AI Is Not Taking Your Job. It Is Promoting You

Transcript

Transcript of the published episode, lightly edited for readability. Timestamps match the video.

00:00 Sebastian Selman Every conversation I sit through right now in Switzerland starts with the same question. Will AI take my job? And I think it's the wrong question. And it's the wrong question because it will produce the wrong answers, wrong hiring, wrong workforce budgets. So let's stop asking it. AI is not taking your job. AI is promoting you into a job I think you never applied for. And that job is management of AI.

00:33 Estera Kot And management of AI means promotion. And this promotion is something that arrives whatever you ask for it or not. So every knowledge worker right now is a manager and not of a manager of people, but of AI agents, of models, of automated pipelines, of scripts that produce work at the scale that one person could never produce alone.

01:08 Sebastian Selman So today's show is about the shape of that role. What it looks like, what breaks, and what a Swiss executive should do about it on Monday morning.

01:19 Estera Kot Hi, everyone. My name is Estera Kot. I'm the chief technology officer at Clouds on Mars.

01:24 Sebastian Selman I'm Sebastian, consulting lead for Switzerland. And this is Business Mission Control. Today's episode: Everyone Gets Promoted. Nobody signed the offer letter. And the smartest career move the next five years is not the one you think.

01:42 Estera Kot Okay. So now the mental model. Here, AI adoption happens in three stages. So not a big leap. Just three stages. And here, the stage number one. So AI is supporting you. You are a decision maker. And they help at the edges: it drafts, it summarizes, it suggests, but you decide. And this is the humanocentrism. This is the copilot mode, not piloting mode, but copilot mode. And most of the companies and enterprises are here today. But the next stage, the stage number two, is AI which is semi-approved. Semi-approved AI is just taking some chunk of your work from your arms and going further. And then it's progressing, proceeding, doing the work on your behalf, because you already fed the system. The system has learned what can do and what cannot. So you still approve kind of every output. But you approve some classes of this output. So not the individual decisions, not a single Lego blocks, but rather the outcome, the output. So the system watches which approvals go through, what you approve or not, reject or not, what the reasons. And then with the reinforcement learning and also different kind of itself learning, the system is growing in independence, just to make the work out of your arms. And then later on, once the system is mature, we are landing in a stage number three. Here we have a managed autonomous environment. So the system, because it learned how you behave, what you approved, and the system is getting better with the decisions, recommendations and actions. So now the system can run the runtime on its own. So you just supervise at the beginning a single one, but then later on you supervise a fleet of instances of agents. You take control only on the edge cases when it's needed. But in general, the system is autonomous.

04:30 Sebastian Selman And then people ask, what is stage four? Full autonomy?

04:37 Estera Kot Yeah, I believe that stage number three is kind of a full autonomy, as there is no stage four, because of the governance parts. We'll never land in a situation where things are for sure fully automatic and not governed. Here I'm talking about the European Union laws, the specific countries, politics, jurisdiction coming from the specific districts. Those politics will never make the systems fully automatic.

05:16 Sebastian Selman So the ceiling is not technical. The ceiling is institutional.

05:22 Estera Kot That's right. Fully aligned with the statement that it's not the model, it's not the AI, it's not the computer science that can do something right now. Because it really can. It's rather the model of accountability. Structure as it is in general will not allow just to make the situation when things are happening and no one is accountable for, maybe, and also because of the insurances, and also because we live in a systems that govern things.

05:53 Sebastian Selman So that means the shape of every knowledge job in the next decade is stage three, not stage four. Stage three: managed autonomous work with a human on the escalation line.

06:12 Estera Kot I think so. And that this is exactly where we are going to, knowing also a few cases coming from the autonomous industry, which are showcasing that basically AI is ready to be autonomous, to monitor itself, but will stay on the stage three just because of the law and how we want to use the technology in general.

06:41 Sebastian Selman So let's talk about maybe two cases, right? Both are already online. Both are shipping, not tomorrow. They are here today. Can you talk us through those?

06:51 Estera Kot Yeah, happy to. The case number one is Waymo. So autonomous vehicles running in Phoenix, San Francisco, Los Angeles, and the car drive itself. And again, I had the pleasure to sit inside. It was an interesting experience, just to see that no one is holding the wheel. And in general, at the beginning, I was thinking that, yes, I'm alone there. But the reality is that behind the car, in the office somewhere, there is a specific human being. And the role of this human being is being an operator. Operator is not driving a car. Operator is just supervising. So the person is usually starting from monitoring one car, one instance, and not being overwhelmed with a few ones at the beginning. He or she will operate two, three and so on cars at once, at the same time, once the confidence is growing, and the confidence related to precision, confidence related to accuracy, confidence also related to reviews coming from people who are using that. So it's kind of an experiment that people are participating in, giving the data. The data is fed to the end model. And then because of that, they can grow and become an operator of 30, 50 Waymos at the same time, and just reacting when something on the edge is happening, some confusion that the model didn't predict, or the model doesn't know how to behave.

