ad with reality. If you want curious AI people on your team, you have to let them be curious AI people once they start.

It could also be that we already treat these skills as something internalized—something we assume people have without needing to vet or make explicit. Think of the email example above. No one asks candidates today whether they can send an email or use the internet. We assume they have those skills.

I would love for that to be the case, but I don’t think that’s what we see here.

My hypothesis is that we still don’t see this as a core competency in HR, and I worry about that. (I’ve written about this several times this summer.)

As a candidate, I would be worried if AI wasn’t mentioned anywhere in the recruitment process. (And preferably not as “we don’t allow it here”.)

The job ad is a great way to signal to candidates and the rest of the organization that this matters to you.

It also helps raise the bar for your current team, new hires, and yourself.
Use it.

If you haven’t yet subscribed, join the 11600+ smart, curious HR folks by subscribing here:

Subscribe now

Curious to learn more about Claude Cowork and ChatGPT Work?
I run the third cohort of the popular Cowork for HR starting on the 27th of August.
Three sessions, and the first 10 to sign up get a 30-minute private coaching session with me.

Join now

Happy Tuesday,

I know it’s the middle of summer…but autumn is getting close, and what better than a few of my predictions for what’s to come in the upcoming months?

1. The gap widens.

The organizations that started early with AI, that have been on this question for a while, that built things, shipped things, and developed capability will come back this autumn with rocket fuel.

This summer’s releases (Fable 5 and ChatGPT 5.6) amplify everything they already have in place. The ones who positioned themselves well didn’t do it by talking about AI. They did it by doing things.

And the distribution right now is wildly uneven. On one end you have all of Silicon Valley going “this is great, let’s go.” The capability in some places is insane. I’ve had glimpses into environments that are fully native in this. Where people just run entire departments and even companies with AI. It’s something else entirely.

On the other end, a lot of organizations are still standing on the platform, stamping their feet or torn between what models to choose, if any, and what should be allowed or not.

This autumn, I think the difference starts showing for real. The ones who did the work will get leverage from it.

We’re about to find out what that leverage looks like.

2. Every leadership consultant has discovered AI.

You can feel this almost everywhere. In PR outlets, on LinkedIn, and in requests that I get for collaborations. Everyone in the traditional leadership space is engaged in this now. The problem is that many of them aren’t practitioners.

Leadership still matters. As we saw a couple of weeks ago, a good manager sits in a decisive position for whether this goes well or not. So, on one hand, it’s great that these leadership consultants are now waking up to the fact that AI is an important and natural part of workplaces. But you need to get practical. You need to explore what’s possible and know the technology, hands-on.

Agencies are reaching out wanting to build capability. Great, of course. But this technology has been around for almost four years. Now you start?

If you are considering buying AI + leadership courses during this fall, be very stringent around their AI capabilities. What have they done? What have they built? Do they understand what’s possible and what’s not possible? Always happy to help with second opinions here; just reach out, and I’ll help you for free. This is important.

(How can I write this, for free? Won’t I get spammed with requests? No, I won’t because most of you won’t reach out even if you should…)

3. Cost drives open source.

Token cost stays in focus. The trend is open source, driven by the hunt for the lowest possible cost. Locking yourself into a single model is not a great idea. Spread your bets.

My most advanced clients have built their own harness around this and could now swap to Kimi K3 the day it drops. Big upfront investment in hardware. But as open source catches up as it has this summer, running it gets cheap.

One more thing. If you’re stuck in Microsoft with Copilot at work, you’re limited. I like Microsoft, they do a lot of good work. But you can’t practice on what’s happening outside that wall.

I’ve said it before, and I’ll say it again: Buy your own computer. Get your own license on ChatGPT or Claude. It’s probably the best career investment you’ll make this half year. Not doing it is professional negligence, and that’s not hyping it; it’s just what it is.

Because how are you supposed to lead an organization if you don’t understand the capability that exists in the world right now?

Those are the three big ones, what did I miss?

