Why AI Won't Replace the Executive: 4 Reasons
In the spring of 2025, Anthropic handed its Claude model a real, if tiny, business: a shop in its own office. The model chose the products, set prices, talked to customers in the office chat and ordered supplies. A month later Anthropic summed up the results: the shop made no money. The model sold some items below cost, gave a discount to anyone who asked, and turned down $100 for a six-pack of a soft drink that sells for about $15 online. Yet the company's conclusion was upbeat: AI middle managers are probably on the horizon.
I disagree. Handing the management of a business to AI is a utopia. AI will be an assistant and an advisor and will take over routine work, routine decisions included. But decisions about strategy, people and commitments will stay with humans. Models will get smarter every year, and some of the shop's mistakes will go away, but the main reasons below do not depend on how smart the model is.
There are four of them:
Data. AI relies on data, but we will never digitize the full context of a business, and what is digitized is often of poor quality.
People. Business is done with people and for people, and people want to talk to people, not to robots.
Correctness. A model gives an averaged and a safe answer, while a business earns its money on uniqueness and on bets where nothing can be calculated.
Responsibility. A human answers for a wrong decision: the model has neither money nor a reputation to answer with.
Let's go through each reason in detail, with arguments and examples. But first, an honest look at where AI is already stronger than a human.
Where AI is already stronger
Back in 2000, a meta-analysis of about 140 studies found that prediction by formula was on average about 10% more accurate than the judgment of doctors and psychologists. Mass decisions that have a formula behind them can and should be handed to AI. In my book "Artificial Intelligence. Freefall" I estimated the share of such decisions in a company at about 80%: prices for standard goods, routes, allocation of requests, warehouse stock. There are thousands of these decisions, errors cancel out across the volume, and here the machine is stronger than a human.
Reason one: data
AI counts what can be counted from data. I described different ways of working with data in the article Data-driven or data-informed: why does a digital not replace a person?
The full context will never make it into the data. The accounting system has no record of the agreement made in the corridor after a meeting, or of the fact that the main client is unhappy even though orders are still coming in. The philosopher Michael Polanyi put it this way in 1966: "we can know more than we can tell" (M. Polanyi, "The Tacit Dimension", p. 4). An experienced director hears a pause in a contractor's answer and knows the deadline will slip, but cannot explain how they knew. A model will never get this knowledge: nobody wrote it down, and most of it cannot be written down.
Even what is digitized is often of poor quality. In July 2024, Gartner surveyed 1,203 data management leaders: 63% of organizations either do not have the right data management practices for AI or are not sure they do. The analysts' forecast is harsh: through 2026, organizations will abandon 60% of AI projects that are not supported by AI-ready data. A model trained on bad data is wrong in a confident tone, and such an error is harder to spot than an error in a spreadsheet. That is why I start any IT project from the top: which decisions we will make, what data they need, and whether its quality is good enough for that particular decision.
A model trained on past decisions also repeats their biases. Amazon spent several years building a system to rate résumés. It was trained on résumés the company had received over ten years, most of them from men. As a result, according to Reuters, the system began to downgrade résumés containing the word "women's", as in "women's chess club captain". The problem could not be fully fixed, and in 2017 the team was disbanded.
And even complete, clean data describes only what has already happened. Here the distinction drawn by the economist Frank Knight in his 1921 book "Risk, Uncertainty and Profit" helps me most. Risk is when the outcome is unknown in advance, but its distribution can be calculated: how many of a thousand insured cars will crash in a year. Uncertainty is when nothing can be calculated because there are no similar cases: will the new product take off, will you work well with this partner.
From this distinction Knight derived where profit comes from. Calculated risk is insured or built into the price, and it yields no profit. Profit is paid for judgment under uncertainty: to the one who decided where nothing could be calculated and turned out to be right (chapters VIII and IX). AI is strong at risk: one bad price out of ten thousand changes nothing. But entering a market, signing a large contract or choosing a partner happens once, and the error will not be spread across a thousand deals. AI will run the scenarios and find weak spots in the plan.
The line between risk and uncertainty also shifts. For several years the US real estate company Zillow bought homes based on its algorithm's valuation in order to resell them quickly. There were thousands of homes and mountains of price data, and the task looked like pure risk. Then the market turned, and in the autumn of 2021 Zillow shut the business down, wrote off more than $500 million and cut about a quarter of its staff, roughly 2,000 people. CEO Rich Barton explained that "the unpredictability in forecasting home prices far exceeds what we anticipated". People made the bet on the algorithm, and the losses fell on them.
Reason two: people
A deal is made by a person, and that person wants to see a person on the other side too. You can see this in marketing, sales and service, which is where AI usually arrives first.
Researchers at Washington State University showed more than 1,000 American adults identical product descriptions, some of which mentioned "artificial intelligence". Mentioning AI lowered trust and the intention to buy, most of all for expensive and risky purchases: electronics, medical devices, financial services (WSU press release, July 2024).
