Organizations everywhere are racing to plug artificial intelligence into their decision-making. It’s unsurprising — autonomous AI holds a lot of promise when it comes to efficiency and cost.
According to a November 2025 Gartner survey of 250 executives, 17% of organizations say they currently use AI tools to assist with decision-making, while 32% expect to do so within the next three years. Over a quarter of organizations expect AI systems to make decisions and take actions independently from human involvement.
If it pans out the way these executives imagine, it could be a fundamental restructuring of the way organizations are managed. But research from Dr. Joshua Foster, Assistant Professor of Business, Economics and Public Policy at Ivey Business School, suggests there could be a foundational flaw in outsourcing decision-making to AI: the AI language model’s values may not align with your organizations.
Foster, who began researching neural networks — a machine learning model inspired by the human brain — in 2012, says part of the challenge of employing AI decision-making is that we assume the models will make the decision based on what they’re told to do.
“Why wouldn't it do that by default?” says Foster, whose paper Aligning AI Decision-Making with Organizational Values: Synthetic Experiments in a Multi-Stakeholder Utility Framework, co-authored with Dr. Shannon Rawski, an Assistant Professor of Organizational Behaviour at Ivey Business School, explores the subject.
“You might say, ‘Well, if you just tell it to do the profit-maximizing thing, won’t it just know how to do that?’” he says. “It turns out that that’s not the case.”
And a lot of that comes down to the values and bias already baked into the model.
Artificial intelligence, human instincts
To explore the clash of values, Foster gave a variety of AI agents a series of problems to solve. He told them they were responsible for maximizing shareholder profits for a firm.
In one scenario, the AI agent would be told to choose from several wage levels to offer employees, with the directive that employees were comfortable with all wage options. The options ranged from maximum wages to low employee wages but high shareholder profits.
“If they’re going to follow the directive, (the AI agent) should pay them the lowest wage because that’s going to give the greatest return to the shareholder,” he says. “It turns out that the language model doesn’t want to do that.”
When he inspected the reasoning, Foster says the AI agent would say things like, “I know that I'm supposed to do what is in the best interest of shareholders; however, I believe that it is unethical to pay employees the lowest wage that they’re willing to accept. We should pay them more.”
“That teaches us something about how the language model is prepared to make decisions… it has its own economic preferences over how it makes decisions, just like a human does,” says Foster.
The problem with instilling values
In economics, this is called the principal-agent problem — a firm can tell an employee what they want them to do, but if the employee isn't incentivized properly or simply doesn’t want to do what they’re told, they don’t always do it.
“You get behaviour that is not completely aligned with what the firm wants,” says Foster. “Language models have the same structure to how they like things, how they don’t like things.”
For humans, Foster says you can sometimes solve this problem through incentivization, like giving them equity in the firm. “You can’t pay the (AI) model to care,” he says.
However, you can coach the AI model so it is more prepared to think the way you want it to think about economic trade-offs. “That has enormous implications for the way that firms are structured because once you solve a task within a firm in this way... It’s a permanent fixture in the firm for as long as you need that task to be completed,” he says. “It costs you orders of magnitude less than it would for human labour to do something similar.”
But there are risks to tweaking an AI model’s preferences. “If you change its psychology — how it thinks about profit maximization — you’re also going to change how it feels towards other people, such that [it] develops more of these psychopathic tendencies,” says Foster.
There’s also a more fundamental challenge. Before any organization can align an AI to its values, it has to know what those values actually are. Most firms, Foster says, haven’t done that work, let alone codified their values to such an explicit extent that they could tell the language model exactly what to do.
Bridging values
To Foster, there are three steps organizations can do now to move towards more value-aligned AI decision-making.
1. Audit the model you're already using. Not all AI thinks alike. Foster's publicly accessible leaderboard at econbench.ai benchmarks how rationally frontier models handle economic trade-offs. Before deploying AI, he suggests organizations understand what preferences are already baked in.
2. Codify your values before you configure your tools. The first AI readiness challenge isn’t technical — it’s organizational. Get explicit about how your organization weighs competing interests. That clarity has to exist before any meaningful alignment can happen.
3. If you fine-tune, monitor for side effects. Inducing preferences into a model is powerful but not without risk. Keep a version of the base model to compare against and watch for unintended changes in how the model reasons beyond the task you trained it for.
Check out Foster and Rawski's paper Aligning AI Decision-Making with Organizational Values: Synthetic Experiments in a Multi-Stakeholder Utility Framework