If you haven’t yet subscribed, join the 11500+ smart, curious HR folks by subscribing here:

Subscribe now

Curious to learn more about Claude Cowork and ChatGPT Work?
I run the third cohort of the popular Cowork for HR starting on the 27th of August.
Three sessions, and the first 10 to sign up get a 30-minute private coaching session with me.

Join now

In Sweden, we have a saying. Don’t cross the bridge to get water.
And that is very applicable when it comes to working with AI, in many different ways, but we’re going to talk about one very specific way today- meta-prompting.

When you meta-prompt, you let the AI do the work for you, and how I usually do that depends a bit on the task I have in front of me, but I’m going to show you both ways.

Let’s get to it!

Help me with this.

This is a fairly easy yet powerful one; I run this whenever I’m in a slight time hurry but still want good results. What I’m doing here is letting the AI do the heavy lifting and creating the prompt for me.

The more you describe what you want the AI to do, the better. If it’s about anything related to a document you have, be sure to upload the document as well.

So, for example, if you want to create a dashboard that visualizes your employee engagement data, make sure you upload the data set.

Let’s see how a meta prompt would look in that instance.

Help me write a prompt for creating an employee engagement data dashboard that I can show to my executive team. 

The data you should visualize is attached. 

Ask me any clarifying questions before creating the prompt.

In ChatGPT, this would look like this (and it would look very similar in whatever tool you are using as well).

And sure enough, it’s following instructions!

Instead of using the proposed answers, you can also give your feedback and guidance on what you want.

After you’ve answered all the questions, the AI will give you the prompt.
In this case… the prompt is long (which it usually is when meta-prompting like this)

Create a polished, interactive employee engagement dashboard as a working web app using the attached Excel workbook.

## Purpose

The dashboard will be presented to an executive leadership team. It should help them quickly understand:

1. The overall state of employee engagement.
2. Where the organisation is performing well.
3. Where intervention is most needed.
4. Which workplace drivers are associated with engagement, advocacy and retention.
5. What concrete actions leadership should consider.

The tone must be executive, concise and evidence-based. Avoid HR jargon, generic advice and unnecessary detail.

## Data source

Use the attached Excel workbook.

Use:

* `Raw Data` as the primary analytical source.
* `Survey Items` for question definitions, constructs and scales.
* `README` for methodology and responsible-use notes.
* The existing `Dashboard` sheet only as a calculation cross-check. Recalculate all results directly from `Raw Data`.

The file contains 252 synthetic respondents from Nordholm Systems AB.

Do not invent historical data, benchmarks, response rates or external comparisons. There is only one survey wave, so do not create trend charts.

## Build requirements

Create a fully functioning, responsive single-page web application, not a static image or wireframe.

The main view should be suitable for presentation on a 16:9 executive meeting screen. Allow users to explore further without making the initial view crowded.

Include an option to upload a replacement Excel file using the same data structure so the dashboard can be reused.

All filters, calculations, charts and narrative insights must update dynamically.

## Executive overview

At the top, show six compact KPI cards:

* Number of respondents
* Engagement index, scale 0 to 6
* Driver index, scale 1 to 5
* eNPS
* Average intent to stay, scale 1 to 5
* Percentage flagged as at risk

For each KPI:

* Show the value clearly.
* Show the scale where relevant.
* Include a brief plain-language interpretation.
* Do not label a score as good or bad using an invented benchmark.
* When filters are active, clearly indicate the filtered sample size.

Below the KPI cards, generate a concise executive summary containing no more than five sentences:

* Overall situation
* Most important concern
* Strongest organisational asset
* Most important segment difference
* Recommended leadership focus

The summary must update when filters change.

## Main dashboard sections

### 1. Engagement overview

Show:

* Distribution across Highly engaged, Engaged, Neutral and Disengaged
* Average Vigor, Dedication and Absorption scores
* The number and percentage in each engagement band

Use a horizontal stacked bar or similarly compact visual. Avoid pie charts.

Explain that the engagement score uses a 0 to 6 scale.

### 2. Organisational comparison

Create a sortable comparison table or horizontal bar chart by department.

