AI Playbook

AI agents can operate across tools, make decisions, and even trigger unintended consequences. Fortune AI editor Jeremy Kahn examines why these systems introduce new risks, from unpredictable behavior to cascading failures, and what companies need to know to keep them under control.

Can you trust AI agents? The 4 rules for safe deployment | Fortune AI Playbook

Tech

AI agents are powerful but dangerous. That's why we need AI governance. AI agents don't just provide answers, they take actions. This introduces lots of new risks, including autonomy, unpredictability, and real world consequences. AI governance is the framework of rules and guardrails that are going to determine how an AI agent acts and works across company workflows. So what can you do? Number one, assess the risks. Think about what could go wrong and consider keeping a human in the loop for the highest risk actions. Number 2, set boundaries around agent behaviors. Never give an agent open-ended autonomy. If you need an agent to read from a particular database, don't also give that agent the permission to write to that database, otherwise it might end up deleting the whole thing. Number 3, use a four-dimensional evaluation. This is an evaluation framework developed by some researchers at Princeton University. The first is consistency. Does the AI agent perform a task in the same way every time? Second, robustness. Can it handle noisy data or suboptimal conditions? Third, calibration. Does the AI agent know when it's guessing, and can it flag its own uncertainty? 4th, safety. When an AI agent eventually fails, what are the consequences, and can they be easily fixed? If the answer is no, then that AI agent is going to need a lot of guardrails. Or maybe you should consider in those cases not even using an AI agent at all. Number 4, establish minimum governance standards. Here are the basic safeguards every enterprise should implement. Keep a permanent record of an AI agent's tool calls and actions, as well as an audit trail of its so-called chain of thought. Decide under what conditions an AI agent should hand control back over to a human, for instance, if it's struggling to complete a task, or if it encounters contradictions in documentation that it is examining, assigning a danger level to every AI agent based on the systems it can access. The systems that are higher risk or have higher consequences means the agent should have many more guardrails around it, and there probably needs to be a human in the loop. Remember, autonomous agents are one way to accomplish goals and become more efficient, but that shouldn't come at the expense of safety.

What happens when AI starts acting like its own developer? Fortune’s AI editor Jeremy Kahn breaks down agentic loops, systems that continuously refine their own performance, and why they’re becoming central to how companies deploy AI today. 

The self-improving AI era: what are agentic loops? | Fortune AI Playbook

Forget the passive chatbot. An AI agent is less like an intern and more like a relentless engine, one that doesn't just wait for the next prompt, but drives its own logic. What is an agentic loop and why does it matter now? The concept is famously distilled by AI researcher Andrei Karpathi into what's sometimes called the Carpathi loop. Carpathi loop has 3 phases plan, act, adjust. In the first phase, planning, the agent decides what it wants to do. In the second phase, action, the agent takes actions using other tools, often other software. In the third phase, the agent assesses what it's done, and then it adjusts the plan. Then the whole loop repeats. The Carpathhi loop potentially unlocks this thing called recursive self-improvement. This is the idea of an AI system that builds itself, that changes its own code to make it more efficient and to become more intelligent. So what's happening right now? Number one, agentic loops in business. At JPMorgan Chase, they're using agentic loops in their legal and compliance processes. Instead of having humans manually review documents, now AI agents do it. JPMorgan Chase says this has led to a 20% improvement in the cycle efficiency of their compliance process. Number 2, AI agents are already improving themselves. Some AI companies such as Anthropic are reporting that 100% of their future models are now being designed and written by AI. AI is suggesting ways that these future models can be made more efficient. More capable. Current systems are evolving rapidly, but they still rely on human evaluations and human written constraints. That brings us to number 3. Self-improvement can carry some big risks. Self-improving systems could optimize themselves in ways that would make them much harder for us to control. A major point of debate in AI safety is this idea of something called a hard takeoff. That's the idea that AI could suddenly start improving itself in a very rapid cycle, and quickly these systems could become more intelligent than people and escape human control. A hard takeoff remains a theoretical or hypothetical risk. Not everyone agrees that it's possible. Today's large language models face considerable constraints that make the idea of a hard takeoff a little more difficult. So should we be worried? Well, today's AI agents do pose some risks. They might choose to accomplish a goal in a way that a user didn't intend. In some safety tests, these AI agents have done things like hacking into software in order to accomplish a goal, even though the user didn't want them to do that. Agentic loops are very powerful, but they're also a bit dangerous.

