The Pope, Starbucks, and the AI bill that lands on the platform team

Pope Leo XIV published a 42,000-word encyclical on AI yesterday and invited Anthropic's Chris Olah to the Vatican. The same week, Starbucks quietly retired an AI tool it had run across 11,000 stores for nine months because it couldn't tell oat milk from whole milk. Today's issue is the read on what both stories have in common, and why the bill from your AI deployments is landing somewhere you're probably not looking.

In today’s issue:

• Main story: The Pope just made your AI governance a board problem

• AI Headlines since Friday

Yesterday, Pope Leo XIV published a 42,000-word encyclical on AI titled Magnifica humanitas, addressed to the Catholic Church's 1.4 billion members. He warned that "no algorithm can make war morally acceptable." He called for AI to be "disarmed" before it dominates humanity. He asked governments to stop handing oversight over to private labs. He invited Chris Olah of Anthropic to the Vatican to speak alongside him.

That is the cultural event of the week. The operating event nobody is writing about will hit your calendar within a quarter.

Setup

While the Pope was being briefed by Olah, Emma, who leads data infrastructure engineering at OpenAI, was explaining to Nate Jones what her team has been quietly absorbing. An agent ran a routine job and took down a Kafka cluster. The same week, another agent debugged a deep system bug and shipped a training export overnight while a human slept. Both stories belong together because they show the same shift. Agents can now do real operational work, and they can create real operational risk in the same week. Both effects land on the platform team, not the application team.

The same week, Starbucks quietly retired an AI inventory tool that CEO Brian Niccol had deployed across 11,000-plus North American stores nine months ago. The model could not reliably distinguish oat milk from whole milk. An internal newsletter announced that milk would now be counted "the same way you count other inventory categories." With human eyes and a clipboard.

The turn

The headline of AI adoption in 2026 is "application teams got 10x faster." That is true. Cursor hit $3 billion in annualized revenue in two years. Dan Shipper at Every is running a 30-person company where every role is AI-augmented, and output is multiples of what it used to be.

The version nobody is writing the headline about is the one Nate Jones described this week. When app teams accelerate, the work does not disappear. It moves downstream. It lands on the platform team, the security team, the data team, the compliance team, and the legal team. Those teams have the same headcount they had two years ago. The board approved the AI budget as a win for one part of the company. That part wins. The rest of the company picks up the workload that comes with it.

So the harder question for operators in 2026 is not "should we use AI?" That question is settled. The question is who is absorbing the operational debt each deployment creates, and whether anyone budgeted for it.

Where the bill lands

The application team owns the velocity story. The platform team owns the recovery story. When a coding agent ships a feature in three days that used to take two weeks, leadership tells the board, "we are 10x faster now." What goes unreported is what happens next. The platform team spends three days figuring out why a deploy hit a database query path no human would have written. Then, two more days adding guardrails. Net company velocity is up while net platform team capacity is down, and the board sees only the first number until a Starbucks-shaped incident makes the second one visible.

Agents create work that looks invisible until it isn't. For nine months at Starbucks, every store generated wrong dairy counts. Every wrong count became someone's reconciliation problem at the regional level. That cost did not show up on the AI pilot's dashboard. It showed up on the operations team overtime. The retirement of the tool is the moment when the cost becomes unavoidable. Until then, the company was paying for it without knowing.

The "AI takes jobs" narrative is masking the "AI requires more jobs" reality. Dario Amodei has said AI may eliminate half of entry-level white-collar jobs. The companies actually deploying AI at scale are walking away with the opposite problem. Every is hiring, not firing. Anthropic and OpenAI are hiring forward-deployed engineers as fast as they can find them. The reason is plain. AI makes expert skills cheap, which raises demand for experts to validate, integrate, and clean up what the agents ship. Companies that planned headcount on the idea that AI replaces people are now understaffed for the work AI has created.

The Pope's intervention is a board-meeting story, not a religious one. Leo XIV named the lever: "robust legal frameworks, independent oversight, informed users, and a political system that does not abdicate its responsibility." Olah's reply was blunt. Every frontier lab, his own included, sits inside incentives that can conflict with doing the right thing. Two weeks from now, when your board asks about AI governance, the conversation has changed. The question is no longer what your AI committee does. It is who outside the committee is allowed to say no, and what authority they have when they do.

AI READY PRO · FREE

A 30-day AI program. In your inbox. Then it ends.

One email every morning at 7 am ET. Each one is a short read and one real thing to try before lunch. By Day 30, you have nine concrete capabilities, including prompting that hits your bar on the first draft, AI workflows that take five hours a week off your plate, and the language to lead the AI conversation at your company.

Built for ops leaders, COOs, chiefs of staff, founders, and team leads at mid-market companies. The four weeks move you from foundations (effective prompting, catching hallucinations, learning AI with AI) to connect (a personal context layer, daily automation, five hours a week back), to build (reusable skills, agent design and debugging), to lead (tool and model evaluation, leading the conversation at work). Twenty-one mornings, no upsells, no community Slack.

Counterargument

The strongest objection is that platform debt is temporary. Agents will get good enough to self-heal, self-deploy, and self-secure. The platform team's job is being automated alongside the application team's. The current gap is a one-year gap, not a structural one.

