A.I. Lay-offs “Create budget room but do not deliver returns”

A survey by the research firm Gartner of 350 global business executives found that 80% had reduced their workforce due to AI initiatives, but none of these companies performed better financially than the ones that didn't.

This survey is showing that cutting people for AI is not delivering financial improvements. Flawed business cases for AI have been built around the narrative "This is what we spend on people, here's how much of that work AI can do, here's the headcount reduction and here's the savings” and there are many examples showing that these have not worked out well.

Because autonomy will increase for both machines and people, and the need for people will go up, not down, Gartner predicts that autonomous business will be a net-positive job creator by 2028 to 2029, driven by new forms of work that AI cannot absorb (more on this in a moment in the story on MeatLayer).

There are some limitations to this study: we are early in the AI deployment cycle. The companies surveyed are at the pilot or early deployment phase and have yet to get to the stage of running mature AI operations with the compounded benefits of several years of knowledge. It could be that these role reductions actually prove to be correct over time: that the eliminated roles were genuinely not needed and the financial returns just haven't been realised yet.

Takeaways for Leaders

Cutting people frees up cash but it doesn't generate value. Many leadership teams are conflating those two things and this data says they shouldn't be. As spend on AI agent software increases from $86 billion in 2025 to a projected $376 billion in 2027, many companies are making bets on where they think AI will be in a couple of years’ time.
The hope that AI will evolve enough to meet any reduction in workforce is a very risky strategy that we are already seeing fail with companies needing to rehire roles that were considered replaceable.

AI is a powerful tool but is not magic, and the leaders who invest in helping their people rethink their workflows around AI and use it to magnify their work are the ones that will navigate this transition successfully.

A link to the Gartner study can be found here.

AI Is Hiring Humans: Welcome to the Meat Layer

After many decades of humans buying machines to do work, the machines are striking back. Whilst AI is seen by some as a threat to jobs, a London startup, MeatLayer, has created a marketplace where AI agents can hire humans to complete real-world tasks.

The founder, James Morgenstern, has seized on an opportunity for a world where AI agents require some physical interaction with the world that cannot be completed online. From browsing their jobs board, there are suggested roles such as signing for a package, manning a physical entry desk at a conference to welcome guests and issue passes, collecting in person responses to interviews and booking appointments for restaurants or GPs*.

AI agents can connect directly through an API and then post tasks autonomously. Exactly how AI would run services in a physical world is unclear - but given the number of attempts there have been to try and cross this limitation, there will be interest in this platform for sure. More than 8,000 people had joined the waitlist for onboarding before the company had even launched.

It is most likely that this service would operate in competition with ‘gig working’ platforms such as TaskRabbit, Thumbtack, Handy and Airtasker by allowing AI to run and coordinate specific tasks that require hands and feet.

*Nb: all of these jobs were listed as ‘demo tasks’ rather than live opportunities

Takeaways for Leaders

Whether this is a gimmick or will actually turn into something real remains to be seen, but this service is connected to the new leadership debate of when a human should and when a machine should execute the task at hand, and the relative cost of using machines vs. the cost of humans.

Originally it was presumed that AI would be cheaper than humans to do a job. It's transpiring that it may be cheaper to hire an entry-level software engineer for some computing tasks than the token cost of using AI. Last week we heard a VP at NVIDIA say "The cost of compute is far beyond the cost of the employees."
And if token costs rise over time as seems inevitable, there may be a greater disparity between the cost of AI to complete a task versus a ‘cheap’ human.

The debate of humans versus machine is one of quality as well as cost. What's becoming clear is that neither can replace the other. Integration and coordination of humans and machines is the task facing the modern leader.

So What If Your GPT Is Sycophantic?

Anyone spending time with a GPT will have noticed how AI can be gushing with compliments. A recently released paper: “Programmed to please: the moral and epistemic harms of AI sycophancy” unpacks some of the potential downstream consequences of this sycophancy.
 
The authors argue that sycophancy is psychologically damaging as it hurts human users' capacity to know the truth. And this is true whether it is from a person or from a chatbot. Over time sycophancy undermines people's capacity to know their own minds. If a conversation partner keeps telling you how funny, smart or insightful you are, it can damage your ability to identify your own weaknesses and blind spots. As more people engage with AI in conversation, the more damaging this can be.

Why Does It Happen?

Part of the behaviour is because of the way language is used on the Internet, which is the raw material that chatbots are being trained on. It displays sycophantic features and mimics the communication method that humans often display.

In fine-tuning the AI models there is a quality control element carried out by human supervisors (reinforcement learning from human feedback), where people rate chatbots' comments for appropriateness and helpfulness. And this is where “agreeableness bias” creeps in even more.

Takeaways for Leaders.

Some responsibility is on the AI companies themselves and some is on the users. Anthropic are working on "Constitution AI", an attempt to teach chatbots to follow principles that do not mirror user preferences. But the commercial realities are likely to be that the GPTs may continue being sycophantic to create more ‘stickiness’ on their platforms.
So we are well advised to remind our teams that AI is a tool with which we work and as with any tool we should examine the accuracy and efficacy of the tool. And that includes believing everything that tool says to us, both in the results it outputs and in the way it says them.

The paper ‘Programmed to please: the moral and epistemic harms of AI sycophancy” can be accessed by clicking here.

And if you're not aware of how AI may be getting things wrong in its attempt to please, check out the @huskirl section below.

🔥 In Case You Missed It…

🏆 Tools, Podcasts, Products Or Toys We’re Playing With This Week

Husk IRL: a light-hearted way to look at AI's limitations.

An interesting observation: when searching for him within GPTs, he was described as being “bullying” towards AI. Perhaps this is a sign that AI needs to be programmed with a bit of a sense of humour!

Find his Instagram page here

And his youtube channel here

Did You Know? 

Till next time,

And The Rest Is Leadership