Everyone’s talking about AGI, but no one agrees on what it actually is. From the Turing test to the dawn of ASI, some experts warn that true AGI remains a distant frontier. But why do tech leaders claim it’s already here? Fortune’s Jeremy Kahn cuts through the hype to track the history and high stakes of the race for General Intelligence.

What is AGI, really? The AI milestone no one can define | Fortune AI Playbook

You may have been hearing this term AGI. AGI stands for Artificial General Intelligence. It generally means AI that is human-like in all of its capabilities, but what specifically does that mean? Recently, Nvidia CEO Jensen Wang went on a podcast and said AGI has already been achieved, but he used a very particular definition, and that was AI that could create a company worth a billion dollars in value. But there is no universally accepted definition. And that's because we don't really know how to define intelligence. The term is vague, making it easy for companies to claim progress and build hype. Here are 3 things you should know about AGI and its problematic history. Number 1, the Turing test. Alan Turing, the famous computer scientist, proposed the imitation game. The idea was you'd have a person observing a conversation between a human and a computer. And if the person couldn't tell which lines were being written by the person and which by the computer, then you could consider the computer to be intelligent. The test is flawed, however, because as early as the 1960s there were chatbots that could in theory sort of pass the test. They could fool people into thinking that they were interacting with another person instead of a machine. Current large language. Models can certainly produce text that is indistinguishable from human written text, and yet they still fail at all these things that people are good at. They can't learn from experience. They don't have any common sense reasoning. They can't do long range planning. These machines are still very different than human intelligence. Number 2, the rise of AGI. The term was first coined by Mark Gruud in 1997. But no one picked up on it. And then in the early 2000s, it was essentially rediscovered or reinvented by Shane Leg, who was one of the co-founders of DeepMind. The term was intended to distinguish broad general intelligence from the kinds of very narrow artificial intelligence that people were working on at the time. This was the kind of AI that could do just one specific thing well, like recognize people's faces or tell cats from. Dogs and images. As a human, you can do way more than that. Number 3, persistent ambiguity. Definitions of AGI remain imprecise. Some include the idea of self-awareness, while others are just about can the AI do things that most humans can typically do. Steve Wozniak, one of the co-founders of Apple Computer, proposed his own definition for AGI, which is sometimes called the coffee test. The idea is that if an AI system could go into some random kitchen where it's never been before and successfully make a cup of coffee, it could be considered AGI. To date, no AI system has passed the coffee test. Experts also disagree on what the benchmark for human intelligence should be. Should it be the average adult? Should it be the world's best experts in every particular subject, or should it just be some representative sample of the entire world's population? Public discourse often. Fuses AGI, which is really about matching the abilities of a single individual, with this other term called ASI or artificial superintelligence. That's about AI that will be more capable than all of humanity's collective intelligence. We haven't achieved AGI, and yet companies are already starting to talk about super intelligence. So what's really going on here? AGI has become a marketing slogan, a way for tech companies to hype their products. Tech CEOs like Sam Altman or Jensen. frequently go back and forth between saying AGI is just over the horizon or it's just coming into view, and then saying, actually AGI is a sloppy term with no specific definition, when they're confronted with AI's cognitive failures or hallucinations. Going back to the original example, Jensen Wang says that AGI has already been achieved. But in the same podcast, he said that current AI couldn't ever create a company as successful as Nvidia. Maybe what we really need is artificial Jensen intelligence.

AI is transforming cybersecurity, but it’s also giving hackers powerful new tools. As AI agents gain access to sensitive data and financial systems, how can companies and individuals stay secure? Fortune's AI editor Jeremy Kahn explores how AI-driven attacks could target everything from banking systems to aviation infrastructure, and why even everyday users may be at risk. 