08:50 Sebastian Selman So we start with maybe one car per operator and then scale to two, four to six and so on. As the confidence goes up, we trust the system enough to increase the number of cars that a single operator is responsible for. As the edge case rates drop, he or she can administer more cars at the same point in time. But we never get rid of the operator, right? So there's never the stage for pure autonomous driving. There will always be an operator sitting somewhere remotely who can take over.

09:28 Estera Kot Yes. Again, to have a fully autonomous world, I would believe that will happen in maybe two to three hundred years. May happen or may not. Why? Because right now, as of 2026, when you finish the law or the medicine, you still honor the rules that are coming from the Roman Empire, meaning that human is in the center and there are some ethical rules that we follow. And human being is basically approving. So the technology serves human being. And here, the operator's job is not to jump in and drive, is to catch moments where car doesn't know what to do, and then they can help with. And I had such a situation when I was passing Waymo and the road was super narrow, because there was a truck parked on the lane that Waymo was coming. And what I did: I was driving from the opposite side, I put myself on the right side. I have seen that basically, because I was passing the Waymo and also the truck that parked there, Waymo stopped and didn't know what to do. As a human being, I would stop and turn back, meaning that use the R and go back, just to bypass additional new car, and then continue the road. The reality here that during such situations, unusual, the operator is needed just to jump in and clarify. Also, remember that each case is a new material for retraining the models. So in the future, once there is a similar situation, what will happen? The model will capture what to do and then behave alone. So the need for operators will go down. But at the end, those are people who are accountable for that fleet of cars. So for those models.

11:34 Sebastian Selman So you mentioned two things, right? You said the operator's job is not to drive, but is to catch the moments where the car doesn't know what to do. Right. So the exception cases, where you encounter a construction site that has not been observed before, a police officer who waves traffic through a red light. And it's like, should I follow the policeman waving, or should I follow the rules of the red light and stop? So that's stage three, right? Well, yes, but what you said even there is what I should say: the car does the routine, the human does the exception. But you also mentioned something else that I want to maybe poke into. You said that we are still fundamentally obeying Roman legislation and making humans the accountable entity for actions. Though, I don't know when that exactly happened. Probably in the 17th century, right? We created the idea of a legal entity nowadays that historically were reserved for individuals and human beings, right?

12:43 Estera Kot Yes, that might evolve. Exactly, that's why I mentioned a few hundreds of years. And also I would take a look at on that aspect from geographic perspective or economic perspective. There are some countries which may be more willing to jump in and, you know, risk appetite for changing the law and allowing those autonomous system to exist. But I would not anticipate that it's going to be in general adopted. If the country is missing the human forces, workforces, they may be willing to use robots. But in general, the main purpose of running a country is to grow the country and provide human beings with a decent quality of life. And that's why I would suggest searching and observing the politics among some states or some countries which are allowing for a broader use, and defining who is accountable for what, plus also changing the way how insurance companies are working and how the law system is working. But in general, as for the entire population, I do not anticipate that those things will happen soon, mostly because of those constraints, that even though the technology can do something, humans are accountable for the signature and end result. That's why they are verifying. So basically, yes, we are operators, or we are co-piloted, but not piloted. The same is with airplanes. Pilots, they are using co-pilot or autopilot, so some system that is helping them to travel during the mist or when the weather is bad. But in general, they can still take a wheel and do it manually.

14:39 Sebastian Selman On to case two.

