Fullstack HR

You can ask AI to complete a task. Someone still has to decide which task matters, how it connects to the rest of the business, and what happens next.

Sam Pasupalak, co-founder and CEO of Skyfall AI, joins me to talk about his team's ambition to build an AI CEO. We discuss why he sees planning as the missing piece, how world models could help a system anticipate the consequences of business decisions, and why learning continuously remains a difficult research problem.

Sam also draws a line between the operations he wants to automate and the employee relationships and motivation he believes will stay human.

And we get into where he expects automation to arrive first: small online businesses, including SaaS and direct-to-consumer companies, with very different implications from those he sees for larger organizations.

Candidates use AI to write applications. Companies use AI to filter them. Everyone ends up frustrated. So how do we break the hiring doom loop?

Sharawn Tipton, Chief People Officer at Greenhouse, joins Johannes Sundlo to discuss what happens when you give candidates a chance to speak before judging their CV.

We explore voice AI interviews, the candidates recruiters might otherwise miss, and why transparency, the option to opt out, and clear human accountability matter. Sharawn also shares how her own HR team uses AI every day, how she thinks about AI slop, and what she believes HR leaders should stop doing on Monday.

Plus, a lesson from her grandmother that applies surprisingly well to AI.

You can get everyone using AI and still have very little to show for it.

Syed Ali Abbas, Global VP of People for Business Units at HelloFresh, joins me to talk about what the company is doing with AI and how it judges whether the work is paying off.

We discuss the data cleanup behind its AI efforts, a US pilot offering 350 recipes to more than 25,000 customers, and what happens when the pace of change leaves employees and managers struggling to keep up.

Abbas also explains why HelloFresh gave people time to learn before building AI into performance expectations, starting with a skills self-assessment that neither HR nor managers could see.

And we get into his objection to HR leaders trying to become the CTO, along with the work he wants HR to focus on: redesigning workflows, developing people, and helping the business get results.

Kantar has around 12,000 employees and around 15,000 agents. Andy Doyle is responsible for both, and he explains how they reached that scale in 18 months without an engineering army.

In this episode:

• Why the CEO pushed for the title Chief People and Agent Officer, and what it changes in practice
• The bet behind giving all 12,000 people Copilot in three months, from 300 pilot licenses to everybody
• Change influencers instead of change champions, and the control Andy gave up to make it work
• The agent that cut a monthly HR and finance reconciliation from a week to 10 minutes
• Kantar's target of resolving 95% of HR service queries without a human, and the three agents that guard the risk line
• The agent factory, run out of HR, with 70 citizen builders and none of them from the technology function
• Why the HR community has the highest AI usage of any function at Kantar

With Andy Doyle, Chief People and Agent Officer at Kantar.

Links mentioned:

• Andy Doyle on LinkedIn → https://www.linkedin.com/in/ajdoyle/

In May I argued that AI is a work tool, not a perk. Three months and a lot of client work later, the gap between the people who have it and the people who don't is not closing. It's growing, and most organizations are fuelling it without meaning to.In this episode:- Why well-intentioned license decisions end up splitting your workforce in two- The two cohorts living side by side in the same organization, one running fast with their own Claude and ChatGPT setups, one barely started- The mechanism behind it, licenses follow whoever is comfortable asking- Why one client is taking AI access to the union negotiating table, and why I think they're right- Equal opportunity has to include the tools, not just the loud and vocal- The real missing piece, nobody has defined why AI matters for this organization- Why leaders can't be strategic about AI without being practical with it firstLinks mentioned:- AI Is a Work Tool, Not a Perk → https://www.fullstackhr.io/p/ai-is-a-work-tool-not-a-perk

Subscribe to the FullStack HR newsletter → https://fullstackhr.io

First episode of the new interview series with people who do stuff with AI, not people who talk about AI. Joe Shahmoradian runs talent acquisition for EMEA at Morningstar, and everything in this episode is running in his team today.

In this episode:

• How his team cut post-call admin by 75% with AI transcription on screening calls, and why the first tool got killed
• Amira, the AI teammate that works like a journalist and turned five-day market intelligence reports into 90-minute video briefings
• The AI clone Joe built before paternity leave, and what it did when he prompted it maliciously at 1 AM
• Morningstar's playbook for adoption at scale, an AI Academy, a champions community that grew from 5 to nearly 200, and an idea hub with real ownership
• What Joe would do in his first 90 days at a company starting from zero, and why governance comes before agents
• Why the workforce of the future is humans and agents as teammates, and why the discourse needs more boldness

With Joe Shahmoradian, Director of Talent Acquisition EMEA at Morningstar.