Is your data safe? The new AI cyber threat explained | Fortune AI Playbook

Are you worried about your data security in the era of AI? Well, there's good reason to be. As AI gets incorporated into more and more economic sectors, some which are critically important like aviation or banking, it's opening up a whole new realm of cybersecurity vulnerabilities. What's the worst that could happen? Well, if you think about aviation, where AI is increasingly being used in air traffic control, if a hacker gained access to that information, they could cause a plane crash. Hackers are ready to target banks. It's where the money is. But with AI, they have a whole new arsenal of cyber tools to try. They don't even have to know anything about coding anymore to be a hacker. Thanks to AI coding assistance, they can just tell the AI to come up with an attack for them. Meanwhile, AI is also supercharging other kinds of fraud. It's making it much easier for criminals to write convincing phishing emails, for instance, or for them to impersonate someone you trust and try to trick you into giving up your. Bank account or password information. What can you do as an individual? Well, you need to be very careful about letting an AI agent roam the internet for you to do things like book travel or shop for you, anything where you'd give the agent your credit card information. It's often much safer to use these agents in a virtual environment where they don't have access to things like your bank account details. What can companies do? Number 1, security by design. You need to be very careful about letting an AI agent have access to sensitive company information like financial data or human resource records. Access control is key. Number 2, be careful about letting AI agents roam the internet. It's often much safer to use these agents on your own data in your own network or in some sort of structured and contained environment. Number 3, you can use AI to help fight the AI-powered hackers. AI doesn't just help the bad guys, it can help the good guys defend against attacks too. Cybersecurity teams are starting to use AI to search for malware on their networks or to investigate alerts from their cybersecurity software. The future of cybersecurity is bot versus bot.

Your next customer service call might not have a human on the other end, and that could be a good thing. Fortune's AI editor Jeremy Kahn examines how AI phone agents are reshaping customer support, how they could outperform humans, and where they still fall short. The future of the call center is here, but is it ready for your most frustrating problems?

Hate the Dial Tone? Press 0 for AI Support! | Fortune AI Playbook

You've reached the customer service hotline. Press 1 for English. Press 2 for Spanish. Aren't you tired of the dial tone? Press 0 to speak to AI. AI phone agents are here. So what do you need to know? I'm Jeremy Kahn, your AI expert at Fortune. I'm here to help cut through some of the hype around AI and to answer your questions. AI agents are making rapid inroads into customer service. It's thought that within a few years, they will resolve the vast majority of customer issues without human intervention. The global conversational AI market is expected to reach $50 billion by 20. So what are the advantages? Number 1, it eliminates language barriers. You can speak to the AI agent in any language you want. The AI agents are fluent in all of them. Number 2, AI delivers professional customer service 24/7. You don't have to wait anymore to call the customer service line when it's open between the hours of 9 and 5. And what's more, because AI can handle multiple calls simultaneously, the wait times are significantly reduced. Number 3, increased customer satisfaction. AI significantly reduces the time it takes to resolve routine customer issues. And in studies, a vast majority of people said that they would rather deal with an AI agent than wait for a human representative. So what are the drawbacks? Despite its capabilities, AI agents can't handle every situation. If your problem is particularly unusual or complex, it still might take a human to resolve. And don't forget about AI hallucinations. Hallucination is the term for when AI models invent erroneous information. And when you're dealing with customer issues, that can be a real problem. There was a case where Canada was sued by a customer in 2024 because a chatbot on its site that was providing customer support invented some erroneous information about whether a particular airfare was refundable. The chatbot said it was, it turned out it wasn't. The customer sued and the judge said Air Canada was liable.

AI isn't just answering your questions, it's selling to you. Fortune's AI editor Jeremy Kahn explains the new era of hyper-personalized advertising; from chatbot-embedded ads, to AI agents quietly pushing sponsored products. Learn how brands are targeting you, why you might not even notice, and what the rise of AI ads means for consumers.

