This roundup is pulled together from several weeks of articles I’ve read that I think are worth sharing. Read side by side, these nine pieces cluster around three questions: what AI actually costs to run at scale, where AI still needs humans in the loop, and what separates the organizations that capture real value from the ones that don’t.

The Economics of AI at Scale

A pattern is emerging: AI cost is no longer an afterthought, it’s a design constraint. Engineering leaders are actively routing work to different models based on task complexity (“modelmaxxing”), finance teams at major employers are throttling usage outright because bills have outpaced budget, and the more structural argument — that deployment, not the model itself, is where the real moat is — reframes the whole conversation. The takeaway: raw access to a frontier model buys very little on its own. What matters is the discipline built around how, when, and how much it gets used.

Tokenmaxxing is so over. It’s all about modelmaxxing now.

Business Insider — Posted on July 4

Twice a week, Morgan Linton tells his 16 engineers which AI models to use and when. Business Insider spoke to Linton, the Lake Tahoe-based chief technology officer of AI startup Bold Metrics, 50 minutes before his engineering team’s standup. He planned to tell one team to use Claude Fable on low, …

Companies Are Throttling Employees’ AI Use Because It’s Too Expensive

404 Media — Posted on July 2

Sources and leaks from Amazon, Adobe, Atlassian, Citi, and more show what is really happening with AI right now: companies are trying to rein in AI …

Models Aren’t the Moat. Deployment Is

HackerNoon — Posted on June 27

Notes from the field on what actually wins in enterprise AI. “The Guide notes that the most expensive failure mode of an enterprise AI company is, …

---

AI Doesn’t Replace the Human Layer

Each of these pieces pushes back on the idea that AI can simply be dropped in as a stand-in for people. Calling agents “coworkers” can dull the human instinct to check their work and blur who’s actually accountable. The fact that AI vendors keep having to embed real engineers with real customers to make deployments work undercuts the pitch of AI as an on-demand utility. And the well-worn statistic that 70% of transformations fail — a number that hasn’t budged in decades — is a reminder that the hard part of change was never really about the technology. Human judgment, oversight, and behavior are still the constraint, not the model’s capability.

AI agents are not your “coworkers”

MIT Technology Review — Posted on June 29

Marketing AI agents as digital employees may make human workers worse at spotting errors and more likely to offload accountability. Imagine coming in …

Why are big AI companies embedding engineers with customers, and what does that mean?

Fast Company — Posted on May 26

If intelligence were a true utility, you wouldn’t need to send people to every customer to make the faucet work. The promise of frontier AI has always sounded like a utility: abundant intelligence, available on demand, as easy to access as electricity, water, or cloud computing. The metaphor is …

We found the real reason 70% of transformations fail

Fortune — Posted on May 21

Corporate transformations fail much more often than they succeed. The failure rate is around 70% and this figure has not improved in decades. Since …

---

What It Actually Takes to Capture Value: Organizational Learning

Satya Nadella, CEO of Microsoft where I work (so naturally I’m attuned to his messaging), has some really interesting new framing. In the two pieces is the clearest articulation of an alternative to the “pick the best model” race: build learning loops — systems that get better every time they’re used — rather than betting everything on a single vendor’s model. It’s an argument that points to organisational capability, not a procurement decision or any specific technology or model.

Organizational knowledge might be making a comeback. The total collection of information, skills, experiences, and expertise held within a company. Key types include explicit knowledge, tacit knowledge, and embedded processes. It provides business value by improving performance and supporting daily operations. And it should not be ceded to AI model vendors but rather combined with AI to create new competitive advantage.

Microsoft CEO Satya Nadella On The Future Of AI: ‘Don’t Let The Models Eat Everything’

NDTV Profit — Posted on June 15

The last thing any of us want is a world where every company across every sector is ceding value to a few models that eat everything they see, …

New Memo: Microsoft’s Nadella Argues Learning Loops Beat Models

Forbes — Posted on June 15

Microsoft CEO Satya Nadella just articulated something most enterprises miss about their own AI future: the real competition isn’t which model you pick. It’s whether your organization learns from what it builds. Here’s the concept. A learning loop is a system that gets better every time it’s used, not …