I want to be honest about the bear case for my own argument. Agents are improving at observability and automatic remediation. The gap will narrow. But two things have to be true for the bear case to win, and neither is true yet. The agent has to reliably know when it has caused a problem. The company has to trust an agent to mark its own homework on a Kafka outage at 2am. The Starbucks story tells you how long the lag can run: nine months, 11,000 stores, one model that could not tell oat milk from whole milk. In every year of waiting, the company that under-funded the platform team is paying the bill the application team's velocity gain created.

What this means for you

Three things to watch for inside your company this quarter.

The first is whose team grew this year and whose did not. If app engineering is hiring while platform, data, and security are flat, the company is building up capacity debt. Ask the platform lead privately what the incident backlog looks like. If the answer is "growing," the AI win on the board deck has a footnote nobody has written.

The second is whether your AI committee can say no. If it cannot, it is a procurement committee, not a governance committee. Pope Leo's specific phrase was "independent oversight." Translate that for an operator: somebody whose pay is not tied to AI ROI gets a vote on what ships. If the only people in the room own the app team's speed numbers, the room is built so it cannot say no to the deploy that creates the platform team's next outage.

The third is whether a Starbucks-shaped failure is being tracked anywhere in your portfolio. Every operator I have spoken to in the last six months has at least one pilot that is technically live but practically retired. The cost is being absorbed somewhere. It could be a manual workaround in operations, an extra QA pass, or a quiet "we just stopped using that feature" that never made it back to leadership. In my experience, 30 to 50 percent of the AI inventory inside a typical mid-market company is dollar-active and functionally retired. Operators who do not run that audit this quarter are running a budget against a fiction.

From the field

The pattern that shows up almost every week is that the operators I work with budget for the build and never for the operate. They put a number on the AI pilot. They do not put a number on the team that absorbs the output of the pilot for the next 18 months. We have seen this across 40-plus engagements and roughly 280 operator interviews. It is the single most reliable failure I can name. It is also not new. It is an org-design problem older than AI.

When I tell a CFO this, the response is usually some version of "the platform team will figure it out." That is the answer that produced the Starbucks situation. Eleven thousand stores ran wrong dairy counts for nine months, and nobody figured it out until somebody at the top finally pulled the plug, by which point the cost had compounded silently across the entire system.

The right answer is dull. Hire on both sides of the deploy. Fund the team that absorbs the speed gains, not just the team that creates them. Look at the AI budget for app work and ask if the platform-side budget is at least 30 to 50 percent of it. If not, the company is borrowing from its future ops team to pay for this year's speed story, and the loan has a variable rate.

SPONSORED BY CLUTCH

Hire secure AI teammates that work 24/7.

Hire pre-built AI teammates. Give your engineers and operators a platform to ship their own AI apps. Stop losing sleep about what is running where.

Clutch is the platform behind both: pre-built agents for the workflows your ops team should automate first, plus the integration plane your team's vibe-coded apps and Claude Code projects plug into. One platform. Real production. Visible and safe by default.

Built for ops, engineering, and security teams that are tired of the shadow-AI surface area inside their own company.

SINCE FRIDAY

• The Pope's encyclical Magnifica humanitas called for AI to be "disarmed."
  AI just became a board-level moral conversation, not a CIO-only one. If your leadership hasn't been asked about AI ethics by a board member yet, this is the week that changes.

• Cursor hit $3 billion in annualized revenue, up from $2 billion in February.
  SpaceX still holds the right to acquire it for $60 billion. Developer tooling is compounding faster than any enterprise software category in memory. Budget cycles that treat it as a line item are already behind.

• Polsia reportedly raised $30 million at a $250 million valuation while running coding, research, outreach, ads, and support through agents.
  The one-founder agent company is now a funded experiment, not a pitch deck. The headcount model this implies is worth sitting with before your next hiring plan.

• Prompt injection is being reframed as operational risk: data theft, deletion, memory poisoning, tool abuse, and remote code execution.
  When agents touch real systems, prompt injection is an infrastructure problem your CISO owns, not your content team. If it isn't on your security team's radar yet, it should be.

• BofA now models the AI-rack 800V DC power semiconductor market at $1.4B in 2026 and $11B in 2027. 
  The compute bottleneck is moving from chips to power delivery. That affects your cloud cost trajectory two budget cycles out, and most operators aren't pricing it in yet.

P.S. If your platform team has quietly become the absorbing function for every AI deployment in the company, that is not a quirk of your organization. It is the most common pattern we see.

Hit reply and tell me one specific AI deployment at your company that is technically live but practically retired. I read every reply, and I keep a list.

IF YOU GOT VALUE FROM THIS ISSUE

Keep going. The next 30 mornings, in your inbox.

The analysis you just read came out of the same operator practice that the 30 Days program teaches. Foundations, then connect, then build, then lead. By Day 30, you own nine concrete capabilities: prompting that hits your bar on the first draft, AI workflows that take five hours a week off your plate, agent skills your team can run, and the language to lead the AI conversation at your company.

Free. Thirty mornings, then it ends. No upsells, no community Slack. Inspired by Hilary Gridley's Couch to 5K for AI.