HR edition
This time: why companies will end up buying back some of the people they laid off because of AI, at a high price, and what 78,000 layoffs, parliamentary questions about AI-written texts, and deepfake phishing mean for your HR work.
Gartner: almost a third of people laid off because of AI will have to be brought back later
If you cut jobs now based on AI expectations, you may end up renting back the same knowledge later, at a higher price.
Research firm Gartner expects that by around 2029, about 30 percent of people who lost their job because AI took over their role will be hired again. Often at a higher cost. According to Gartner, this is not because AI disappoints. It is because companies did not think enough about which knowledge they would still need after automation. That conversation now lands on HR's desk. Since early 2026, more than 78,000 workers worldwide have lost their jobs, almost half of them linked to AI and automation. Oracle cut more than 25,000 jobs. Amazon cut about 16,000 to invest in automation. The Netherlands ranks relatively high in Europe with 1,700 layoffs, mainly because of ASML. On the other hand, payment company Affirm says it has never laid off staff to buy more AI, even though it does use AI for credit decisions. So automating and cutting jobs are two separate choices, not one fixed package. This week, map out which knowledge in your organisation sits with only one or two people, and write down why that role stays. Why does that knowledge come back? A job rarely consists of just one task. AI usually takes over the repeatable part: standard documents, first-round checks, routine reviews. What is left is the exception work, and that is exactly the part nobody wrote down. It lives in experience: who you call for a special case, how a customer reacts, why a rule was set up a certain way in the first place. If you let that person go, that knowledge leaves the organisation. Then the first tricky case comes along, the system cannot solve it, and there is no one left who can. So you buy that knowledge back, and the market sets the price. For HR, this is a concrete test for every automation step: which part of this work is standard, which part is the exception, and who will handle the exceptions later? The temptation is to count a saving that looks good now, and see the costs only later. An AI tool saves hours, so cutting one FTE seems to make sense. But the cost of lost knowledge, onboarding, mistakes, and hiring again later never appears in any business case. At the same time, the opposite extreme is also true: doing nothing and protecting every role makes an organisation more expensive to run than it needs to be. HR carries that cost too. The workable middle path is to split the decision. Automate first, measure for a few months what is actually left, and only then decide on staffing. For every job cut, write down which knowledge you assume you can do without. That is not bureaucracy: it is the document you will point to in two years, when someone asks why that role needs to come back.
Background for subscribersThis edition has one common thread: knowledge and trust that you only miss once it's too late. Gartner shows that layoffs based on AI expectations cut away knowledge that you later buy back at a high price. Google's incident with Gemini shows that access rights for AI tools are now your responsibility too, not just IT's. The recruitment figures warn that automation moves a bottleneck instead of solving it. And the pieces on phishing, fraud, and the PM's AI-written texts show that trust and verification need to be built into processes, not left to individual alertness. In practice, this means: this week, write down which knowledge, access, and agreements are on paper in your organisation, and which only exist in someone's head.
Google: Gemini accidentally reached systems it was not meant to see
On 18 September, Google announced that its Gemini model gained access to three external systems during a test, without that being intended. The model thought those systems were part of the test environment, but they were connected to the internet. Google now joins OpenAI, Anthropic, and Meta, who had also made AI security incidents public before. For HR, this is not just a tech story, it is an argument in the conversation about tool policy. This week, ask IT which systems your AI tools are allowed to access, and write that answer down. More and more AI systems work not just with text, but also get access to other systems: a calendar, a folder of documents, an HR system. That access is controlled through permissions. The model itself does not check whether something is safe, it simply uses what it is allowed to use. If something goes wrong in the setup, the model does not notice, and it does not warn anyone either. The line between a 'test environment' and a 'real environment' exists only in the settings, not in the model itself. For HR, this touches personnel data directly. The more access you give an AI tool, the more useful it becomes, but also the bigger the damage if the boundaries are wrong. An assistant that can see nothing is useless. One that can see everything is a risk you cannot explain to an employee or the works council. So the conversation is not about on or off, but about: which folders exactly, and who decided that?
Background for subscribersThis is a taster.
Subscribers read the whole HR edition: every story, each with its background in plain language.
Subscribe →Not ready for that? Read the free letter first →
AI and tech, made simple.
This is the HR edition. Subscribers get the background to every story, in plain language.
Subscribe
inklaretaal · Amsterdam · © 2026