Include:

* Respondent count
* Engagement index
* Driver index
* eNPS
* Intent to stay
* Percentage at risk
* Difference from the company average

Allow the user to switch the grouping between:

* Department
* Team
* Location
* Work arrangement
* Job level
* Tenure band

Clearly display sample size beside every group.

Do not automatically interpret small differences as meaningful.

### 3. Engagement drivers

Display all 12 driver scores ranked from lowest to highest:

* Manager support
* Recognition
* Role clarity
* Influence
* Development
* Team support
* Psychological safety
* Workload manageable
* Work-life balance
* Change clarity
* Meaning at work
* Tools and resources

For each driver, show:

* Average score on the 1 to 5 scale
* Difference from the overall driver average
* Correlation with Engagement Index
* Correlation with Intent to Stay
* Correlation with Recommend 0 to 10

Use Pearson correlation and label it clearly as an association, not evidence of causation.

Add a driver-priority matrix:

* Horizontal axis: average driver score
* Vertical axis: correlation with Engagement Index
* Label every driver
* Distinguish between potential priorities, strengths to protect and lower-priority issues
* Explain the logic in one short sentence

Do not use a single unexplained composite score. Leadership should be able to see whether something is a priority because it scores poorly, is strongly associated with an outcome, affects many people or appears repeatedly in comments.

### 4. What requires attention

Create a section titled **Leadership priorities**.

Identify the three most important issues using a balanced assessment of:

* Low driver scores
* Associations with engagement, eNPS and retention
* Percentage of employees affected
* At-risk concentration
* Comment-theme frequency
* Organisational segments where the issue is most pronounced

For each priority, provide:

* The issue
* The evidence
* The groups most affected
* A concrete proposed leadership action
* A suggested executive owner or function
* The first practical next step
* One leading indicator to monitor
* One employee-outcome measure to monitor

Recommendations must be specific to the data. Do not provide generic recommendations such as “improve communication” without describing what leaders should actually change.

Separate organisation-wide actions from local department or team actions.

### 5. What should be protected

Create a section titled **Strengths to protect and scale**.

Identify the three strongest aspects of the employee experience based on:

* High scores
* Positive relationships with engagement or retention
* Consistency across departments
* Positive comment themes
* Strong results in particular organisational groups

For each strength, explain:

* Why it matters
* Where it is strongest
* What leadership should continue doing
* How the organisation could learn from strong groups without assuming that their practices caused the result
* A warning sign that would indicate the strength is weakening

### 6. Retention and advocacy

Show:

* Promoter, Passive and Detractor distribution
* eNPS by organisational group
* Intent-to-stay distribution
* At-risk percentage by organisational group
* Groups combining lower engagement, negative eNPS and higher risk

Do not expose individual respondents or Manager IDs.

### 7. Employee comments

Analyse `Comment_Theme` and `Open_Comment`.

Show:

* Number of non-empty comments
* Theme frequency
* Positive and negative themes
* How themes differ across departments
* Whether qualitative themes support or contradict the quantitative findings

Do not show a long comment feed on the executive homepage.

Provide a separate expandable comments section with a small number of representative anonymised comments. Do not show comments from groups below the confidentiality threshold.

Do not claim sentiment accuracy beyond what can reasonably be inferred from the text.

## Filters

Include compact filters for:

* Department
* Location
* Work arrangement
* Employment type
* Job level
* Tenure band
* Age band
* Gender

Include a prominent **Reset filters** button.

Filters should update the complete dashboard, including KPI cards, charts, priorities and narrative summaries.

Show active filters clearly.

## Confidentiality and responsible reporting

Use a default minimum reporting threshold of seven respondents.

When a selected group contains fewer than seven respondents:

* Suppress scores and comments.
* Display “Insufficient responses”.
* Do not expose individual-level records.
* Do not allow combinations of filters that reveal a very small group.

Make the threshold easy to change in the code.

Treat demographic comparisons as descriptive only. Do not make assumptions about individuals or protected groups.

