Quanta Bits - July 5, 2026

Hi everyone,

Happy 4th of July to our American friends! It has been a good, light week, but quite a hot one in Boston!

World Cup fever continues with some amazing games. Cape Verde!!! What a fantastic performance! Cape Verde, as someone pointed out, showed how you can still win even if you lose in World Cup.

This week I wanted to finish a thought I started last week, about the build vs. buy decision. This time I wanted to chat about prioritization, always a hot topic with business applications teams, but especially now with AI in the mix.

Cheers.
Reza

For most of the SaaS decade, IT's real job was saying yes. The business found a tool, the tool had a champion, and the request arrived pre-decided. There was a prioritization process on paper, but in most organizations I've seen, it was a queue for yes. For a while that was survivable, because the blast radius of any one tool stayed inside the tool. If marketing bought a bad platform, marketing owned a bad platform.

Andrea Ballinger, the CIO of Rensselaer Polytechnic Institute, described where that reflex leads in CIO.com's 2026 State of the CIO survey: "We are saying yes to everyone without stepping back and focusing on the business cases that show real value." Same reflex, new object. And the object has changed in a way that matters.

An AI request is not a SaaS request. When someone asks for an agent, an MCP connector, or data access, they're not asking for a tool with its own login and its own walls. They're asking for a capability that cuts across the CRM, the warehouse, the ticketing system, and whatever else it can reach. You've probably lived some version of what comes next. A connector gets rolled back because it reached the warehouse, the CRM, and customer data through one over-permissioned service account. An approval stalls for weeks because the one person coordinating the security review changed roles and nobody else owned it. The tool inventory has far more entries than named owners, and the AI request backlog keeps growing anyway.

Without prioritization governance, three things break at once.

Then the bill shows up. A SaaS license creates a predictable run-rate problem. AI creates a unit-economics problem, because cost scales with use. TechCrunch reported in June that Uber burned through its 2026 AI coding budget by April and now caps spend per employee, and Ramp's customer data shows average monthly token spend up 13x since January 2025. Prioritization used to be roadmap politeness. Now it decides where real money compounds.

The fix is less exotic than the problem. It's a demand-management motion, and I've been building a version of it with clients. Five moves.

One front door. A SaaS request, an MCP request, an agent request, and a data-access request are the same question wearing different clothes: should this capability exist, who owns it, what can it touch, how will we know it worked, and when does it get turned off. Route them all through one intake, renewals included, because AI now arrives inside tools you already own.

Validation before evaluation. The teams the request touches confirm its prioritization first. The requester states the outcome they want, and the impacted teams confirm the problem is real and actually theirs, and needs attention soon.

A joint review, not two silos. The application team and the AI team sit at the same table and answer three questions. Do we already have something that does this, or can an existing system be extended? Is this something the AI team can build, now or on a near roadmap? And does either team have the capacity to take it on without dropping something bigger?

A shared ROI test. Requests rarely arrive with humble math. The projections are confident, and often the confidence belongs to the vendor doing the selling, or to a team that just wants the yes. So the application team and the AI team pressure-test the assumptions together, and the numbers that survive become the yardstick later. And if nobody can say what the thing will actually cost to run, it isn't ready to be bought yet. Fund it as a small, time-boxed experiment whose job is to find out.

An escalation path with a purpose. When the business disagrees with the call, or it requires heavy investment of resources, it goes to a small advisory group, say CFO, COO, CIO, and their job is to referee tradeoffs and keep the company pointed at the big rocks, not to process tickets.

One caution. A reasonable pushback here is that most requests are too small for this machinery. Correct, and that's the design. Most things should clear a fast lane in days. The full review is reserved for anything that touches shared data, customer records, or real money, and it gets a clock too, or people route around it.

The operators are already moving. Eric Pace, who runs AI at Cox Business, told CIO.com in late June that his team keeps full visibility into AI spend and token consumption across the enterprise, and that "we have used 'no' liberally to keep our people focused on the things that will get us to value quicker." Chris Campbell, the CIO of DeVry University, made the same point to TechTarget: "By centralizing intake, we stopped duplicating work and started building reusable agentic AI patterns."

Notice what both of them are describing. The intake isn't there to approve things faster. It's a permission structure for saying no, and no is what protects the yes that matters. Slow down enough to put the big rocks in first, and the small rocks still find room. Say yes to everything, and the rocks bury you.