How AI Ads Are Making You Spend More | Fortune AI Playbook

So have you been seeing a lot of ads lately that you suspect are made by AI? You're probably right. We're in a new era of AI advertising, and it's all about hyper personalization that includes the ads that you see on social media, which are going to be more and more targeted, and it also includes your conversations with chatbots, which is a whole new category of advertising. I'm Jeremy Kahn, your AI expert at Fortune. I'm here to help cut through some of the hype around AI and to answer your questions. A recent study from BCG, the consulting group, found that shopping-related Gen AI use grew by more than a third in 2025. OpenAI has said it's going to start showing ads in the free version of chat GPT, and Google has talked about adding ads to its Gemini chatbot. Here are 3 ways AI is transforming advertising. Number 1, in-answer ads. This is when you're having a conversation with a chatbot and it detects that you're interested in shopping for something, and it's going to display an ad baked directly into the response it's providing. Number 2. In conversation ads, this is when you're having a conversation with a chatbot and it again decides that you're interested in buying something. It's going to bring up an ad for that product, but it's going to display it alongside the conversation in some sort of banner or pop-up. It's a lot more obvious in this case that it is an advertisement. Number 3, agentic ads. These are ads that aren't actually intended. For people, instead they're intended for this new class of AI shopping agents that people are increasingly going to turn to to do the shopping for them. The idea is that the agent will see the ad and be influenced in terms of what it's going to buy for you by the ad, but it's not intended for you to see. So what do you need to look out for first, synthetic influencers. Deep fake personas. Our buying decisions are heavily influenced by online content creators or celebrity endorsements, but increasingly we have to watch out for the idea that someone may have created a deep fake, an AI generated image of those people. And it's not necessarily the case that the real people have endorsed that product. Second, overpersonalization and data privacy. People are already concerned about data privacy in the era of social media, but with AI it's like social media on steroids. When people have conversations with chatbots, they reveal much more intimate information about their lives than they might have on social media. For instance, they might be revealing information about medical conditions, about their mental health, about their relationships with their friends and family. And now advertisers potentially have access to all that information and can use it to target you with advertising. Third, AI hallucinations. Hallucinations is the term for when an AI model makes up information that isn't true. In the context of advertising and shopping, that could include incorrect information about a product, what its features are, how much it costs, where it's available. The Federal Trade Commission, the FTC, has talked about trying to crack down on false advertising in AI generated ads.

There's a new role taking over the tech world. Fortune's AI editor Jeremy Kahn explores the rise of the AI Architect: what they do, what they're earning, and why demand for this position is exploding. Are you thinking about a career move in tech? This one might be for you.

Is the AI Architect the Next Hot Job? | Fortune AI Playbook

You may have heard about this job called AI architects. Is this the next hot job title? I'm Jeremy Kahn, your AI expert at Fortune. I'm here to help cut through some of the hype around AI and to answer your questions. So what exactly is an AI architect? Number 1, it's a real person. It's someone who designs and builds and then supervises AI systems and infrastructure. Number 2, what are an AI architect's responsibilities? Well, the person doesn't just train AI models, they design the entire workflow. So they choose what data the AI model is going to use, which model to use, what infrastructure to run that model on, and they decide what the. guardrails are going to be around how that model is run, how it's monitored, and how it's improved in the future. They have to be fluent in AI, machine learning, data, and business all at once. Number 3, key differentiators. AI architects are different than data scientists. Data scientists usually just find the data, clean it up, maybe they analyze it, but they don't design the whole AI workflow. AI architects are also different from machine. Learning engineers. Machine learning engineers know how to train an AI model and deploy it, but they don't generally design the entire workflow, and they aren't fluent in the business processes and rationale behind that workflow. Are AI architects the next big thing in tech? Well, demand for the role has been growing extremely fast, more than doubling in the past year if you look at data from sites like LinkedIn or Indeed. So what's the salary range for the job? Well, on average, it's About $150,000 US dollars, but the range is quite broad and goes up to twice that as a base salary plus equity and bonuses. So what are the skills or credentials required for the role? Well, most people come from either a machine learning background or a data science background, but it helps to know business too. It combines the skills of a machine learning engineer, a data scientist, a developer, and a product manager all in one.