Include a small methodology note stating:

* The data is synthetic.
* The survey structure is illustrative and not a validated version of UWES, COPSOQ or another published instrument.
* Correlations do not establish causality.
* Results should be followed by employee dialogue before decisions are made.

## Visual design

Use an executive, modern and restrained design.

Use:

* Warm off-white background
* Charcoal text
* Muted coral or orange as the main accent
* Soft teal for positive indicators
* Muted red only for genuine risk indicators
* Generous whitespace
* Clear hierarchy
* Minimal chart decoration
* Accessible contrast
* Direct labels where possible

Avoid:

* Excessive gradients
* Decorative illustrations
* Large blocks of text
* Traffic-light colouring of every metric
* Gauge charts
* 3D charts
* Multiple pie charts
* Overly playful design
* Automatically colouring every below-average score red

Use colour together with labels and icons so meaning does not depend on colour alone.

## Interaction and presentation features

Include:

* A presentation mode that prioritises the executive overview
* Tooltips explaining each metric and scale
* Download or print-friendly executive summary
* Export of the currently filtered summary as PDF or a print-ready page
* A visible timestamp or survey-period label
* A short “How to read this dashboard” explanation

## Analytical rules

Use unrounded values for calculations and round only for presentation.

Presentation rounding:

* Index scores: two decimals
* Percentages: one decimal
* eNPS: whole number where possible
* Correlations: two decimals

Do not infer statistical significance merely from ranking differences.

Do not create external normative benchmarks. Company averages may be used as internal reference lines.

When identifying strong or weak departments, consider sample size and multiple measures rather than relying on one metric.

## Calculation validation

Before finalising the dashboard, verify that the unfiltered results are approximately:

* Respondents: 252
* Engagement Index: 3.77 out of 6
* Driver Index: 3.48 out of 5
* eNPS: +5
* Intent to Stay: 3.93 out of 5
* At risk: 32.5%
* Highly engaged: 16.7%
* Engaged: 45.6%
* Neutral: 34.1%
* Disengaged: 3.6%

The lowest company-level driver scores should be approximately:

* Workload manageable: 3.03
* Change clarity: 3.20
* Recognition: 3.31

The highest company-level driver scores should be approximately:

* Meaning at work: 3.84
* Team support: 3.82
* Tools and resources: 3.76

The drivers most strongly associated with engagement should include psychological safety, manager support and recognition. Verify this from the raw data instead of hardcoding the conclusion.

Potential organisational hotspots to investigate include:

* Sales, with lower engagement, negative eNPS and a relatively high at-risk percentage
* Finance, with negative eNPS and lower manager support
* Engineering, where advocacy and intent to stay may require attention
* Product, which appears comparatively strong on engagement and eNPS

Treat these only as validation clues. All displayed findings must be generated from the workbook calculations.

## Final output

Build the working app first.

After building it, provide a concise explanation covering:

1. The dashboard structure.
2. The main calculation logic.
3. The confidentiality rules.
4. The three most important findings in the unfiltered data.
5. Any technical or methodological limitations.

Do not ask additional questions. Make reasonable implementation decisions while following these requirements.

But why do this? As you see above, the prompt is usually way, way longer than if you were just to type “build me an employee engagement dashboard,” and my experience is that this yields a way better, more competent result.

Sure, it takes about 5-10 extra minutes upfront, but you have to spend less time adjusting after you get the initial results.

By the way, this is the result from the prompt. A nice-looking dashboard with all the info, as described. Dummy data, of course but I would say it’s rather nice-looking.

The ramble

That was the first of the two that I mentioned, typing and adding data. I do that from time to time, but this one I do almost daily.

What I do is that I start Wispr Flow and then talk naturally about what I want to do, the things I would want the AI to do, and in the end I add something like this.

Turn this in to a strong, cohesive prompt to solve the task that I just described. 

Together with a ramble, it looks something like this.

And the output is this prompt:

Write a compelling promotional message announcing the third cohort of Cloud Co-work, beginning on 27 August and running for a few weeks.

