What You’ll Find This Week

AI software takes about twice as long to buy as ordinary software. Sixteen to twenty weeks against seven to ten, with 58% of AI purchases pulling seven or more people into the decision. All of that scrutiny lands before the money moves, and almost none of it lands after.

Brice Challamel ran Moderna's AI rollout before OpenAI hired him as its Head of AI Strategy and Adoption. Ten months ago he asked what the "I" in ROI actually stands for, and his answer was that most of what companies call AI investment is consumption: licenses, compute, pilots, consultants, with nothing durable left at the end.

This week: what Moderna's rollout actually produced and what it still can't prove, what a top-down edict settles and what it leaves open, and what happens to a business case once it's done its job.

Here’s what you’ll find:

Don’t Miss Our Latest Podcast

This Week’s Article

"Pop quizz: what's the 'I' in ROI?"

The obvious answer is "investment." But when Brice Challamel posed that question to LinkedIn ten months ago, he offered a different read: most of what companies spend on AI is consumption in the form of "licenses, compute, pilots, consultants." Nothing on the invoice tells you which one you bought. Same line item, same dollar figure, same vendor.

The difference shows up later in whether the company can do something it couldn't do before. Brice’s test for telling them apart is straightforward, "If there's no durable capability or competitive edge at the end, does it deserve to be called an investment?”

When he says most AI spending doesn't earn the word investment, he's grading work he did himself. Challamel ran Moderna's AI rollout as Head of AI Products and Platforms. OpenAI hired him out of that job to be its Head of AI Strategy and Adoption.

Greg Shove, CEO of Section, raised the obvious objection: "someone has to keep the CFO happy." Challamel's answer: "Shouldn't it be the CFO's role to keep everyone else happy?"

Capital projects get reviewed one way: name the scope, price it, project the return, approve or kill it. That works when you know what you're building before you build it. AI budget reviews use the same form and ask for the same number, but they do so at a point when nobody knows yet which use cases will prove impactful. The overhead arrives long before any verdict does.

Levelpath surveyed 300 US procurement and supply chain decision-makers in June 2026 and asked how long enterprise purchases take. For standard software over $10,000, the most common answer was seven to ten weeks. For AI software, sixteen to twenty. Committee size moved the same direction: 58% of AI purchases involve seven or more people, against 43% for standard software, and 28% pull in eleven or more. The three delays named most often were security reviews, vendor evaluations, and contract negotiations.

In short, the category today’s executives describe as “urgent” takes roughly twice the calendar to approve as ordinary software and needs a bigger room to do it.

The same question stops projects that already cleared procurement. Zapier surveyed 835 US managers and above in May 2026 and found that 84% of companies have at least one AI pilot that never reached production, with 38% saying their longest-running pilot had been sitting in testing for more than a year. Asked what blocked deployment, 27% named an inability to measure ROI.

Challamel tested Moderna's own internal chatbot, mChat, against Microsoft Copilot and ChatGPT Enterprise, then picked ChatGPT Enterprise on the strength of its net promoter scores. The company bought the tool its own employees rated highest. Challamel's summary of the approach: "We were never here to fill a bucket, but to light a fire."

What came out of it:

One of those GPTs is called Dose ID. It pulls clinical trial datasets together, visualizes them, and helps study teams select a dose with safety prioritization built into the recommendation. BioPharma Dive covered the rollout and tied it to Moderna's goal of bringing up to 15 new mRNA products to market within five years.

Moderna's stated objective, per OpenAI's case study, was "100% adoption and proficiency of generative AI by all its people with access to digital solutions in six months." Every number above is an adoption number: how many people used it, how often, how many things they built. None of them is a return. Moderna hasn't published one.

The 750 GPTs run on a platform Moderna could stop paying for tomorrow. The 40% of people who learned to build them would still know how. That skill is the part worth calling an investment.

Stéphane Bancel, Moderna's CEO: "If we had to do it the old biopharmaceutical ways, we might need a hundred thousand people today. We really believe we can maximize our impact on patients with a few thousand people, using technology and AI to scale the company."

Bancel made it clear from the outset: “We're looking at every business process, from legal, to research, to manufacturing, to commercial, and thinking about how to redesign them with AI." That's an edict from the top: every function is in scope, the direction is stated. Nobody in the company has to litigate the argument again before exploring AI.

But Bancel named no use cases, funded no specific pilots, and demanded no deliverables. The target was 100% adoption. Nobody wrote down which 750 tools should exist. While the permission came from the top, specification and form came from the bottom up. As a result, the company ended up with tools nobody had thought to ask for before.

The edict settled whether Moderna would use AI, and the adoption numbers followed because employees were given the freedom to determine the how and the what. Whether any of it paid off is a separate question, and Moderna hasn't answered it publicly.

MIT's NANDA report found that 61% of enterprise AI projects were approved on projected ROI that was never formally measured after deployment. Someone built the model. Someone approved it. Nobody opened it again. The projection existed to unlock the budget, and once the budget was unlocked it was finished.

The approval had a date, a meeting, and somebody's signature on it. Revisiting that projection a year later has none of those things, and it can only establish one thing: whether the person who signed was wrong. You can see why nobody is in a rush to schedule that meeting.

Two months ago I wrote that the AI budget pyramid is upside down: companies pour money into tools and infrastructure while starving the organizational work that decides whether anyone ever uses them. The same pattern holds here: money gets watched on its way out the door and forgotten the moment it lands.

Six in ten companies have a business case on file and no idea whether it was true.

If AI becomes infrastructure the way electricity did, every company ends up with the same tools at roughly the same price. Owning them stops meaning anything. Your competitor can buy the same subscription this afternoon, and what they can't buy is the time your people spent learning to build with it.

Pull the list of AI initiatives your company funded in the last eighteen months. For each one, ask what the organization can do now that it couldn't do before, and whether the newly developed capability could survive the cancellation of a contract.

For most of them the honest answer is nothing. There's a subscription, a number of seats, and a renewal date, and nobody can point to anything the company does differently. Saying that out loud is more useful than defending the projection that got it approved.

But a handful of those initiatives will be different. Somebody redesigned a workflow and it stayed redesigned. A team learned to build its own tools and kept building them after the pilot ended. A function stopped waiting on a vendor to ship a feature.

Most of that list will renew on schedule. Nobody will have decided anything.