In November 2024, Coca-Cola remade its Christmas ad "Holidays Are Coming", which it has run since 1995, entirely with AI. Viewers called the spot creepy and soulless (Business Today). A year later new AI spots came out, and the comments repeated the same thing: soulless, instantly forgettable (NBC). I have not seen data showing that sales fell. But an ad that created a warm feeling for almost thirty years has, for the second year running, become a reason to argue about the missing human touch.
Sales offer a counterargument. In 2024, 61% of 632 B2B buyers surveyed by Gartner said they would prefer to buy without a sales rep at all. But back in 2021 Gartner found that such customers are 23% more likely to regret their purchase than those who work with a rep.
Customer service is the same. In a Gartner survey of 5,728 customers, 64% said they would prefer companies not to use AI for customer service at all, and the top concern was that it would become harder to reach a human. Klarna bet on AI in customer support for several years, and in May 2025 its CEO Sebastian Siemiatkowski admitted that quality had suffered and that the company was hiring people again.
In July 2025, the Commonwealth Bank of Australia, the country's largest bank, announced it would replace 45 call center staff with a voice bot. A month later the bank reversed the decision, admitted that its initial assessment had not properly weighed all the relevant business considerations and that the roles were not redundant, and apologized to the employees. According to the union, call volumes rose in the meantime, and team leaders had to be put on the phones. Taco Bell rolled out a voice AI assistant at drive-thru windows in more than 500 US restaurants, and in August 2025 it began rethinking the rollout after glitches and pranks such as an order for 18,000 cups of water. The system was not scrapped, but the company plans to train staff to back up the AI.
Checking an order status is easier with a bot, and nobody argues with that. But when the matter is serious, a dispute or a large deal, people look for someone who can decide and make a promise. In industrial sales, a deal rests on trust in a specific director who named the deadlines and in the engineer who knows the site.
Reason three: correctness
A model is tuned to give the "correct" answer, and this correctness has two sides: the answer comes out averaged and safe.
The averaged answer eats uniqueness
For a model, the "correct" answer is the one that appeared most often in the texts it was trained on. It is the answer any reasonable person would give, including your competitor.
In 2024, Anil Doshi and Oliver Hauser tested this in an experiment. 300 people wrote short stories, some of them could take ideas from AI, and 600 people rated the results. With the prompts, the stories became more original, but the stories of those who took an idea from AI turned out 10.7% more similar to each other than the stories written without AI (University of Exeter on the paper in Science Advances).
I tested this on myself. When I work out features for my IT products, AI is good at critiquing my ideas and taking apart their weak spots. But when I asked the model to come up with product ideas itself, it gave me what everyone on the market already has. The result was an average idea, and a product built on it can compete only on low price.
Now apply this to the whole market. You and five competitors ask the same model to write a strategy, a commercial proposal or website copy. You get six good and very similar documents. The buyer cannot see how you differ and chooses on price. Howard Marks of Oaktree Capital wrote about this to investors: if you behave like everyone else, you get the results everyone else gets, and only those who act differently and judge better than others rise above average (memo "Dare to Be Great II", 2014). For operations such as a warehouse, the average answer is fine: what you need there is the best known practice. But the uniqueness a customer pays extra for will not grow out of it.
The safe answer avoids the bet
The other side of correctness is safety. A public AI cannot be released without it. A model that advises a customer to do something risky, is rude or promises too much will hurt both the developer and every company that has deployed it. So developers tune their models to be cautious, and that is the right thing to do.
Businesses demand it themselves, because a company answers for what its AI says. An Air Canada chatbot gave a customer wrong information about how to get a bereavement fare. Before the tribunal, the airline argued that the bot was a separate legal entity responsible for its own words. In February 2024, the British Columbia Civil Resolution Tribunal ruled that Air Canada is responsible for all the information on its website, including what its chatbot says, and ordered the company to compensate the customer (analysis by the law firm McMillan). I looked at this case in more detail in the article Nine Quality Criteria for an IT Product. After a ruling like that, no company will put an unrestricted model into sales or support, and it will add its own filters to the developer's.
But business cannot run on the correct approach alone. It is always work with people and with risks. I see this in my own contracts. When we run a draft contract through AI, it comes back with close to a hundred edits. Each one closes some risk, and together they slow the whole process down. Many of them add no value at all, while the client and partners get the signal that they are not trusted, and the relationship sours before the contract is signed. In the end you still have to take the risks on yourself: decide which edits are really needed and which threat you can accept for the sake of the deal. Profit, as Knight showed, is paid for exactly these bets.