The message should speak directly to HR professionals and people working with HR, people operations, learning and development, or organizational transformation. Help them understand why practical knowledge of AI is becoming increasingly valuable in their work.

Explain that this cohort has evolved to include hands-on use of ChatGPT for work, along with an accessible introduction to how Codex can support more advanced tasks and automation. Emphasize practical applications: reducing repetitive administration, improving research and communication, accelerating documentation, creating more efficient workflows, and giving HR professionals more time for thoughtful, human-centered work.

Position Cloud Co-work as a supportive, collaborative learning experience—not a technical course. Participants should feel that they can join without being AI experts and learn by applying the technology to real workplace challenges alongside others.

Use a confident, warm, and credible voice. Avoid exaggerated promises, fear-based language, and dense technical jargon. Create a sense of timely opportunity without making the message feel aggressively promotional.

Include:

- A strong opening that captures why HR should understand these technologies now
- A clear announcement that this is the third cohort
- The start date: 27 August
- A concise description of ChatGPT for work and Codex
- Concrete examples of what HR professionals could automate or improve
- A clear explanation of who the cohort is for
- A reassuring statement that no advanced technical knowledge is required
- A compelling call to action inviting people to express interest, request details, or reserve a place

Write the primary version as a polished LinkedIn post of approximately 150–250 words. End with a short, memorable line that reinforces the central idea: AI should help HR spend less time on repetitive work and more time on people.

And there we have it; that’s the prompt I could use to create that promotional message for the course that starts on the 27th. Wink wink, the link is here.

I prefer the latter because I find these models are great at understanding my ramble, taking that into account, and then creating one beautiful prompt out of it. It also gives me the opportunity to look things through after I’ve rambled.
Has it interpreted me in the way that I wanted? Anything that I would like to add or change?

Do I use this all the time? No, I don’t, but I do use it, especially the latter one here, on a daily basis. It’s time-effective and gives me better results, and it’s a nifty little prompt trick to have in your AI arsenal going forward.

Let me know if there’s anything I can clarify or make easier here, should you need it. If not, then meta-prompt away!

One of the perks of writing a semi-large newsletter is everything that lands in your inbox. PR pitches, media monitoring, reports. Most of it is forgettable.

But one thing has struck me lately, there is a lot of “HR must” going around.

HR must lead the AI transformation. HR must own the change. HR must secure a seat at the table.

I have pushed that thesis myself. In this newsletter, on stages, in workshops. I have said it so many times it has become a reflex.

But what if it is wrong?

What if AI should not be treated as a people question at all? What if it should just be pushed out as technology? Install, activate, measure. More tech, less human. No change journey, no workshops, no HR strategy. Just rollout.

That sounds cold. But be honest. A lot of the technology we use today arrived exactly like that. Nobody invited HR when email showed up. Excel never got a change program. It just arrived, and we adapted.

And as I mentioned last week, what if the hype is true? What if agents can handle onboarding, salary reviews, screening, and engagement surveys better than we can? Then the question is no longer how HR should lead AI. The question is what is left.

Does HR need to exist? What is our self-justification in all of this? What is our role when the work that built this function can be automated?

I get a bit despondent at times. Not about the technology, about us. Because we defend our position so relentlessly instead of challenging it.

Every report, every panel, every LinkedIn post lands on the same conclusion.

“HR is more important than ever!”
Is that true, or is it just what we need to hear?

There is also an entire industry that makes its living telling HR how indispensable HR is. (I am part of it.)

Few things in life are static. Why would HR be?

The people function is roughly a hundred years old. Humans organized work for thousands of years without us. In the grand scheme of things, we are a small anomaly, not a law of nature. Yet we act as if our existence were a given.
As if we will always be here.

My point is not that we should shut it all down. My point is simpler. A function that cannot answer why it exists will get the answer from someone else. From a CFO with a savings target. From a CTO with an agent platform. From a CEO who notices the rollout went just fine without us.

So stop defending. Start challenging. Ask the question internally before someone else asks it for you. What do we do that creates value, and what do we do simply because we have always done it?

If we don’t reinvent ourselves, someone else will.