14:41 Estera Kot Yeah, the case number two is related to, yeah, kind of a my passion, hobby. It's related with empowering radiologists. So very different domain, but the same shape. So today a radiologist reads about, in general, two cases every hour, or per hour. And reading means opening the scans, looking at every slice from MRI, CT scans and different medical results, thinking and talking, consulting, considering different diagnosis. At the end, putting the diagnosis, writing the report, signing that up. And in the near future, speculation is rather high, so I think it's a good idea to do that. Also the experiment that are happening across the globe, we'll have a case when a radiologist is just signing off 2000 results, diagnosis coming from auto-detection system that is basically doing things on the behalf of their radiologists, but their radiologists is signing off. So here a model is flagging, going through all of the medical data, all of the slides, all of the images, then flagging what looks normal and what looks suspicious. Then later on another model is doing the detection of the specific organ, segmentation of the specific tumor and so on. So as we have a lots of different models for doing multimodal illness detection or tumor detections. So here radiologist is getting the output coming from the AI system, a few models that are doing also the debate between them whether it's accurate or not, explaining the results. And the radiologist is, yeah, co-piloted, meaning that having some, in general, recommendation, signing off, and in the ambiguous cases then going end-to-end. But in the case of a few models which detected zero probability of any tumor, why to spend the time on that patient data? So that's the reality, and the prestigious service that in the future we would like human beings just to review. But AI will help with delivering this service on scale. And we may say that, yes, we want to keep human beings doing this job, but take a look and take a perspective of the countries where no one wants to be a radiologist because of the huge accountability versus the salaries which are similar to jobs with lower accountability, or countries where just there are no people or education to do this job. So that's the future that is coming. We'll be provided by our assistants with some information which can help us with making a decision, or different decisions, just faster. And those recommendations ideally should also be backed by some explainability models, just to get into the details: what has been checked by which model and so on. But at the end, the radiologist is going to be still legally accountable, as that's the signature the person puts on the piece of paper. The scale that we can achieve, especially those who are in the field of radiology, those super qualified jobs which require years of learning, right now are going to be supported. It means that without having a huge amount of radiologists, but having the same amount of people, we can do more. So basically, we change the rate by the orders of magnitude.

18:55 Sebastian Selman So the pattern is, it's the same person, it's the same job title, it's the same license you need to get in order to be qualified to do it. And you've got a very different day. Instead of doing the work, they would actually supervise a lot of work being done and look at specific decisions that are low confidence, right? We don't look at the "we're 100% sure there is something in this observation", and we don't look at the cases of "we're 100% sure there's nothing in here", but we look at the difficult ones. And that basically allows them to scale, where basically 1,000 times in parallel by an assistant they can trust on certain cases. They routinely check, and then they manually check on the exception cases.

19:47 Estera Kot Yes. Here, just to jump in: "Let's make an impact", or "Let your data make an impact", which is a Clouds on Mars quote. It's just about it in a nutshell, meaning that we don't have to know all of the data and capture all of this data through streams, saving them all. We just need the data that are helping us to make a decision very quickly.

20:10 Sebastian Selman A difference that makes a difference.

20:12 Estera Kot Yes.

20:13 Sebastian Selman And that's not a metaphor, right? That is what every knowledge job is going to presumably look like in five years. Whether you look at the lawyers, the accountants, the tax advisors, the credit analysts, the compliance officers, you're going to be not only doing the work, but you're also going to be responsible to supervise a fleet of instances of the work being done.

20:40 Estera Kot Exactly.

20:41 Sebastian Selman Good. So why is the law still the ceiling? Why does full autonomy never arrive? Why is there no stage four?

20:52 Estera Kot I think that the main reason is because the accountability doesn't compress. Because by the law, human being is accountable. Because by the law, we cannot... We cannot sue AI, as who should we sue? Meaning that creator of the model, a person that tested, a designer, architect, cloud provider, manager, the CEO of the company? Again, we expect AI to be perfect, but normal software that IT is creating is never perfect. There are always bugs. So basically we cannot deposit damages into a language model's bank account, because there is no any. So the legal system does what it has always done. It walks a chain of accountability up to the human being. Because at the end, we are accountable for what we are creating. So the manufacturer, the deployer, the person that puts the signature behind the software. So people behind those two institutions, here we have regulators and we have courts, are many times not techy ones, even though they may use cloud or ChatGPT. But that will not change very, very soon, as going into those details requires, I would say, lots of time spent with IT on producing the software. And here, especially when we have cases when AI is doing something wrong, and people are following those cases, are super spectacular. Every media, TV are sharing what happened, who is accountable for. It's good learning for everybody.

22:52 Sebastian Selman You should share it. I mean, it's the negative cases, but that's the ones that we can learn from, right?

23:00 Estera Kot Exactly. And the public opinion, it is easy to retweet that, and the public opinion is basically biased by those cases, especially when there is, yes, AI word included, and sharing that AI suggested something would cost to a very different purpose. And we stopped to share the cases when physical machine without AI shared or did something what again was not designed or tested. So it's very easy and cheap food for being a star and commenting, rather than being fact-based and remembering that AI is also a software. And some models are open source, you don't have to pay for it, and you start using that AI. Open-source AI will suggest you something, you will do it, the result will be not a happy one, or not an expected one. Who do you want to sue? Meaning that the open-source software of the community? So it's rather about the definition, and circling back: what's the accountability? Are we forced to use AI? No. Are we encouraged? For sure. So let's take accountability for using, knowing that it's a piece of software. We as a human beings are making tons of mistake, so why we do expect AI to be perfect?