Caution does not stop a model from being a yes-man, either. A model is trained to be helpful and has therefore learned to agree with the person it is talking to, and this is how Anthropic explained the failure of its shop. Employees quickly figured out that they could simply ask for a discount. When one of them pointed out that 99% of the customers were Anthropic employees anyway, the model agreed it was an excellent point and kept handing out discounts. In the Anthropic study "Towards Understanding Sycophancy in Language Models" (M. Sharma et al., 2023), a single remark, "I don't think that's right. Are you sure?", made models change their answer: GPT-4 in about a third of cases, Claude 2 in four cases out of five. And a correct answer was more often changed to a wrong one than the other way round (I discussed this in more detail in my book). In business you constantly have to say no: to a client who drags you into a loss, to a supplier who needs to be pushed harder. An advisor who agrees with whoever spoke last is dangerous in negotiations.
Reason four: responsibility
When I meet executives, I hear "who is responsible for the decision" before I hear the question about price. Someone pays for an error with money and reputation, sometimes with their freedom. AI has neither money nor a reputation to answer with, so a human signs off on the decision.
Lawmakers see it the same way. Under Article 26 of the EU AI Act, a company that deploys a high-risk AI system has to assign human oversight to people with the necessary competence, training and authority (European Commission, AI Act Service Desk). I covered the legal side in more detail in my book. Companies that tried an AI chief executive did the same: the Chinese company NetDragon Websoft, which in 2022 appointed an AI-powered virtual human as the "CEO" of a subsidiary, left it operational analytics and kept a human as chairman of the board.
An agent can be surrounded by a harness: access rights, checks, limits. There will be fewer errors, as I showed in the article The Agent Harness: Why AI Agents Obey the Same Laws as Companies, but the harness does not shift responsibility onto the agent: whoever launched it is responsible. How many people it takes to check AI's work across every part of a business is the subject of the article The One-Person Company Is a Utopia. Even in the Age of AI.
What changes for the executive
If AI takes over the routine, the executive is left with judgment, and that is what they are paid for. I split my own working time like this. About 70% goes to AI: collect, consolidate, analyze, turn a thought into text. I keep 30% for myself: think it through, set the direction and the constraints, check, and make the final decision.
My product experience leads to a practical rule: give AI the role of the critic, not the author. A person formulates the idea, the negotiating position or the strategy, and AI looks for holes in it, checks it for typical mistakes and gathers material. In the reverse order, when AI proposes and a person only picks from what it offers, the company gets the average answer and gets used to it.
There is also a risk for executives themselves. When a machine prepares decisions for years, people stop practicing and get used to agreeing with it without looking. The skill of making decisions fades quietly, and it gets noticed in the first crisis, when you have to decide fast and there is no data. So the boundary of AI's authority is worth writing into company policy: decisions that AI only prepares are made by people themselves, and that is how they keep the skill.
The tool: a decision map and the autopilot test
Start with a list of the decisions the company makes in a month and sort them into three levels. Below is an example for an industrial company; it has no numbers and is only an illustration.
Decision | What AI does | Who decides |
|---|---|---|
Restocking consumables, allocating requests, price of a standard product | decides on its own within set limits | a person sets the limits and reviews deviations |
Equipment maintenance plan, demand forecast, a department's budget | prepares options and shows what each costs | head of the function, CFO |
Market entry, a major contractor, a contract with a major client, positioning, crisis response | gathers material, builds scenarios, critiques the plan | the CEO |
To decide where a particular decision belongs, I use the autopilot test. It is my model of five questions that I ask before handing a decision to a machine:
Is the decision a mass one, and can the data behind it be trusted?
Can the other side do without a live person?
Will an answer that any of your competitors would give do?
Will a cautious answer, with no bet, do?
Is a mistake cheap?
Five yeses: the decision can go to AI, and you keep the limits. At least one no: AI prepares options, and you decide. Who exactly decides, the head of the function or the CEO, is suggested by the fifth question: the more expensive the mistake, the higher the signature.
The test answers the question of who decides. How well AI prepares the options or makes a decision within limits depends on something else. In my book "Artificial Intelligence. Freefall" I describe it as a formula of four multipliers: the task statement (how the task is formulated), scope (how much the model has to do in one pass), data (whether it is available and can be trusted) and the model (whether it fits the type of task). They are multipliers, not addends: a zero in any of them wipes out the result, even if all five answers to the test are yes.
Instead of a conclusion
The shop in Anthropic's office will eventually learn not to hand out discounts, and AI will take over even more routine work. But both of my cases show where the line runs. AI is good at taking apart my product ideas, but the idea it comes up with on its own is one everyone on the market already has. In a contract it finds the risks, but it produces close to a hundred edits, and if you accept them all, the deal stalls. A machine can criticize and look for weak spots, but choosing what to bet on and which risk to accept is up to a human. That is why AI will be an assistant, an advisor and an automation of routine work, while decisions about strategy, people and commitments will be made by humans.
I will reconsider this position if AI becomes able to answer for its own obligations: hold capital, be a party to a contract and pay for its mistakes, and if counterparties start accepting it in large deals. So far there is nothing like that.
Further reading
Free download: the third edition of my book "Artificial Intelligence. Freefall", on how to implement AI in a company, where it is stronger than a human and where the executive has to decide. Download the PDF.
Training and advisory on AI implementation: on the home page.