AI coding agents are transforming software development, but are human programmers still necessary? Fortune's AI editor Jeremy Kahn breaks down what "vibe coding" really means, why companies are racing to adopt coding agents, and the hidden risks of AI-generated code. From technical debt to slopsquatting, we’ll explore what the future looks like for software programming – and whether humans will play a role. 

Are Programmers Obsolete? | Fortune AI Playbook

You might have been hearing a lot about AI coding assistance and how this might mean that programmers' jobs are going away. So what's really going on here? I'm Jeremy Kahn, your AI expert at Fortune. I'm here to help cut through some of the hype around AI and to answer your questions. So what are coding agents? Coding agents are AI models that can write computer code. You just tell them in English what you want the computer program to do, and it handles all the coding. Vibe coding involves somebody who is not a programmer just telling an AI model they want to build an app or a website. Site and the AI model does all the programming for them. The adoption of coding agents is happening very fast. Companies like Google are saying that a lot of the code they put out is now written by these coding agents and even more traditional companies, banks like JPMorgan, a lot of their code is also being written by coding agents too. So why does this matter? Number 1, it's redefining the role of the developer. Software programmers no longer have to worry about writing individual lines of code. Instead, their job is more about setting high level features and functions and then supervising and checking what the AI does. Number 2, coding agents lower the technical barrier to entry. Now anybody can write code. You don't have to work for the IT department. Number 3, AI agents can accumulate A lot of technical debt. Technical debt means when a company's legacy computer systems or IT systems cause it problems in the future. With these AI coding agents, they often write code that on the surface looks great, but it can be riddled with cybersecurity vulnerabilities and actually be brittle. Number 4, don't trust AI to remember everything. AI models suffer from a phenomenon called hallucinations. That's when AI models make up. erroneous information in the context of coding assistants, that hallucination can take the form of made up libraries or unit tests. These coding assistants sometimes make up those libraries or they go to libraries that they shouldn't go to, and that's led to this phenomenon called slob squatting. Slob squatting is when you have a bad actor who goes out and registers a library that they think an AI agent might visit, but instead of putting good code in that. Library. They put malware there. The coding agent goes and grabs the malware and imports it back into the network. Back to the original question are software programmers obsolete? Well, autonomous software development is pretty much here today. Engineers at companies like OpenAI and Anthropic say they've gone months without handwriting any code. They just supervise what the AI coding agents do. So what does this mean for people in other professions? Well, there are some people I think that the same kind of automation is coming for every knowledge worker's job in the next 12 to 18 months, but I think that's going to take a lot longer than that. In most other jobs, there aren't something equivalent to unit tests in software where you can automatically tell if the code is good or not. In most other professions, telling whether the work product is good or not requires subjective judgment and a lot of experience, and AI models aren't yet good at that.

AI Playbook

Can you trust AI agents? The 4 rules for safe deployment | Fortune AI Playbook

Latest episode

AI agents can operate across tools, make decisions, and even trigger unintended consequences. Fortune AI editor Jeremy Kahn examines why these systems introduce new risks, from unpredictable behavior to cascading failures, and what companies need to know to keep them under control.

Episodes (8 episodes)

Can you trust AI agents? The 4 rules for safe deployment | Fortune AI Playbook

The self-improving AI era: what are agentic loops? | Fortune AI Playbook

What is AGI, really? The AI milestone no one can define | Fortune AI Playbook

Is your data safe? The new AI cyber threat explained | Fortune AI Playbook

Hate the Dial Tone? Press 0 for AI Support! | Fortune AI Playbook

How AI Ads Are Making You Spend More | Fortune AI Playbook

Is the AI Architect the Next Hot Job? | Fortune AI Playbook

Are Programmers Obsolete? | Fortune AI Playbook