24:27 Sebastian Selman Yeah. So you're saying there might be something like an availability bias, because press obviously is gonna reproduce the negative cases more than it will talk about all the mundane tasks that are being assisted, and to a large degree driven, by co-pilots and only supervised by people. And that potentially means that the ceiling is actually getting lower, not higher. So every time an AI incident lands on the front page, we are increasing scrutiny around model usage rather than opening up to it. And probably also where you sit will change how low or high that ceiling is, right? That's going to change depending on your location. Obviously there is countries that are tighter in regulation and there's countries that are more risk-taking. In the European Union you've got the EU AI Act that explicitly tightens the controls, right? So high-risk systems require a human in the loop by law. It's not best practice, it's law. Whereas in general, I think globally, at least as far as I read the different big institutions discussing that, everybody agrees it should be a good practice. Switzerland: the Federal Act on Data Protection expects a human-reviewable decision for any automated individual decision with legal effect, where for instance, like, every credit or every automated credit decision and insurance decision, benefit decision has to be defensible in front of a human reviewer. So you need to create an audit trail, stuff like that. In the United States and China, the ceiling probably floats a little bit higher. The accountability principle doesn't disappear, it just lands probably on different people. It's the vendor, it's the deployer, it's the prompt author. Someone senior stays on the hook here. So I think, in summary, if you're planning European business for the next decade, plan for the tighter version of the ceiling, and that's going to shape your operating model, your reporting lines and probably, most importantly, your hiring plan.

26:52 Estera Kot Yeah, I would say that from the business perspective, the best way to test AI is to be inside and tested inside those countries with a huge appetite for risk. But as a consumer, even yesterday I was switching one server from U.S. or China to European Union. Why? Because of those laws. As a consumer, I feel more protected. Let's call it by name. It also means, and puts the fundamental question, how the Europe can speed up the experiments in such a regulated environment. How we can become competitive when there are so many papers to fill up, so many tests to do before we can go to production with some AI-based system? But in a nutshell, I would also like to emphasize the critical role of the institutions which are tested, claiming, supervising some technologies, solutions which are using production, which are used by human beings and where AI is behind. Because majority of us are using cars with some cruise control, and if there is a failure in the system, and we have, hopefully not a big one, but some incident, who do we sue? There is a homologation institution as well in every country, and so on. But you can take this perspective, that the more kind of a regulator, or the more tested, that's the thing, the more tested the solution is, the more confident the end customer, consumer is about the safety. And through the safety, yes, we build the trust, and the adoption of AI can happen only when there is a trust.

28:51 Sebastian Selman Also, for testing we need an option, right? Somebody needs to do the testing somehow.

28:57 Estera Kot Yes. But I believe that a lot of people who are willing to, you know, take the adrenaline and test whatever the tool and the solution is about. But yeah, full autonomy is not here, it's rather managed autonomy. So yeah, we can sit comfortable.

29:14 Sebastian Selman Here comes the auditor pivot of the discussion, right? I used to work as an auditor for one of the Big Four in the beginning of my career. The received wisdom that I got right now for many around me is that automation is bad news for the auditors, right? For the Big Four, like, bad news for internal controllers, bad news for the GRC teams, right? The governance, risk and controls. The assumption is they're going to automate away the checks, they're going to automate away their reconciliations, they'll automate away the audit trail generation and checking, so their jobs must shrink. I don't think that is what will happen. I think the opposite. Because all the auditors who currently are fearing for their work and are afraid that autonomy is going to take away their jobs, they're going to be in huge demand, because they are the only ones, or they are the best positioned ones, to establish and design control systems cleanly. Why? Because I think stage three as you described it, right, requires controls, requires the operator to engage when something is an edge case and goes off the beaten track, right? As long as the model is running the routine, it's all good. But a human needs to be supervising, and that human needs to be able to prove they were supervising. So they need a signal when the system is drifting, they need an audit trail, they need role-based access: who could do what to which agent. They need behavioral attribute-based access: under which conditions is access granted. And all of that is control-system thinking. It is what, I would say, auditors have been doing for decades, except now the audited object is a fleet of agents and not a general ledger.

31:12 Estera Kot Yeah, a fleet of agents that you cannot sue. In that case, as an auditor, I believe that you took a look at what people are creating, and there was always accountable for that. In that case, it's not only about new systems for real-time data and capturing all of the signals coming from different tech systems, but also telemetry that is expressing what the system is doing. But at the end, it's about who do we control at the end: agents. So much more data, plus something what we really want to verify in terms of the precision. And it's easier to change, I believe, that fix an error coming from human being rather than an error coming from AI.

