AI tokenomics is a way of thinking about the cost of AI work in terms of tokens (the units AI models use to process information), rather than traditional software licenses or IT line items.

Instead of asking, “What does this AI tool cost as software?”, tokenomics asks, “What does it cost for an AI agent to do this task compared with a human?” That shifts the comparison from IT budget to labor cost, quality, and time.

Key points for leaders:

• New allocation decision: For many tasks, you now have to decide, “Should a human do this, or should an AI agent?” That decision weighs quality, speed, and cost.
• Costs are moving targets: The article notes that what a task costs to run today won’t be what it costs next quarter. As models get better and more efficient, token costs change quickly.
• Tokens are like headcount: Leaders will increasingly manage tokens with the same deliberateness as headcount—deciding who gets tokens, how many, and for what work, and holding teams accountable for outcomes.
• Need for new infrastructure: Most organizations don’t yet have the systems to track and optimize token use. Those that build this capability early and recalibrate as economics shift will have an advantage over those treating token costs as static.

In practice, tokenomics means treating AI capacity as a strategic resource, not just another software expense—and revisiting those decisions frequently as costs and capabilities evolve.

Trust in AI isn’t about believing in “AI” in general; it’s about trusting a specific system to do a specific job.

The article highlights a useful contrast:

• High-trust example: An AI system at Johns Hopkins trained on a single dataset and focused only on reducing unnecessary breast cancer biopsies. It doesn’t write poems or give directions; its scope is narrow and clear, which makes it easier to trust.
• Low-trust example: An AI agent buried in a rental car customer service flow that could be prompted to write code. The system had no clear edges, so users had no basis for trusting what it might do.

From these examples, three conditions for trust emerge:

• 1. Consistent performance: The AI needs to reliably do the job it’s been assigned, not just perform well in demos.
• 2. Working understanding: People don’t need to be ML experts, but they do need a basic understanding of how the system works and what it’s for.
• 3. Accountability: There must be clear ownership and consequences when something goes wrong. The article compares this to commercial aviation, where public investigation and accountability help people get back on planes after a crash.

As AI agents move from generating content to taking real actions, leaders need to design the accountability infrastructure before failures happen, not after. That means:

• Defining where AI is allowed to act and where humans must approve.
• Making it clear who owns the outcomes of AI-assisted work.
• Documenting and reviewing incidents, then adjusting systems and policies.

In short, trust comes from clear scope, consistent performance, and visible accountability, not from broad optimism about AI.

The article makes a clear point: AI returns are driven by leadership decisions, not by the tools themselves. Simply “turning on” AI rarely changes outcomes.

One example comes from Microsoft’s own experience. When the company first rolled Copilot out to its sales force, adoption metrics improved, but business outcomes did not. The issue wasn’t the technology—it was the approach:

• The rollout was treated like a standard product launch, focused on access and usage.
• The underlying sales work itself wasn’t redesigned to take advantage of what AI could do.

The article argues that knowledge work has been largely ad hoc for decades, even in structured fields like legal, finance, and accounting. Workflows are often vague, outputs inconsistent, and goals live in people’s heads rather than in systems.

To get real value from AI, leaders need to:

• Redesign the work, not just add a tool: Break work into measurable steps, decide which steps are best for humans vs. agents, and track outcomes instead of activities.
• Treat AI as a system, not a single model: The article notes that early on, organizations treated “which model to use” as the main decision. Now, value comes from the entire system around the model—data access, context, infrastructure, and how it all fits together.
• Hold enterprise tools to consumer-grade expectations: Employees now use strong AI tools in their personal lives and bring those expectations to work. As one leader put it, “the era of ‘I use this kind of crappy thing because I’m forced to’ is kind of over.”

In practice, if AI isn’t moving the needle, it’s a signal to:

• Revisit how the work is structured, not just how many people have access.
• Clarify what success looks like in terms of outcomes (e.g., fewer manual steps, faster cycle times, better decisions).
• Invest in the system design around the model—data, workflows, guardrails, and accountability.

The organizations pulling ahead are the ones that reimagine their work and systems around AI, instead of treating AI as a plug-in feature.