AI Can Make More Decisions. That Doesn’t Mean It Should

Caroline Chumakov

June 10, 2026 • 5 min read

---

We recently co-hosted an event with Pfizer in Belgium exploring orchestration – making data-driven decisions (human or machine) at speed across​ the entire business ecosystem.​ It’s a continuous cycle of sensing, deciding, acting, and learning that integrates people, processes, technology, and data.  The conversation kept returning to a central tension: Most leaders no longer doubt that AI can take on more work. The harder question is where its authority should stop. 

That question crystallized for me a few weeks later when I read about Emergence World, a recent experiment from AI lab and enterprise tech company Emergence AI, which dropped identical groups of AI agents into parallel simulated societies and let them run for weeks with real-world signals, shared rules, and no resets. Some worlds stayed orderly. Others spiraled into crime, instability, and collapse. (This was due to multiple reasons, including how differently each AI model interpreted the environment and the impact of ecosystem contamination.) 

The experiment brings a dose of reality to some of the questions we debated at our Pfizer event, especially as we delved into human-machine teams and fusion teams. And while the context is different, the Emergence experiment offers a useful reminder: once agents are given autonomy, outcomes depend less on raw capability than on the boundaries around them. 

The Human-Machine Spectrum 

Zero100’s latest research points to a real shift in human-machine working already underway. By 2030, we predict that the center of gravity across supply chain work will move toward machine-led execution, especially in structured, repeatable, high-volume activities like contract lifecycle management.

And productivity gains are real. We surveyed sixty-five companies with more than $1 billion in revenue and found that 41% have already seen more than a 10% improvement in process cycle time and decision speed from AI, while 32% have seen equivalent gains in quality and accuracy. But knowing where that productivity will and won’t show up isn’t obvious – and it’s why we’ve long argued that the future of work tasks will exist across a human-machine spectrum:

But this spectrum introduces space for mistakes. If leaders assume every faster decision is unquestionably a better one, they risk delegating full authority where they should only be delegating a level of execution. Some decisions may look operational on the surface but be profoundly human underneath, involving ethical judgment, strategic tradeoffs, ambiguous context, or high-stakes consequences. 

Where the Language of Productivity Fails  

This consideration gets sharp fast. In the pharma industry, for example, inventory allocation is a functional, data-intensive task. In theory, it’s exactly the kind of work an autonomous system could do well, continuously rebalancing stock, weighing service levels, ingesting disruptions, and optimizing flows in real time. But inventory allocation in this sector isn’t just a planning problem. At times, it’s an access problem. It can shape who gets medicine first, which market waits longer, and how scarcity gets managed when supply does not meet need. 

That is not a workflow to automate. And this is where the language of productivity falls short. In cases like these, the right design requires a human in the loop and a clear division of labor: machines recommend, simulate, and execute within defined limits; people retain the authority to make the tradeoff call. 

The Boundary Is a Leadership Choice  

The most practical way to address the nuances of decision rights is to classify them by consequence, not feasibility: 

• Low-risk, reversible, rules-based decisions should be fully automated.  
• Material but bounded decisions should run with human review by exception. 
• High-stakes decisions involving safety, equity, compliance, or competing harms should stay with people.  

In a recent study of 1,800 workers, Microsoft found that when managers role model AI use, teams don’t just adopt AI faster; they use it more critically, with 22 percentage point increase in critical thinking about AI. That’s the foundation for sustainable AI integration: not blind automation, but informed delegation. 

Where Agentic Autonomy Should End  

The temptation in the agentic era will be to chase productivity to the edge. In many workflows, that will be the right call. But for decisions that shape safety, access, or fairness, purposeful delegation will matter more than technical performance.  This delegation will be the difference between upholding human responsibility and the consequences of leaving ethics to machines. 

Leaders who get this right won’t be the ones who automate the most. They’ll be the ones who know where agentic autonomy ends and who train their people to recognize and enforce those boundaries.