Finn Brooks: Hey, quick question before we start — do you trust numbers that are too tidy? Like, suspiciously round?
Juniper Vale: I mean, depends on what they're measuring. Why?
Finn Brooks: Because NBER just dropped a working paper — six thousand executives across four countries — and the finding is basically: seventy percent of companies are using AI, and almost ninety percent of them report zero productivity change. And I keep staring at that and thinking, is that number too clean to be real, or is that just actually what happened?
Juniper Vale: No, I think that's what happened. And the part that really stops me is the McKinsey side of it — eighty percent of employees are telling surveyors they personally feel faster with AI. Only thirty-seven percent of organizations saw any EBIT impact at all.
Finn Brooks: So workers feel it, companies don't.
Juniper Vale: That gap — that's the whole thing. One clean way to say it: individual speed is real, organizational results are not. And the question hanging over all of it is, where does the speed actually go? Today we're trying to work that out.
Finn Brooks: And I want to get into the companies because — Meta, Walmart, Microsoft, Amazon, Uber, these are not small bets. These are the companies that went hardest and are now the ones retreating most visibly.
Juniper Vale: Right — the speed story and the retreat story are the same story, I think.
Finn Brooks: But here's the part that breaks the 'early days' defense for me — the McKinsey number isn't vague. Eighty percent of employees feel faster. Thirty-seven percent EBIT impact. That's not a measurement problem, that's a conversion problem. The speed is happening, it's just... disappearing somewhere.
Juniper Vale: Okay, think of it like this — a customer service rep at an insurance firm uses AI to draft claim responses forty percent faster. That's real. Her fingers are moving faster, the drafts are going out faster. But three quarters later? Her team's case-closure rate hasn't moved. Customer satisfaction scores, unchanged. The company's combined ratio, unchanged.
Finn Brooks: So she's lapping herself and nobody's catching the baton.
Juniper Vale: Exactly. And this is where the Richmond Fed data gets interesting to me — because these aren't optimistic tech leaders being surveyed. These are CFOs. Late 2025, and sixty-five to seventy percent of them expected no change or only modest gains going into this year. They know.
Finn Brooks: Wait, CFOs specifically? Not — like, not product people, not CTOs?
Juniper Vale: Richmond Federal Reserve CFO surveys, yeah. Finance chiefs. The people whose entire job is — you know — did the money move.
Finn Brooks: No wait, that actually reframes it. Because I kept thinking the gap was about executives being too optimistic or too far from the work, but if the CFOs are already clocking no gains and the number is still sixty-five to seventy percent expecting nothing... the problem isn't awareness. They see the gap. They just don't have — I mean, what even is the connective tissue that would close it? Like, what's the mechanism that turns the rep drafting faster into the combined ratio improving?
Juniper Vale: That's the thing nobody's built yet. The individual speed is real and the org-level result is real — the nothing between them is where this whole story actually lives. And that's what I want to pull apart next.
Finn Brooks: And that nothing — that gap — is actually measurable now. MIT put a number on it: ninety-five percent of generative AI pilots never scale to production with measurable returns. Ninety-five.
Juniper Vale: Wait, that's not 'most fail.' That's almost all of them.
Finn Brooks: Right, and Deloitte's 2026 State of AI report lands in basically the same place — only twenty-one percent of AI projects reach production with demonstrable ROI. Even though worker access to these tools surged fifty percent in the same period. More access, same wall.
Juniper Vale: The retreat is — I mean, S&P Global Market Intelligence tracked this directly. Companies scrapping most of their AI projects went from seventeen percent in 2024 to forty-two percent in 2025. That's not a blip, that's a doubling of retreat in a single year.
Finn Brooks: Okay but the most concrete version of all of this — Walmart. They built Code Puppy, which is their in-house AI agent for spreadsheets and presentations, employees loved it, usage went way past what budget projected, and Bloomberg reported that on June 1st, 2026 Walmart just... capped everyone's monthly token allotment. Like, rationed it.
Juniper Vale: And I don't think that's a failure story. That's almost worse — it's an adoption-outpaced-planning story. The thing worked well enough that people actually used it, and the company wasn't ready for that.
Finn Brooks: So it's not 'AI didn't catch on,' it's — wait, the AI agents part of this is what gets me. These aren't chatbots anymore, they're autonomous systems designed to act with minimal human oversight, and Walmart, Microsoft, Meta, Amazon, Uber — they all deployed at scale before governance was actually in place. The token cap is governance arriving after the fact.
Juniper Vale: Speed first, consequences later — every single time across that whole list.
Finn Brooks: And honestly — which we'll get into — executives can see this is happening and the money keeps flowing anyway, while headcount decisions are already being made that you can't just reverse if the tech doesn't deliver.
Juniper Vale: And the can't-reverse part — that's what I want to sit with, because that's where this stops being abstract. Gartner surveyed 350 executives, published May 5th this year, and eighty percent said they'd already made AI- or automation-linked workforce reductions. Eighty percent.
Finn Brooks: Okay but wait — did those cuts actually help?
Juniper Vale: That's the thing. The workforce reduction rate was nearly identical across high-ROI and low-ROI companies. No measurable link. Cutting people to fund AI did not produce better returns than not cutting people.
Finn Brooks: That is — wait, zero advantage? Like statistically indistinguishable?
Juniper Vale: No measurable link. That's Gartner's language. And yet — I mean, the people are gone. You don't just un-restructure a team because the ROI didn't arrive.
Finn Brooks: So the irreversibility is the actual trap. Because — okay, here's what I can't square. NBER, same paper, executives are forecasting one point four percent productivity growth over the next three years. After three years of basically nothing. Where does that number even come from? That's not — actually, is that hope? Is that just institutional face-saving?
Juniper Vale: I don't think it's hope exactly. I think investment has become a competitive-necessity story that's fully decoupled from the math. You invest because you're afraid of falling behind, not because your own spreadsheets say you should. And once you've cut headcount to fund it — you're committed. Backing out is more expensive than doubling down.
Finn Brooks: Which is why The Guardian's October 2026 framing hit so hard — because when they reported on companies that rushed deployment with no governance frameworks, the consequence wasn't just slow returns. It was active cleanup costs. Errors that someone had to fix, which erased whatever thin gains existed.
Juniper Vale: And that cleanup cost lands on the workers who are still there — the ones who weren't cut. Which means the Gartner finding almost explains itself: cutting staff to fund AI doesn't improve ROI because the remaining workers absorb the error-correction burden the governance gap created. The math was never going to close.
Finn Brooks: And that's — I mean, I can't land on this. The National Bureau of Economic Research surveyed nearly six thousand executives, four countries, three years, and found almost no signal. Not a weak signal. Almost none. That's not a measurement problem at this point, that's just... the answer.
Juniper Vale: I keep trying to find an alternative read and I can't. Eighty-nine percent of organizations, no productivity change. And the forecast is one point four percent growth over the next three years — with no mechanistic reason to believe it arrives differently than the last three years did. I don't know what to call that.
Finn Brooks: Like, is the technology even the point anymore?
Juniper Vale: That's the one I can't shake. Maybe AI adoption became the cover story — for restructuring, consolidation, decisions that were going to happen anyway. And the technology just handed everyone a cleaner sentence to say out loud. I don't know if that's cynical or just accurate.
Finn Brooks: Both, maybe. I'm genuinely not sure either. Which is an uncomfortable place to stop but probably the honest one.