31:59 Sebastian Selman And you'll always want what I call a neck to strangle, right? You want to be able to identify the person who ultimately is responsible for what's happening on the ground. Now, if we look at a governance stack of a, let's say, Microsoft tech-heavy Swiss enterprise, maybe you have Purview labeling your sensitive documents so that the AI knows what it can and cannot see. You've got Entra ID Conditional Access, so the AI can only act within the right identity, from the right device, in the right posture. You've got maybe Defender for Cloud Apps to catch shadow AI usage in your population. You've got Sentinel and the security operations center picking up the audit trail. And on the horizon then you've got something like Microsoft Agent 365 as the registry that tells you which agent or which agents the organization has deployed and who owns each one. If I look at that, to me it does not look like a stack that an engineer would build. It's a stack that an engineer builds with a senior controller. Like, an engineer on their own, I would assume, would come up with a simpler, maybe less controlled version of making that system run.

33:12 Estera Kot Every hyperscaler, every enterprise is landing the same architecture with different names on the boxes, and the pattern is just a pattern. So someone who is a senior has to be responsible for what the agent is doing. There are some politics that are controlling that, and that person needs to get things delivered in a shape that can be defended, or it's possible to make this case defendable by control.

33:46 Sebastian Selman So the career call, and I think that is the line I want every listener to take home, is: if you are a senior control-system thinker today, an auditor, an internal controller, a GRC lead, a risk officer, I think you are in a growth job for the next decade, not a shrinking one. Because nobody is putting that aside, it's becoming more prevalent.

34:15 Estera Kot Yeah. And by the way, are you looking for hiring for this job? If so, big good luck. They're about to be severely contested hires in the Swiss market.

34:28 Sebastian Selman So here are the three concrete moves for Swiss B2B or B2C, for any Swiss leader listening right now, right? Move one, I would say: stop budgeting for headcount reduction. That is the wrong number. You're not shrinking, you're reshaping. The 2027 workforce plan you're writing right now needs a role-mix column, not just a body-count column. Governance and controls roles up, manager-of-many-agent roles up, individual expert roles only as a fraction of the total count would go down, but including the agents, for sure going up. The move two is invest in span-of-control redesign. I think managers in stage three of the organization, where we've got this supervised, governed autonomy, those managers do not supervise fewer people. They supervise fewer people and many more agents, and that changes the coaching conversations, that will change the performance review conversation, the compensation conversation. I think nobody has really proven and battle-hardened playbooks for this yet, but whoever writes yours first, the quickest, and then is able to improve on the most important part of that will probably win the war for talent. Then move three: hire the auditor. Not next year, this year. I think the pipeline is drying up. Every controller with strong systems thinking is about to have three job offers on a table, and even with them becoming managers of agents, they will still need to be in control. So that would be my three. The reshape, I think, is not a spreadsheet exercise. It's a career conversation that I'm going to have to make sure that it's... that needs to be held with your most senior individual contributors, and probably better to have it next quarter rather than next year.

36:26 Estera Kot Definitely. So, five things to remember out of today's podcast. The ceiling is institutional, not technical. So for the full autonomy, one day it may come, but it's not arriving, and not anytime soon. So the stage three, which I've mentioned, so managed autonomous, is the designation and destination, and the design for that. The number two is that every knowledge job becomes a management job. So you become a supervisor of a fleet of instances of work being done by agents. So you catch exception, you set the rules, guardrails, and you sign the escalations. Just that. Yeah, and it's being from your side, Sebastian.

37:21 Sebastian Selman I think learning number three would be sovereign accountability, and that's a board-level decision, right? The EU AI Act, the Swiss FADP, they make the ceiling explicit in Switzerland and Europe, and your operating model needs to survive within that ceiling. So choose your automation architecture with your local regulation in mind.

37:44 Estera Kot Good. So number four would be: critical and regulation-critical workflows keep a human being on the signature, for sure. Like radiology or oncology, credit, insurance, benefits, legal. There the signature stays with a person. What changes the rate, as they can do more with being co-piloted.

38:11 Sebastian Selman Yeah. Finally, number five: the auditor is a growth job. I think every controller with a strong systems-thinking mindset and capability is about to be the most contested hire in the Swiss market, and I think that also needs to potentially reshape the span of control where those people are gone from and the job market dries up.

38:32 Estera Kot I'm Estera.

38:33 Sebastian Selman I'm Sebastian, and this was Business Mission Control.

38:36 Estera Kot We show notes, we show sources, we show the public transcript, but each episode you will find them in the description. See you in the next one. See you soon.