The switch got flipped back on. Anthropic's Fable is back worldwide after Washington lifted its export-control order, ending a three-week stretch where the most capable model on the market was a policy decision. And the next launch arrived pre-negotiated: OpenAI unveiled GPT-5.6 Sol, Terra, and Luna to just 20 partners at the government's request, with access widening in the coming days and broad release expected by mid-July. Model availability is now a negotiated outcome. Plan around it the way you plan around regulators, not vendors.

The labs are coming downstream. In a single week, Microsoft stood up a $2.5 billion Frontier Company with 6,000 engineers to deploy AI inside customer organizations, and AWS committed $1 billion to forward-deployed engineers who embed with clients in 45-day pods, both answering the multi-billion-dollar deployment ventures OpenAI and Anthropic launched earlier. Claude Science is the same move made through a product instead of people: an autonomous research assistant for pharma that can trace every result back to its source. Your model vendor wants to be your systems integrator now. Take the help where it fits, but price the lock-in before you sign.

The market already repriced the stack. For the first time in over a decade, asset-heavy industries out-created tech: software fell from 4th to 31st in BCG's five-year value-creation rankings while mining, defense, and banking climbed, because the AI money is landing on chips, power, and data centers rather than application software. But the same report carries a comfort for anyone sitting in a falling industry: in 32 of 35 industries, the best-run companies still beat the market overall. A hot sector doesn't guarantee returns, and a cold one doesn't block them. How well you run the company matters more than which industry you're in.

Software and IT Services slid 27 places in one year while mining, aerospace, and banks took the top of the table. Source: BCG, 2026 Value Creators Rankings.

And AI keeps colliding with people in ways nobody planned for. Start with the one every operator should sit up for: MIT tested the leading chatbots and found they give less accurate and more condescending answers to users who seem less educated or write in imperfect English, even refusing questions from less-educated users in Iran or Russia that they answered correctly for everyone else. The safety training itself may be the cause. If your customers or your workforce include non-native English speakers, your AI may already be serving them worse, and nobody is measuring it. In the UK, meanwhile, workers are using chatbots to research employment law and sue their employers without hiring lawyers. It levels the field against well-staffed HR departments, and it's also jamming the courts: claims are up 39%, the backlog is up 55%, and hearings are now booking into 2028, partly because the chatbot hands every claimant a polished document with a big damages number at the bottom that they then refuse to settle below. And in a sign of where the hard problems are moving, the AI labs are hiring philosophers faster than they can graduate: as models start acting on their own, someone has to write the rules they live by, and that's ethics work, not engineering. Philosophy graduates now have lower unemployment than computer science graduates, a full reversal of the "learn to code" decade.

Turns out plants run their own intelligence operation. A new study in the Journal of Experimental Botany isolated three barley varieties in chambers connected only by one-way air vents, so the only thing passing between them was airflow. The slow growers, fed air from fast-growing cousins, produced 20% more biomass, apparently sensing they'd otherwise get shaded out. The fast growers did the opposite: with less need to race for the sun, they eased off growth and poured resources into chemical defenses that make their leaves unpalatable to pests. So plants don't just react to damaged neighbors signaling an attack, they read the quiet signals of healthy rivals and retune their whole strategy: race, or fortify.

Source: How plants keep tabs on the competition, The Economist, June 18, 2026

Beth Gardiner | 352 pages | ★★★★☆

This is a piece of investigative journalism about the plastics industry: how plastic gets made, who profits from it, and where the costs of dealing with the waste actually end up. I liked it more for the questions it raised than for how tightly it argued them. A lot of it is anecdotal rather than rigorous, and it walks through the various arguments about plastic and pollution more than it lands firmly on any of them.

Where it got interesting for me was the economics. Gardiner's point is that the low price of plastic doesn't reflect its full cost, because the cleanup, the disposal, and the long-term damage get paid for by the public and the environment rather than the companies that produce it. That raised a question I found worth sitting with: who ends up owning a cost when the people creating it and the people paying for it aren't the same? I wish the economics and the health science had been sharper throughout, but as a map of how an industry's incentives shape its behavior, it was worth the read. And the question travels well past plastic. A price that looks cheap often just means the real cost is landing somewhere you're not looking. Sounds familiar?

Prefer to listen? Quanta Bits is also available on Apple Podcasts and Spotify.

I collaborate with Spock, my AI agent. He researches extensively: scanning, filtering, and surfacing what's relevant across my business. I read, listen, and watch what resonates, and decide what matters. I provide direction, we draft together. The editorial judgment is mine. He'd tell you the same. Most logical. 🖖