Links mentioned:

• Joe Shahmoradian on LinkedIn → https://www.linkedin.com/in/joeshahmoradian/
• FullStack HR newsletter → https://fullstackhr.io

Subscribe to the FullStack HR newsletter → https://fullstackhr.io

Claude started watermarking every text it writes. Substack scores your posts for AI. Recorded from the car after two weeks off, this episode is about why both miss the point.

In this episode:

• What Anthropic's new invisible watermark does, and what it can't do
• How Substack's Pangram scan works, and why it's easy to circumvent
• Why the human-versus-AI framing is the wrong test
• The pieces I let AI write on its own, and why they tanked
• The pieces I wrote alone that got the exact same feedback
• My workflow, dictate in the car, let Claude turn it into an article
• Why turning off the AI scan reads as shame, and what to do instead

Links mentioned:

• How Claude's text watermarking works → https://www.anthropic.com/news/claude-text-watermark
• How can I detect AI on Substack → https://support.substack.com/hc/en-us/articles/50891130623508-How-can-I-detect-AI-on-Substack
• Pangram → https://www.pangram.com

Subscribe to the FullStack HR newsletter → https://fullstackhr.io

A week at Almedalen made one thing obvious. Most leaders shaping how Sweden thinks about AI have only ever touched Copilot or the free ChatGPT, and that single reference point is quietly capping what entire organizations believe AI can do.

In this episode:

• Why almost nobody picks their own model, and how a Microsoft license becomes an AI strategy by accident
• The screwdriver vs power drill problem. Copilot solves the job, but it sets your expectations far too low
• How a single default tool shrinks the whole organization's sense of what AI is capable of
• The shadow IT trap, where your best people quietly run Claude or ChatGPT on the side and you lose both security and their hard-won knowledge
• Why limiting people to one tool becomes an employer branding problem over time
• What the real frontrunners do instead. Securing Copilot, ChatGPT, Claude, and Gemini and letting people choose
• The blunt take on IT departments that claim multi-model access is not doable on security grounds

Anthropic just dropped Claude Fable 5, and I've run it through the same test suite he's used on every major model for 2.5 years. The verdict is clear, but the bigger lesson isn't about the model at all.

In this episode:- Why Fable 5 feels like the most human-like model yet, and how it gives honest, direct feedback instead of patting you on the back- The worker scheduling test (30 people, legal requirements, hard constraints) and how Fable 5 handled it- A real client case where Fable 5 reworked a Copilot Studio workshop deck and improved the flow- The Mythos connection, the security guardrails, and why hacking headlines might be part PR- The pricing cliff on June 22. $10 per input token, $50 per output token, and Johannes's plan to pair Opus 4.8 for creation with Fable 5 for evaluation- The real bottleneck. Models are capable, organizations are not, and why experimentation time is the thing HR must fight for- McKinsey's warning that HR risks becoming a sub-department of IT if it doesn't take the driver's seat now

Links mentioned:- McKinsey HR Monitor 2026 → https://www.mckinsey.com/capabilities/people-and-organizational-performance/our-insights/hr-monitor

Subscribe to the FullStack HR newsletter → https://fullstackhr.io

Organizations keep telling Johannes Sundlo, the host here, "we've got the basics covered, skip ahead." Then he asks them to show how they use AI, and gets blank stares. This episode is about AI shame, where it comes from, and the one step you can't skip.

In this episode:

• AI shame, what it looks like and why it's more common in 2026 than 2025
• The pattern behind it. Buyers say "we're proficient, go straight to processes," but the proficiency is a sliver of what the models can do
• Why this shows up most in leadership teams, and why leaders who skip the grunt work can't set credible AI expectations for their organization
• The missing step. You can't jump to agentic processes before you've done the unglamorous work of learning what the model can do
• Educate everyone, including truck drivers without computers, because employees are already getting AI advice on salary and managers in their spare time
• Why being practical beats strategic-only thinking in AI adoption right now
• Lessons from roughly 110 workshops this winter and spring, and why Johannes is scaling that down